Analyzing EEG
-
Upload
norma-khoirun-nisa -
Category
Documents
-
view
13 -
download
1
description
Transcript of Analyzing EEG
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 111
Letters
Epileptic seizure detection by analyzing EEG signals using
different transformation techniques
Mohammad Zavid Parvez n Manoranjan Paul
Centre for Research in Complex Systems School of Computing amp Mathematics Charles Sturt University Bathurst NSW 2795 Australia
a r t i c l e i n f o
Article history
Received 13 August 2013
Received in revised form4 May 2014
Accepted 24 May 2014
Communicated by Bo ShenAvailable online 2 June 2014
Keywords
EEG
Epilepsy and seizure
Ictal and interictal
EMD
LS-SVM
a b s t r a c t
Feature extraction and classi1047297cation are still challenging tasks to detect ictal (ie seizure period) and
interictal (ie period between seizures) EEG signals for the treatment and precaution of the epileptic
seizure patient due to different stimuli and brain locations Existing seizure and non-seizure featureextraction and classi1047297cation techniques are not good enough for the classi1047297cation of ictal and interictal
EEG signals considering for their non-abruptness phenomena inconsistency in different brain locations
type (generalpartial) of seizures and hospital settings In this paper we present generic seizure
detection approaches for feature extraction of ictal and interictal signals using various established
transformations and decompositions We extract a number of statistical features using novel ways from
high frequency coef 1047297cients of the transformeddecomposed signals The least square support vector
machine is applied on the features for classi1047297cations Results demonstrate that the proposed methods
outperform the existing state-of-the-art methods in terms of classi1047297cation accuracy sensitivity and
speci1047297city with greater consistence for the large size benchmark dataset in different brain locations
amp 2014 Elsevier BV All rights reserved
1 Introduction
The human brain processes sensory information which is
received by externalinternal stimuli Human brain is an organic
electrochemical computer as neurons exploit chemical reaction to
generate electricity [1] Electroencephalogram (EEG) is a graphical
record of ongoing electrical activity which measures the changes
of the electrical activity in term of voltage 1047298uctuations of the brain
through multiple electrodes placed in the different location of the
brain [2] In the clinical contexts the main diagnosis of EEG is to
discover abnormalities of brain activities referred to the epileptic
seizure A seizure occurs when the neurons generate uncoordi-
nated electrical discharges that spread throughout the brain and
epilepsy is a recurrent seizure disorder caused by abnormal
electrical discharges from brain cells often in the cerebral cortex
[3] Another clinical use of EEG is in diagnosis of coma brain deathencephalopathies sleep disorder etc Moreover EEG can be used
in many applications such as emotion recognition [4] video
quality assessment [5] alcoholic consumption measurement [6]
sleep stage detection [7] change the brainwaves by smoking [8]
mobile phone usages [9] etc
People may experience abnormal activities in sensation move-
ment awareness and behavior during seizure as a result they
cannot perform normal task Ictal and interictal both are medical
condition of seizure where ictal represents the period of seizureand interictal represents the intermediate period between two
seizures Note that interictal is signi1047297cantly different from normal
non-seizure (mainly for healthy people) signal in terms of signal
characteristics A prediction of seizures (ie ictal) from interictal
could aware a patient to put away from the next seizure and also
can make sure better treatment and precaution A number of
existing EEG features extraction and classi1047297cation techniques
[11ndash17] are able to classify seizure and non-seizure EEG signals
for a small size dataset almost perfectly However those methods
are struggling to provide acceptable classi1047297cation accuracy for ictal
(ie seizure period) and interictal (ie period between seizures)
EEG signals due to the non-abruptness and inconsistence phenom-
ena [10] of the signals in different brain locations Moreover it is
more challenging to get acceptable label of accuracy by a particularmethod if we do not have any domain knowledge of the available
EEG signals due to different types of seizures (ie generalized or
partial) patients hospital setting brain location and artifact Note
that ictal signals are considered the EEG signals during the seizure
period when a patient shows abnormal activities whereas inter-
ictal signals are considered as non-ictal signals between two
seizure periods of an epileptic patient Thus we can consider the
characteristics of interictal signal as a middle stage of non-seizure
and ictal signals (although a patient may show normal brain
activities similar to the non-seizure signals during the interictal
period)
Contents lists available at ScienceDirect
journal homepage wwwelseviercomlocateneucom
Neurocomputing
httpdxdoiorg101016jneucom201405044
0925-2312amp 2014 Elsevier BV All rights reserved
n Corresponding author
E-mail address mparvezcsueduau (MZ Parvez)
Neurocomputing 145 (2014) 190ndash200
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 211
Existing methods [11ndash18] used small epilepsy dataset [19] (for
detailed information of the dataset see Section 31) The small
dataset with duration 236 s has seizure (ie ictal) and non-seizure
signals which can be distinguished by their visual phenomena
such as magnitude of amplitude and changing rate of frequency
(see in Fig 1(a)) For the non-seizure signal the amplitude is low
and the frequency is high while the nature of seizure signal is
totally opposite (see Fig 1(a)) Fig 1(b) and (c) shows the ictal and
interictal EEG signals from a large dataset [22] The 1047297gure
demonstrates the non-abrupt phenomena (ie not easily distin-
guishable between ictal and interictal based on amplitude and
frequency) of the ictal and interictal signal for both cases of Frontal
and Temporal lobes compared to the dataset in [19] Moreover
EEG signals from different locations exhibit different phenomenal
activities for an ictal and interictal period Thus the classi1047297cation
of ictal and interictal EEG signals is challenging compared to that
of seizure (ie ictal) and non-seizure EEG signals It also demon-
strates that the characteristics of EEG signals are varied fromdataset to dataset
Birvinskas et al [17] used DCT (discrete cosine transformation) to
reduce the EEG signals and then extracted features from the
reduced signals to classify EEG signals They achieved good
classi1047297cation accuracy by keeping low frequency DCT coef 1047297cient
for feature extractions Worrell et al [33] claimed that the seizure
onset frequencies are in the range of gamma frequency ie high
frequency band (typically between 30 to 100 Hz) They explicitly
mentioned that seizure onset frequencies are on the range of
frequency 74376 137 and 23776 28 Hz for Frontal and
Temporal lobes respectively They also observed that dominant
spectral peaks are observed in the high frequency spectral com-
ponents for the Frontal lobe and in the lower gamma frequency for
the Temporal lobe during the seizure period Other literature such
as Panda et al [11] Ocak et al [13] and Ghosh-Dastidar et al [12]
used other frequency bands such as delta theta alpha and beta
with gamma band to extract features for EEG signal classi1047297cation
From the above 1047297ndings and the characteristics of the different
datasets we 1047297nd that the frequency range of seizure onset can be
varied based on the brain location seizure type (generalpartial)
patient and hospital setting (eg capturing machine quality
artifacts etc) However we observe that the signal has larger
spikes ie high amplitude in the ictal (ie seizure) period
compared to the interictal period Thus based on the temporal
and spatial features of raw EEG signals and the above 1047297ndings of
the recent literatures our hypothesis is that the distinguishing
features between ictal and interictal signals would be in the high
frequency spectral area irrespective of the brain location patient
type of seizure and hospital setting
To verify our hypothesis we perform an experiment to 1047297nd the
gamma frequency of signals which provides the maximum ampli-
tude using ictal and interictal signals in different brain locationsThe results for 200 EEG signals from ictal and interictal are shown
in Fig 2 The 1047297gure shows that relatively high gamma frequencies
ie high frequencies in the signals provide the maximum ampli-
tude of ictal compared to the interictal signals for both Frontal and
Temporal lobes from different patients We use the maximum
amplitude criteria for the gamma band selection in this analysis as
our observation and Worrell et als observation indicate that
dominant spectral peaks for the seizure onset are in the high
frequency (ie gamma band) Thus the features extracted from
gamma frequencies ie the high frequencies should provide better
classi1047297cation accuracy between the ictal and interictal irrespective
of brain location patients type of seizure and hospital setting
Based on the above mentioned observation in this paper we
proposed generic seizure detection techniques which use different
Fig 1 Ictal and interictal EEG signals from different datasets and different brain locations (a) Seizure and non-seizure EEG signals from the small dataset [19] the 1047297rst row
and remaining four rows indicated the seizure and non-seizure signals respectively [19] (b) Ictal (patient one 7th datablock channel one) and interictal (patient one 95th
datablock channel one) of Frontal lobe signals [22] and (c) Ictal (patient 1215th datablock and channel one) and interictal (patient 12 35th datablock and channel one) of
Temporal lobe signals [22]
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 191
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 311
established transformations and decompositions to get different
statistical features from the high frequencies component of the
signals to classify the ictal and interictal signals using the least
square support vector machine (LS-SVM) [23] Although we use
a number of established transformation and decompositions we
use weighted high frequency coef 1047297cients to extract the features In
the experiment we investigate the strength of each transformation
and decomposition for the capability of classi1047297cation accuracy as
well as the combined strength through different feature extrac-
tions We use a large dataset [22] consists of raw ictal and
interictal signals (no explicit artifact removal technique is applied)
from different brain locations different seizures and patients to
verify the performance of the proposed methods against two
state-of-the-art methods The experimental results demonstrate
that the proposed generic methods outperform the existing
methods in terms of accuracy (ie overall) sensitivity (ie ictal
or positive) and speci1047297city (ie interictal or negative) with greater
consistence for the challenging ictal and interictal EEG signals
classi1047297cation
The contributions and novelties of the proposed methods are in
the capacity to (i) classify challenging ictal and interictal signals
irrespective of brain location patient type of seizure and hospital
setting with a certain range of artifact tolerance (ii) identify that
high frequency component of EEG signals has the distinguishing
features to classify ictal and interictal EEG signals (iii) formulate
a weighing vector for high frequency coef 1047297cients to provide
individual importance of different coef 1047297cients (iv) provide the
rationality and way of extracting high frequency components from
each transformationdecomposition technique (v) to analysis EEG
signals for seizure detection best of our knowledge we are the 1047297rst
one to use two dimensional SVD (vi) identify the distinguishing
features (provided pictorial comparison) for ictal and interictal
signals for classi1047297cation and (vii) provide comprehensive result
analysis using receiver operating characteristics (ROC) curve Whis-ker-box graphical view of classi1047297cation outcome and classi1047297ca-
tion performance in terms of sensitivity speci1047297city and accuracy
2 Existing methods for seizure detection
Existing feature extraction and classi1047297cation methods based on
wavelet [11ndash13] and Fourier transformation [14] were employed
for the detection of seizure in EEG signals Panda et al [11]
computed various features such as energy entropy and standard
deviation (STD) using discrete wavelet transformation (DWT) and
used support vector machine (SVM) as a classi1047297er Dastidar et al
[12] applied wavelet transformation to decompose the EEG signals
into different range of frequencies and then extracted three
features such as STD correlation dimension and the largest
Lyapunov exponent (quantifying the non-linear chaotic dynamics
of the signals) and applied different methods for classi1047297cation
Ocak [13] proposed a fourth level wavelet packet decomposition
method to decompose the normal and epileptic EEG epochs to
various frequency-bands and then used genetic algorithm to 1047297nd
optimal feature subsets which maximize the classi1047297cation perfor-
mance Polat et al [14] extracted feature using fast Fourier trans-
form (FFT) and then classi1047297ed using decision making classi1047297er The
techniques [111214] have used small dataset [19] (for detailed
information of the dataset see Section 21) Sample examples of
the dataset are shown in Fig 1(a) The classi1047297cation accuracies
were 9120 [11] 9670 [12] and 9872 [14] respectively Ocaks
[13] experimental dataset was of the same pattern with the
dataset in [19] and the classi1047297cation accuracy was 98
Other feature extraction and classi1047297cation methods [15ndash18]
based on empirical mode decomposition (EMD) principle component
analysis (PCA) and genetic algorithm (GA) have been proposed
Liang et al [15] extracted EEG features where the approximate
entropy(ApEn) analysis was evaluated for its ability to analyze
multiclass EEGs and the performance of ApEn analysis was
enhanced by incorporating the power of EEG sub-bands or auto-
regressive models PCA and GA were compared for their ability to
select features by applying various linear and nonlinear methods
for classi1047297cation The maximum accuracy was 9867 Pachori [16]
decomposed EEG signals into intrinsic mode function (IMF) using
EMD and then computed mean frequency for each IMF by Fourierndash
Bessel expansion [20] to differentiate seizure and non-seizure EEG
signals Zhang et al [21] proposed an approach for the pattern
recognition of four complicated hand activities such as grasping
a football a small bar a cylinder and a hard paper based on EEG
signal in which each piece of raw data sequence for EEG signal is
decomposed by wavelet transform After forming a matrix they
applied singular value decomposition (SVD) on the matrix to 1047297ndthe singular value feature vectors These features are used as input
to arti1047297cial neural network to discriminate four hand activities and
they got 8900 classi1047297cation accuracy
Recently Bajaj et al [18] extracted amplitude modulation and
frequency modulation bandwidth as features from IMF using EMD
Among the existing contemporary techniques Bajaj et als tech-
nique is the latest and the best in terms of performance They used
the LS-SVM [23] technique for the classi1047297cation of seizure and
non-seizure EEG signals using the small dataset [19] They
obtained 980 to 995 accuracy using radial basis function (RBF)
kernel and also obtained 995 to 100 accuracy using Morlet
kernel
We have observed that the existing methods [11ndash18] exhibit
almost perfect classi1047297cation accuracy for the small dataset of
Fig 2 Gamma frequencies which provide the maximum amplitude for ictal and interictal signals (from a standard dataset [22])
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200192
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 411
seizure and non-seizure EEG signals However their performance
is below the expected results using the large dataset [22] for ictal
and interictal EEG signal classi1047297cations For example when we
applied the Bajaj et als [18] feature extraction and classi1047297cation
method to differentiate ictal and interictal EEG signals for the 1047297rst
IMF in the large dataset from Epilepsy Center of the University
Hospital of Freiburg [22] the classi1047297cation accuracy is 7790 for
Frontal lobe and 8020 for Temporal lobe which is far below the
accuracy 9850 obtained using the small dataset [19] of seizureand non-seizure EEG signals
The paper is organized as follows the details of the proposed
four methods with dataset feature extractions classi1047297cation
techniques and computational time are described in Section 3
the detail experimental results are explained in Section 4 while
Section 5 concludes the paper
3 Proposed methods
In this paper we propose four methods based on the various new
features extracted from DCT DCTndashDWT SVD and IMF We use the
transformation and decompositions to separate high frequency
components and then apply innovative ways to extract different
statistical features When we extract features from the high fre-quency component (ie IMF and detail signals respective) using EMD
and DWT we take total length of the high frequency components
and we do not use any weighting for them as all the coef 1047297cients
represent the high frequency components As SVD and DCT provide
coef 1047297cients of all frequencies we need to select a subset of the
coef 1047297cients which can represent high frequency components We
investigate the strength of each transformation and decomposition
for the capability of classi1047297cation accuracy as well as the combined
strength through different feature extractions The dataset individual
method their corresponding rationality and detailed methodology of
feature extractions are described in the following sub-sections
31 Dataset
Small Dataset [19] The existing seizure and non-seizure feature
extraction and classi1047297cation schemes [11ndash18] except [13] used
small dataset [19] The dataset consists of 1047297ve subsets each subset
contains 100 single-channels recoding and each recoding has
236 s in duration captured by the international 10ndash20 electrode
placement scheme ie 32 electrodes [14] Among them two
subsets are taken from healthy volunteers two subsets are taken
from seizure free intervals and another subset is also taken during
seizure period [18] (see sample examples in Fig 1(a)) Ocak [13]
used different datasets with the same pattern of small dataset
where 300 were normal and 100 were epileptic patients
Large Dataset [22] Seizure and brain discharges are much more
complicated than what can be seen in the scalp Thus in the paper we
have used intracranial EEG dataset using grid strip and depth electrodes
rather than scalp EEG The dataset used in the paper was recorded at
Epilepsy Centre of the University Hospital of Freiburg Germany [22] The
database contains invasive EEG recordings of 21 patients suffering frommedically intractable focal epilepsy The data obtained by Neuro1047297le NT
digital video EEG system with 128 channels 256 Hz sampling rate and
16 bit analog-to-digital converter According to the dataset [22] epileptic
EEG signals can be classi1047297ed into (i) ictal (ie seizure) (ii) preictal (a
certain period before seizure onset) (iii) interictal (period between
seizures) and (iv) non-seizure (period without any seizure activity after
a certain time of interictal) Normally duration of ictal is varied from few
seconds to 2 min The ictal records contain epileptic seizures with at
least 50 min of preictal data preceding each seizure The median time
period between the last seizure and the interictal data set is 5 h and
18 min and the median time period between the interictal data set and
the 1047297rst following seizure is 9 h and 36 min [29] As suggested in the
dataset we have treated preictal and ictal period as an ictal signals in
our experiments We do not consider any non-seizure signals into the
interictal period as we carefully select the same duration of interictal
signal (with the ictal signal) from the above duration of interictal signal
32 Feature extraction using DCT
DCT is a transformation method for converting a time series
signal into basic frequency in such a way that the DCT coef 1047297cients
are arranged from low frequency to high frequency components
Low frequency components represent the coarse signals and high
frequency components represent the detail signals DCT is widely
used in the image processing and video coding areas to compress
image signals based on their frequencies [34] As the ictal and
interictal EEG signals have different amplitudes and frequencies
(not visually separable) thus the most distinguishable featuresshould be located in the high frequency components of DCT
coef 1047297cients see in Fig 3(a) As the EEG signal is non-linear and
non-stationary in nature thus for real time processing of EEG
signals DCT may not be correct to directly correspond to the
frequency analysis however if we segment the EEG signals in time
window and apply DCT on them to 1047297nd DCT coef 1047297cients we can
avoid non-linear and non-stationary nature of the signals In our
experiment we apply DCT on the benchmark dataset [22] and take
Fig 3 (a)High frequency DCT coef 1047297cients and their energy and entropy for the signals of Temporal lobe the 1047297rst two rows show the high frequency DCT coef 1047297cients (ie
25) of ictal and interictal signals while the last row shows the energy and entropy of using the high frequency DCT coef 1047297cients (ie 25) (b) High frequency DWT
coef 1047297cients and STDs of high frequency DCT and DWT coef 1047297cients for the EEG signals of Temporal lobe the 1047297rst two rows show that the high frequency DWT coef 1047297cients of
ictal and interictal signals while the last row shows the STD of the high frequency DCT and DWT coef 1047297cients of ictal and interictal EEG signals
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 193
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 511
the higher frequency component (ie the last quartile (Q4) ie last
25 of DCT coef 1047297cient) Then we extract statistical features such as
entropy and energy of the weighted higher frequency coef 1047297cients to
classify ictal and interictal signals
The extraction of the statistical features from the Q4 is not
straightforward We use heuristics to provide different weight to
the different coef 1047297cients and investigate the classi1047297cation accu-
racy Although we try different type of weighting approximations
the experimental results show that exponential weight performswell The weight w i for the i-th position coef 1047297cient is selected as
wi frac14 β e β xi where xi x j frac14 1
Number of Selected Coeff icients We also try
different β and test for the classi1047297cation results We 1047297nd that
β frac142 provides the best results for the dataset Fig 4 shows an
example of weights According to the proposed weight we provide
more weight to the relatively low frequency coef 1047297cients and less
weight to the relatively high frequency coef 1047297cients within the
higher frequency area of the overall signal as sometimes the high
frequency coef 1047297cients in the tail would be very small
We de1047297ne weighted entropy and energy as follows
Energy frac14 sumn
i frac14 1
W i X ietht THORN
2and Entropy frac14 sum
n
i frac14 1
W iY 2i etht THORNlnethY 2i etht THORNTHORN eth1THORN
where n is the length of signal X (t ) is the higher frequency DCTcoef 1047297cients and Y (t ) is the normalized value of X (t )
The representation of ictal and interictal data for different higher
frequency DCT coef 1047297cients and their entropy and energy are given in
Fig 3(a) It can be observed that the changing rate of frequencies is
higher for interictal signal than ictal signal from Temporal lobe It is
also interesting to observe that the amplitude of the higher frequency
DCT coef 1047297cients for ictal signals is larger compared to that of interictal
signals Note that although the original ictal and interictal EEG signals
are not visually separable the high frequency DCT coef 1047297cients of ictal
and interictal signals are visually separable The third row of Fig 3
(a) also demonstrates that the weighted energy and entropy extracted
from the higher frequency DCT components have different values for
ictal and interictal signals Thus weighted entropy and energy would
be effective features for ictal and interictal signals classi1047297
cation
33 Feature extraction using DCT ndashDWT
DWT is a transformation and widely used in the image and signal
processing research to decompose the imagesignals into different
frequency and bands DWT can be speci1047297ed by low pass (LP) and high
pass (HP) 1047297lters named as standard quadrature mirror 1047297lters [28] The
cut-off frequency of these 1047297lters is one-fourth of the sampling
frequency The bandwidth of the 1047297lter outputs is half the bandwidth
of the original signal which generates downsampled output signals
without losing any information according to the Nyquist theorem [13]
The down-sampled signals from LP and HP 1047297lters refer to 1047297rst-level
approximation and details of the original signal respectively [24] DWT
is sometimes more effective compared to DCT to separate signals into
low and high frequency components when the signal is longer Thus
we also investigate the effectiveness of the DWT to extract some other
statistical features (such as STD) to classify ictal and interictal signals
In our second approach we apply single-level one dimensional DWT
on the EEG signal to 1047297nd the high frequency components from the
signal Then we calculate the STD from the detail frequency (ie highfrequency) components of the DWT coef 1047297cients and use as a feature
We use the whole high frequency component for feature extraction as
DWT separates the signals into low and high frequency components
Another feature is extracted using the STD of the weighted high
frequency DCT coef 1047297cients (similar to Fig 4) Two STD features from
DCT and DWT higher frequency components are then used for the
classi1047297cation purpose Note that we use different primary functions of
the wavelet transformation including Daubechies (db) series We 1047297nd
that db16 is the best for our experiment
In the experiment we take only the high frequency DWT compo-
nents The 1047297rst two rows in Fig 3(b) show the higher change of
frequency and lower amplitude for interictal signal compared to ictal
signal from Temporal lobe The third row of Fig 3(b) represents the
STDs generated from the high frequency DCT coef 1047297cients (weighted)
and DWT for ictal which are higher than that of interictal signal Thus
the STDs extracted from the high frequency DCT (weighted) and DWT
coef 1047297cients can be good features for classi1047297cation
34 Feature extraction using SVD
SVD decomposes two dimensional signals into three matrixes in
such a way that the middle matrix (ie singular value) has unique
properties to separate signals The middle matrix containing the
square root of eigenvalues and diagonalized contains much less non-
zero coef 1047297cients compared with the coef 1047297cient matrices of other
transformations [35] The values of diagonal matrix are with des-
cending order from top-left to bottom-right As the original ictal and
interictal signals are almost of the same nature we do not expect asigni1047297cant different values of other two matrixes after SVD however
we expect some distinguishable values on the extreme part (ie
bottom-right) of diagonal matrix Thus any feature extracted from
the bottom-right singular values could be used to classify ictal and
interictal signals To the best of our knowledge we are the 1047297rst to use
two dimensional SVD to analyze EEG signals for seizure detection
To extract the features using SVD we reshape the EEG signals
into a two dimensional matrix The column vector of EEG signal is
rearranged to square matrix for the computation of singular value
components using SVD From the singular values we take the
weighted Q4 (weight is similar to Fig 4 but the number of selected
coef 1047297cients are different) of the non-zero diagonal values to
calculate the STDs and energy and treat as the two features to
classify ictal and interictal signals SVD allows transforming any
given matrix AAC mxn to diagonal form using unitary matrices ie
A frac14 U Σ V H where Amn is a matrix U mm and V nnare orthogonal
matrix and Σ mn matrix with non-negative diagonal entries
The columns of U are called left singular vectors and the rows
of V H are called right singular vectors If A is a rectangular m n
matrix of rank K then U will be m m and Σ will be
Σ frac14
λ1 0 ⋯ 0
0 λ2 ⋯ 0
⋮ ⋮ ⋱ ⋮
0 0 ⋯ λn
⋮ ⋮ ⋮
0 0 ⋯ 0
2666666664
3777777775
eth2THORN
where λ1Z λ 2Zhellip λn-1 Z λnFig 4 Weight for different coef 1047297cients of selected high frequency coef 1047297cients
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200194
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 611
The data distribution of non-zero singular values of the ictal
and interictal signals from the Frontal lobe is shown in Fig 5(a)
The top part of the 1047297gure clearly shows that the Q4 singular values
of interictal signals are larger than that of the ictal signals The
bottom part of Fig 5(a) shows the STD and energy extracted from
the weighted Q4 of non-zero singular values from ictal and
interictal signals respectively The 1047297gure exposes that the STDs
extracted from the ictal and interictal signals have distinguishable
values thus they can be used for ictal and interictal signal
classi1047297cation
35 Feature extraction of IMF using EMD
The main strength of EMD is its ability to analyze nonlinear and
non-stationary signals like EEG signals When we apply EMD on
a signal the signal is decomposed into a 1047297nite number of signals
into IMFs which are ordered from higher frequency component to
lower frequency component The number of IMFs in a signal
depends on the local characteristics of the signal rather than
pre-de1047297ned number of IMFs This feature makes each IMF crucial
to identify some characteristics of the original signal Since the
IMFs are the representations of different frequenciesamplitudes
of the original signal and the distinguishing features between ictal
and interictal EEG signals lie on the frequencyamplitude our
hypothesis is that any feature extracted from IMFs will be an
effective feature to classify ictal and interictal successfully We also
observe that the ictal and interictal signals are deferred in higherorder frequency components thus our other hypothesis is that
features extracted from the 1047297rst few IMFs should be enough to
classify ictal and interictal signals successfully Through the
decomposition we keep only few mode functions of the original
signals for the 1047297nal feature extraction however we are able to
keep the important mode functions for the classi1047297cation purpose
Experimental results reveal that our both hypothesis are correct as
extracted features from the 1047297rst IMF successfully classify ictal and
interictal signals (see Section 4) Thus IMFs can be used in
classi1047297cation applications where the original signals are non-
linear and non-stationary and distinguishing features exist on
the different frequenciesamplitudes
Each IMF satis1047297es two following conditions (i) the number of
extrema and the number of zero crossing are identical or differ at
most by one and (ii) the mean value between the upper and the
lower envelope is equal to zero at any time
The EMD algorithm can be summarized as follows
1 Extract the extrema (minima and maxima) of the signal xetht THORN
2 Interpolate between minima and maxima to obtain
ℓminetht THORN and ℓmaxetht THORN
3 Calculate local mean metht THORN frac14 ℓminetht THORN thornℓmaxetht THORNfrac12 =2
4 Extract the detail detht THORN frac14 xetht THORN metht THORN
5 Check detht THORN is an IMF according to the conditions which are
mentioned above If yes repeat the procedure from the step1 on the residual signal r etht THORN frac14 xetht THORN metht THORN If no replace xetht THORN with
detht THORN and repeat the procedure from step 1
The pictorial scenario of generating IMFs is shown in Fig 6 The
signal xetht THORNcan be represented as a combination of IMFs and
residual component
xetht THORN frac14 sumL
i frac14 1
dietht THORN thornr Letht THORN eth3THORN
where di and r L are the i-th IMF and L-th is residual signal
We observe that features extracted from the 1047297rst two IMFs
provide the best classi1047297cation results To reduce computational
time we consider only the 1047297rst IMF from the EMD and then
calculate STD of IMF to capture the variation of the signal Byignoring other IMFs we do not sacri1047297ce any signi1047297cant classi1047297cation
accuracy The strength of a signal is distributed in the frequency
domain relative to the strengths of other ambient signals is central
to the design of any 1047297lter intended to extract or suppress the signal
To 1047297nd the strength of the signal we apply power spectral density
(PSD) on the 1047297rst IMF Then determine the STD feature from
normalized PSD The calculation of PSD is as follows
P ethwiTHORN frac14 1
N DethwiTHORN 2
eth4THORN
where DethwiTHORNis the discrete Fourier transform of IMFdietht THORN Normalized
PSD can be de1047297ned as
pi frac14 P ethwiTHORN=sum
i
P ethwiTHORN eth5THORN
Fig 5 (a) Non-zero singular value of SVD for Frontal lobe The 1047297rst row shows that the non-zero singular value distribution for ictal and interictal signals from Frontal lobe
and the last row shows the STD and energy of the last 25 non-zero singular values of ictal and interictal EEG signals (b) The 1047297rst IMFs of ictal and interictal signals with the
extracted features the top two rows represent the 1047297rst IMF from the ictal and interictal signals of Temporal lobe bottom row represents STD of the 1047297rst IMF and STD of the
normalized PSD (power spectral density) of the same IMF of ictal and interictal EEG signals
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 195
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 711
In the experiment we extract two features such as STD of the
1047297rst IMF and STD of normalized PSD of the 1047297rst IMF From the 1047297rst
two rows in Fig 5(b) it can be observed that the 1047297rst IMF of ictal
signal carries the higher amplitude compared to the interictal
signal
The third row of the 1047297gure represents the STD of the 1047297rst IMF
and STD of PSD of the same IMF from using ictal and interictal
signals The STDs for both cases of interictal signal are signi1047297cantly
larger than that of ictal signal Thus the two new features
extracted by EMD method can be good features for classi1047297cation
Table 1 presents the summary of all corresponding features use
for ictal and interictal classi1047297cation by each method
36 Classi 1047297er
We extract a number of features from different transforma-
tionsdecompositions To classify ictal and interictal signals we
need to use a classi1047297er The goal of a classi1047297er is to 1047297nd patients
states such as ictal (class 1) and interictal (class 2) using machine
learning approaches with cross-validation The challenge is to 1047297nd
the mapping that generalizes from training sets to unseen test
sets For the cross-validation data are partitioned into training set
and test set In our experiment we use 4-fold cross validation
where each dataset was randomly divided into 4 splits where
three of them were used for training and the fourth one used for
testing
To classify the ictal and interictal signals we extract various
features from DCT DWT SVD and IMF of EMD For classi1047297cation we
use SVM-based classi1047297er as the SVM is one of the best classi1047297ers in
the EEG signal analysis SVM is a potential methodology for solving
problem in linear and nonlinear classi1047297cation function estimation
and kernel based learning methods [25] It can minimize the
operational error and maximize the margin hyperplane as a result
it will maximize the classi1047297cation performance [25] LS-SVM is the
extended version of SVM and it is closely related to regularization
networks and Gaussian process and it also has primal-dual inter-
pretations [23] The major drawback of SVM is in its higher
computational burden of the constrained optimization program-
ming However LS-SVM can solve this problem [26]
The classi1047297er is a LS-SVM which learns nonlinear mapping
from the training set features xifrac141hellipnT where nT is the number
of training features into the patients state ictal ( thorn1) and interictal( 1) For unbiased classi1047297cation results [27] we divide the whole
1000 trials into four subsets in each subset we randomly select
75 samples for the validation training set and the remaining 25
for the validation testing set Let f yigi frac14 C 12 designate the LS-SVM
validation test outputs mapping to class 1 or class 2 The equation
of LS-SVM can be de1047297ned in [18] as
f eth xTHORN frac14 sign sumN
i 1
α i yiK eth x xiTHORN thornc
eth6THORN
where K ( x xi) is a kernel function α i are the Lagrange multipliers c
is the bias term xi is the training input and yi is the training
output pairs Linear RBF and Morlet kernel are used in our
experiments and RBF and Morlet functions can be de1047297ned in [18]
Fig 6 Extracted IMFs by EMD from Temporal lobe (Patient 12 channel one 15th datablock for ictal and 35th datablock for interictal)
Table 1Extracted features of proposed and existing methods
Proposed methods Feature-1 Feature-2
Q4WHF_SVD (Quartile four weighted high frequency component using SVD ) STD on 25 weighted high frequency
singular values
Energy on 25 weighted high frequency
singular values
Q4WHF_DCT (Quartile four weighted high frequency component using DCT ) Energy on 25 weighted high
frequency DCT coef 1047297cients
Entropy on 25 weighted high frequency DCT
coef 1047297cients
FIMF_EMD (First IMF using EMD ) STD on the 1047297rst IMF (ie high
frequency components)
STD on normalized PSD
Q4WHF_DCT ndash HF_DWT (Quartile four weighted high frequency component
using DCT and high frequency from DWT)
STD on 25 weighted high frequency
DCT coef 1047297cients
STD on detail signals (ie high frequency
components) after apply DWT
Combined Combined all above features
Existing methods Feature-1 Feature-2
EMD[18] Amplitude modulation bandwidth Frequency modulation bandwidth
DCT [17] Features from 6 of quantized DCT coef 1047297cients at lower frequency
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200196
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 811
as
keth x xiTHORN frac14 expeth x xi2=2s
2THORN eth7THORN
where s controls the width of the RBF function
keth x xiTHORN frac14 prodd
k frac14 1
cos ω0etheth xk xki THORN=aTHORN
h iexpeth xk xk
i =2a2THORN eth8THORN
where d is the dimension a is the 1047298exible coef 1047297cient and xki is the
k-th component of i-th training data Matlab codes for LS-SVMclassi1047297er are available at httpwwwesatkuleuvenacbesista
lssvmlab
37 Computational time
Feature extractions using IMF through EMD take longer time
due to two-fold iterations to 1047297nd all IMFs In the 1047297rst iteration
EMD decomposes a signal and checks whether the signal makes an
IMF and in the second iteration it subtracts the current IMF from
the signal and continues for the next IMF Even if we stop the
iteration for the 1047297rst IMF the EMD-based technique takes a full
1047297rst iteration for calculating the IMF The experimental data
reveals that the proposed technique using DCT DWT and SVD
takes less computational time compared to the EMD-based tech-nique as DCT DWT and SVD transformation is faster compared to
the EMD decomposition Thus the proposed methods (except
FIMF_EMD method ie First IMF using EMD) take insigni1047297cant
time compared to the existing method [18]
4 Experimental results
Sometimes EEG signals have line noise and other kinds of
artifacts due to muscle and body movements Notch 1047297lter and
independent component analysis (ICA)-based method are recom-
mended to remove the line noise and artifacts respectively for
suf 1047297cient amount of EEG signals [30ndash32] However in this experi-
ment we do not use any explicit noiseartifacts removal technique
to see the tolerance of the proposed method against noise
artifacts
We propose four methods and the experimental results of the
proposed methods are compared with existing methods [1718]
and corresponding features are presented in Table 1Our ultimate
target is to classify ictal and interictal EEG signals using LS-SVM
with different kernels such as RBF and Morlet where the values of regularization parameter and kernel parameter are 10 and 6
respectively for both kernels We also use linear kernel In this
paper four methods are proposed based on the new features of
ictal and interictal EEG signals captured from Frontal and Temporal
lobes using DCT DWT SVD and IMF of EMD to see the strength of
individual transformationdecomposition techniques for seizure
detection Then we also test the classi1047297cation accuracy by combin-
ing all features In our experiment we use 200 signals from ictal
and 800 signals from interictal EEG signals After training and
testing of all features we calculate average value from the four
subsets to compute sensitivity speci1047297city and accuracy [18] The
sensitivity speci1047297city and accuracy are de1047297ned as
Sensitivity frac14 ethTP =ethTP thornFN THORNTHORN 100 eth9THORN
Specificity frac14 ethTN =ethTN thornFP THORNTHORN 100 eth10THORN
Accuracy frac14 ethethTP thornTN THORN=ethTP thornTN thornFP thorn FN THORNTHORN 100 eth11THORN
where TP and TN represent the total number of detected true
positive events and true negative events respectively The FP and
FN represent false positive and false negative respectively
The technique in [18] claims that the second IMF provides
better classi1047297cation results while they use frequency and ampli-
tude modulation bandwidth features for the dataset [19] We
conduct experiments using 1047297rst four IMFs however we provide
results using only the 1047297rst IMF presented in Table 2 as it carries the
Table 2
Sensitivity speci1047297city and accuracy for different features of ictal and interictal EEG signals from Frontal and Temporal lobes
Classi1047297er Brain location Criteria Existing method Proposed method
Kernel Lobe EMD[18] DCT[17] Q4WHF_SVD Q4WHF_DCT FIMF_EMD Q4WHF_DCT ndashHF_DWT Combined
RBF Frontal SEN 2941 870 6666 100 9583 7678 4914
SPE 8051 7973 8220 8130 8186 8337 8382
ACC 7790 7810 8150 8160 8220 8300 7980
Temporal SEN 6667 400 100 9824 9710 9747 9647
SPE 8028 9863 8351 8473 8571 8665 8710
ACC 8020 7970 8420 8550 8650 8750 8790
Average SEN 4804 635 8333 9912 9647 8713 7281
SPE 8040 8918 8286 8302 8379 8501 8546
ACC 7905 7890 8285 8355 8435 8525 8385
Morlet Frontal SEN 3382 000 6383 8889 9333 6000 2442
SPE 8101 7978 8216 8126 8112 8396 8042
ACC 7780 789 8130 8140 8130 8180 7560
Temporal SEN 8462 200 100 9841 9012 9368 9468
SPE 8354 9900 8430 8527 8618 8632 8775
ACC 8360 7960 8510 8610 8650 8690 8840
Average SEN 5922 100 8192 9365 9173 7684 5955
SPE 8228 8939 8323 8327 8365 8514 8409
ACC 8070 7925 8320 8375 8390 8435 8200
Linear Frontal SEN 2895 2728 8095 100 100 100 8750
SPE 8035 8061 8131 8089 8155 8081 8223
ACC 7840 766 8130 8110 8190 8100 8240
Temporal SEN 8421 600 9800 9722 9811 9800 9836
SPE 8528 9613 8411 8288 8437 8412 8509
ACC 8520 7810 8480 8340 8510 8480 8590
Average SEN 5658 1664 8948 9861 9906 9900 9293
SPE 8282 8837 8271 8189 8296 8247 8366
ACC 8180 7735 8305 8225 8350 8290 8415
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 197
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 911
best result compared to other IMFs For the RBF kernel function
the proposed methods outperform the state-of-the-art method
consistently and the average sensitivity using the proposed meth-
ods such as Quartile four weighted high frequency component using
SVD (Q4WHF_SVD) Quartile four weighted high frequency compo-
nent using DCT (Q4WHF_DCT) First IMF using EMD (FIMF_EMD)
and Quartile four weighted high frequency component using DCT and
high frequency from DWT (Q4WHF_DCTndashHF_DWT) are 8333
9912 9647 and 8713 for RBF kernel respectively whereas
the average sensitivity using the state-of-art-methods [17] and
[18] are 635 and 4804 respectively Using Morlet kernel the
average sensitivity of the proposed methods is higher compared to
the existing methods [17] and [18] Moreover we justify our result
applying linear kernel Table 2 also presents that in most of the
cases the proposed methods with linear kernel provide better
results compared to the proposed methods with Morlet and RBF
kernels in terms of sensitivity Table 2 also presents that the
proposed Q4WHF_DCT method is the best among all proposed
methods in terms of sensitivity It is also interesting to note that
the accuracy of the proposed Q4WHF_DCTndashHF_DWT method is
reasonably consistence in terms of three criteria for RBF and
Morlet kernel According to Table 2 all proposed methods always
outperform the existing methods [1718] in terms of sensitivity (i
e seizure and ictal detection rate) however in some cases the
existing methods [1718] outperform one or more proposed
methods in terms of speci1047297city (interictal detection rate) For
seizure detection sensitivity is more important compared to the
speci1047297city detection as the cost of failure detecting ictal is heavier
compared to interictal In overall classi1047297cation the proposed
methods are also better compared to the existing methods We
combine all features to classify ictal and interictal EEG signalsHowever combining is not good because LS-SVM may not prop-
erly combine all available features to make a good prediction
model The average sensitivity of the proposed methods is 9136
(average result of all kernel and lobe) whereas the sensitivities of
the state-of-the-arts methods are 5461 [18] and 799 [17]
respectively
The performance of the LS-SVM is evaluated by (ROC) plot
shown in Fig 7(a) ROC illustrates the performance of a binary
classi1047297er system where it is created by plotting the fraction of true
positives from the positives ie true positive rate (TPR) vs the
fraction of false positives from negatives ie false positive rate
(FPR) with various threshold settings TPR is known as sensitivity
and FPR is one minus the speci1047297city or true negative rate Fig 7(a)
shows that the proposed four methods are carrying good
classi1047297cation results than that of the state-of-the-art method
[18] using dataset from Frontal lobe of the third subset of training
and testing EEG signals
Fig 7(b) shows the distribution of normalized feature values
(ie energy amp entropy in the Q4WHF_DCT method STD in the
Q4WHF_DCTndashHF_DWT method STD amp energy in the Q4WHF_SVD
method and STD in the FIMF EMD method) of ictal and interictal
from the Frontal lobe dataset [22] The 1047297gure shows that the range
of feature values for ictal is different from that of the interictal
signals This phenomenon is the other evidence why we use those
features to classify ictal and interictal EEG signals in the proposed
methods
First column of Fig 8 shows classi1047297cation comparison using the
proposed methods (ie FIMF EMD and Q4WHF_DCTndashHF_DWT)
and the state-of-the-art method [18] Fig 8(b) and (c) shows the
LS-SVM classi1047297cation results using the proposed FIMF EMD and
Q4WHF_DCTndashHF_DWT against the state-of-the-art method (ie
EMD Fig 8(a)) We generate 1047297rst column of Fig 8 from Frontal lobe
for the second subset of training and testing and obtain the
classi1047297cation accuracy 9120 for the Q4WHF_DCTndashHF_DWT
method 8720 for the FIMF_EMD method and 8080 for the
EMD method It can be concluded from these 1047297gures that the
proposed method based on the DCTndashDWT and the IMF outperform
the state-of-the-art method (ie EMD) The same classi1047297cation
comparisons are also provided in second column of Fig 8 using the
state-of-the-art method based on the EMD and the proposed
method based on the SVD and DCT respectively We consider
Temporal lobe for the fourth subset of training and testing obtain
classi1047297cation accuracy 9120 for Q4WHF_SVD 8960 for
Q4WHF_DCT and 8520 for EMD [18]Thus it is concluded that
the proposed methods based on DCT and SVD also outperform theEMD method [18]
5 Conclusion
Unlike seizure and non-seizure EEG signals classi1047297cation of
ictal and interictal EEG signals is a challenging task due to non
abrupt phenomena (ie visually non-separable) between ictal and
interictal EEG signals Moreover the features of ictal and interictal
signals are not consistence in different brain locations patient
types of seizure and hospital setting during epileptic period of a
patient Thus the classi1047297cation performance for ictal and interictal
EEG signals using the existing features extraction and classi1047297cation
techniques is not at expected level Our initial observation
Fig 7 (a) The receiver operating characteristics (ROC) curves of the third subset of training and testing EEG signals by three proposed methods ( Q4WHF_SVD FIMF_EMD
and Q4WHF_DCTndashHF_DWT) against the EMD method using LS-SVM with RBF kernel from Frontal lobe and (b) Normalized distribution of feature values extracted from ictal
and interictal EEG signals using DCT DWT SVD and IMF based method where 200 samples are taken from each of ictal and interictal categories in Frontal lobe
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200198
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1011
indicates that high frequency components of EEG signals have
distinguished characteristics To improve the accuracy in this
paper we propose novel feature extractions strategies from
the high frequency components based on a number of estab-
lished transformation and decompositions The average sensi-
tivity (which is the measure for seizure) of the proposed
methods is 9136 (average result of all kernel and lobe)
whereas the sensitivity of the state-of-the-arts methods are
5461 [18] and 799 [17] respectively The experimental
results show that the proposed methods outperform the state-
of-the-art methods in terms of accuracy sensitivity and speci-
1047297city for the ictal and interictal EEG signals with greater
consistence from the large benchmark dataset in two different
brain locations
Fig 8 First column represents the classi1047297cation of ictal and interictal EEG signals from Frontal lobe for the second subset of training and testing using (a) the state-of-the-art
method against (b) the proposed IMF and (c) the DCT-DWT method Second column represents the classi1047297cation of ictal and interictal EEG signals from Temporal lobe for the
fourth subset of training and testing using (d) the state-of-the-art method against (e) the proposed SVD and (f) the DCT methods
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 199
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1111
References
[1] Neurology Applied How Science is Bringing Music Instruction Back toExpressive Development URL langhttpwwwthomasjwestmusiccomneurolo
gyappliedhtmrang (accessed 270812)[2] EEG Signal The science behind Bo-Tau URL langhttpwwwbo-taucoukindex
phpaction=the_science_behind_bo_tau rang (accessed 270812)[3] Epilepsy-SeizuresURL langhttpwwwhealthcommunitiescomepilepsy-sei
zuresabout-seizuresshtml rang (accessed 270812)[4] M Soleymani T Pun M Pantic Multi-modal emotion recognition in response
to videos IEEE Trans Affect Comput 3 (2) (2012) 211ndash223[5] S Scholler S Bosse MS Treder B Blankertz G Curio K-R Muumlller
T Wiegand Toward a direct measure of video quality perception using EEGIEEE Trans Image Process 21 (5) (2012) 2619ndash2629
[6] W Di C Zhihua F Ruifang L Guangyu and L Tian Study on human brainafter consuming alcohol based on EEG signal in Proceedings of the Interna-tional Conference on Computer Science and Information Technology (ICCSIT)5 406ndash409 2010
[7] E Estrada H Nazeran F Ebrahimi and M Mikaeili EEG signal features forcomputer-aided sleep stage detection in Proceedings of the IEEE EMBSConference on Neural Engineering 669ndash672 2009
[8] ZM Hana1047297ah KFM Yunos ZH Murat MN Taib S Lias EEG brainwavepattern for smoking behaviour after horizontal rotation treatment in Pro-ceedings of the IEEE Student Conference on Research and Development 2009
[9] ZH Murat R Shilawani SA Kadir RM Isa and MN Taib The effects of
mobile phone usage on human braingwave using EEG in Proceedings of the13th International Conference on Modelling and Simulation 2011
[10] K Lehnertz RG Andrzejak J Arnhold T Kreuz F Mormann C Rieke
G Widman CE Elger Nonliner EEG analysis in Epilepsy its possible use forinterictal focus localization seizure ancipation and prevention J ClinNeurophsysiol 18 (3) (2001) 209ndash222
[11] R Panda PS Khobragade PD Jambhule SN Jengthe PR Pal and TK GandhiClassi1047297cation of EEG signal using wavelet transform and support vectormachine for epileptic seizure diction in Proceedings of the InternationalConference on Systems in Medicine and Biology (ICSMB) 405ndash408 2010
[12] SG Dastidar H Adeli N Dadmehr Mixed-band wavelet chaos-neural net-
work methodology for epilepsy and epileptic seizure detection IEEE TransBiomeical Eng 54 (9) (2007) 1545ndash1551
[13] H Ocak Optimal classi1047297cation of epileptic seizures in EEG using waveletanalysis and genetic algorithm Signal Process 88 (2008) 1858ndash1867
[14] K Polat S Gunes Classi1047297cation of epileptiform EEG using a hybrid systembased on decision tree classi1047297er and fast Fourier transform Appl Math
Comput 187 (2) (2007) 1017ndash1026[15] SF Liang HC Wang WL Chang Combination of EEG complexity and spectral
analysis for epilepsy diagnosis and seizure detection EURASIP J Adv SignalProcess 853434 (2010) 1ndash15 (Article ID)
[16] RB Pachori Discrimination between ictal and seizure-free EEG signals usingempirical mode decomposition Res Lett Signal Process 293056 (2008) 1ndash5(Article ID)
[17] D Birvinskas V Jusas I Martisius and R Damasevicius EEG dataset reductionand feature extraction using discrete cosine transform in Proceedings of theEuropean symposium on computer modeling and simulation (EMS) 199-2042012
[18] V Bajaj RB Pachori Classi1047297cation of seizure and non-seizure EEG signalsusing empirical mode decomposition IEEE Trans Inf Technol Biomed 16 (6)(2012) 1135ndash1142
[19] Small dataset Epilepsy dataset (text format) langhttpepileptologie-bonndecmsfront_contentphpidcat=193amplang=3ampchangelang=3rang (accessed 280412)
[20] J Harrison Fast and accurate bessel function computation in Proceedings of the19th IEEE international symposium on computer arithmetic 104ndash113
2009[21] X Zhang W Diao Z Cheng Wavelet transform and singular value decom-
position of EEG signal for pattern recognition of complicated hand activitiesLect Notes Comput Sci Digit Hum Model 4561 (2007) 294ndash303
[22] Large dataset EEG Data set from Epilepsy Center of the University Hospital of
Freiburg langhttpepilepsyuni-freiburgdefreiburg-seizure-prediction-projecteeg-databaserang (accessed 100612)
[23] KD Brabanter P Karsmakers F Ojeda C Alzate and JD Brabanter KPelckmans B De Moor J Vandewalle JAK Suykens LS-SVMlab ToolboxUsers Guide version 18 Katholieke Universiteit Leuven 2011
[24] S Sanei JA Chambers EEG Signal Processing John Wiley amp Sons Ltd UK
2007 (Centre of Digital Signal Processing Cardiff University)[25] S Abe Support Vector Machine for Pattern Classi1047297cation Second Edition
Springer London Dordrecht Heidelberg New York 2010
[26] H Wang and D Hu Comparison of SVM and LS-SVM for regression inProceedings of the international conference on neural networks and brain 1279ndash283 2005
[27] H Xing M Ha B Hu D Tian Linear feature-weighted support vectormachine Fuzzy Inf Eng 1 (3) (2009) 289ndash305
[28] A Subasi EEG signal classi1047297cation using wavelet feature extraction and amixture of expert model Expert Syst Appl 32 (4) (2007) 1084ndash1093(Elsevier)
[29] JR Williamson DW Bliss DW Browne JT Narayanan Seizure predictionusing EEG spatiotemporal correlation structure Epilepsy Behav 25 (2) (2012)230ndash238
[30] TP Jung S Makeig C Humphries et al Removing electroencephalographicsartifacts by blind source separation Psychophysiology 37 (2000) 163ndash178
[31] W De Clercq A Vergult B Vanrumste WV Paesschen SV Huffel Canonicalcorrelation analysis applied to remove muscle artifacts from the electroence-phalogram IEEE Trans Biomeical Eng 53 (12) (2006) 2583ndash2587
[32] J Ma P Tao S Bayram V Svetnik Muscle artifacts in multichannel EEGcharacteristics and reduction ClinNeurophysiol 123 (8) (2012) 1676ndash1686
[33] GA Worrell L Parish SD Cranstoun R Jonas G Baltuch B Litt High-frequency oscillations and seizure generation in neocortical epilepsy Brain127 (2004) 1496ndash1506
[34] M Paul W Lin CT Lau BndashS Lee A long term reference frame for hierarchicalb-picture based video coding IEEE Trans Circuits Syst Video Technol (2014)httpdxdoiorg101109TCSVT20142302555
[35] Z Gu W Lin CT Lau BndashS Lee and M Paul Two dimensional singular valuedecomposition (2D-SVD) based video coding in Proceedings of the IEEEconference on image processing (IEEE ICIP-10) 181ndash184 2010
Mohammad Zavid Parvez received BScEng (hons)Degree in Computer Science and Engineering fromAsian University of Bangladesh in 2003 and MScDegree in Electrical Engineering with emphasis onSignal Processing from Blekinge Institute of TechnologySweden in 2010 Parvezs research interests includebiomedical signal processing and machine learning
He has more than three year industry experience asa software developer Moreover he was involved sev-eral software projects in Denmark Currently he is aPhD student and teaching staff in Charles Sturt Uni-versity
Parvez is a Graduate student member of IEEE He hasreceived full funded scholarship from the Faculty He has published several journaland conference papers in the area of EEG signal analysis and cognitive radio
Manoranjan Paul received BScEng (hons) Degree inComputer Science and Engineering from BangladeshUniversity of Engineering and Technology (BUET) Ban-gladesh in 1997 and PhD Degree from Monash Uni-versity Australia in 2005 He was an Assistant Professorin Ahsanullah University of Science and Technology Hewas a Post-Doctoral Research Fellow in the Universityof New South Wales in 2005ndash2006 Monash Universityin 2006ndash2009 and Nanyang Technological Universityin 2009ndash2011 He has joined in the School of Comput-ing and Mathematics Charles Sturt University (CSU) at2011 Currently he is a Senior Lecturer and AssociateDirector of the Center for Research in Complex Systems
(CRiCS) in CSUDr Paul is a senior member of the IEEE and ACS He is in editorial board of three
international journals including EURASIP Journal on Advances in Signal Processing(JASP) Dr Paul has served as a guest editor of Journal of Multimedia and Journal of Computers for 1047297ve special issues He has been in Technical Program Committees of more than 25 international conferences Dr Paul obtained more than A$1Mcompetitive grant money including Australian Research Council (ARC) DiscoveryProject He has published more than 65 refereed publications including journalsand conferences He organized a special session on ldquoNew video coding technolo-giesrdquo in IEEE ISCAS 2010 He was a keynote speaker on ldquoVision friendly videocodingrdquo in IEEE ICCIT 2010 and tutorial speaker on ldquoMultiview Video Codingrdquo inDICTA 2013 His major research interests are in the 1047297elds of video coding computervision and EEG Signals analysis Dr Paul received Vice Chancellor ResearchExcellence Award in Faculty level at 2013
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200200
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 211
Existing methods [11ndash18] used small epilepsy dataset [19] (for
detailed information of the dataset see Section 31) The small
dataset with duration 236 s has seizure (ie ictal) and non-seizure
signals which can be distinguished by their visual phenomena
such as magnitude of amplitude and changing rate of frequency
(see in Fig 1(a)) For the non-seizure signal the amplitude is low
and the frequency is high while the nature of seizure signal is
totally opposite (see Fig 1(a)) Fig 1(b) and (c) shows the ictal and
interictal EEG signals from a large dataset [22] The 1047297gure
demonstrates the non-abrupt phenomena (ie not easily distin-
guishable between ictal and interictal based on amplitude and
frequency) of the ictal and interictal signal for both cases of Frontal
and Temporal lobes compared to the dataset in [19] Moreover
EEG signals from different locations exhibit different phenomenal
activities for an ictal and interictal period Thus the classi1047297cation
of ictal and interictal EEG signals is challenging compared to that
of seizure (ie ictal) and non-seizure EEG signals It also demon-
strates that the characteristics of EEG signals are varied fromdataset to dataset
Birvinskas et al [17] used DCT (discrete cosine transformation) to
reduce the EEG signals and then extracted features from the
reduced signals to classify EEG signals They achieved good
classi1047297cation accuracy by keeping low frequency DCT coef 1047297cient
for feature extractions Worrell et al [33] claimed that the seizure
onset frequencies are in the range of gamma frequency ie high
frequency band (typically between 30 to 100 Hz) They explicitly
mentioned that seizure onset frequencies are on the range of
frequency 74376 137 and 23776 28 Hz for Frontal and
Temporal lobes respectively They also observed that dominant
spectral peaks are observed in the high frequency spectral com-
ponents for the Frontal lobe and in the lower gamma frequency for
the Temporal lobe during the seizure period Other literature such
as Panda et al [11] Ocak et al [13] and Ghosh-Dastidar et al [12]
used other frequency bands such as delta theta alpha and beta
with gamma band to extract features for EEG signal classi1047297cation
From the above 1047297ndings and the characteristics of the different
datasets we 1047297nd that the frequency range of seizure onset can be
varied based on the brain location seizure type (generalpartial)
patient and hospital setting (eg capturing machine quality
artifacts etc) However we observe that the signal has larger
spikes ie high amplitude in the ictal (ie seizure) period
compared to the interictal period Thus based on the temporal
and spatial features of raw EEG signals and the above 1047297ndings of
the recent literatures our hypothesis is that the distinguishing
features between ictal and interictal signals would be in the high
frequency spectral area irrespective of the brain location patient
type of seizure and hospital setting
To verify our hypothesis we perform an experiment to 1047297nd the
gamma frequency of signals which provides the maximum ampli-
tude using ictal and interictal signals in different brain locationsThe results for 200 EEG signals from ictal and interictal are shown
in Fig 2 The 1047297gure shows that relatively high gamma frequencies
ie high frequencies in the signals provide the maximum ampli-
tude of ictal compared to the interictal signals for both Frontal and
Temporal lobes from different patients We use the maximum
amplitude criteria for the gamma band selection in this analysis as
our observation and Worrell et als observation indicate that
dominant spectral peaks for the seizure onset are in the high
frequency (ie gamma band) Thus the features extracted from
gamma frequencies ie the high frequencies should provide better
classi1047297cation accuracy between the ictal and interictal irrespective
of brain location patients type of seizure and hospital setting
Based on the above mentioned observation in this paper we
proposed generic seizure detection techniques which use different
Fig 1 Ictal and interictal EEG signals from different datasets and different brain locations (a) Seizure and non-seizure EEG signals from the small dataset [19] the 1047297rst row
and remaining four rows indicated the seizure and non-seizure signals respectively [19] (b) Ictal (patient one 7th datablock channel one) and interictal (patient one 95th
datablock channel one) of Frontal lobe signals [22] and (c) Ictal (patient 1215th datablock and channel one) and interictal (patient 12 35th datablock and channel one) of
Temporal lobe signals [22]
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 191
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 311
established transformations and decompositions to get different
statistical features from the high frequencies component of the
signals to classify the ictal and interictal signals using the least
square support vector machine (LS-SVM) [23] Although we use
a number of established transformation and decompositions we
use weighted high frequency coef 1047297cients to extract the features In
the experiment we investigate the strength of each transformation
and decomposition for the capability of classi1047297cation accuracy as
well as the combined strength through different feature extrac-
tions We use a large dataset [22] consists of raw ictal and
interictal signals (no explicit artifact removal technique is applied)
from different brain locations different seizures and patients to
verify the performance of the proposed methods against two
state-of-the-art methods The experimental results demonstrate
that the proposed generic methods outperform the existing
methods in terms of accuracy (ie overall) sensitivity (ie ictal
or positive) and speci1047297city (ie interictal or negative) with greater
consistence for the challenging ictal and interictal EEG signals
classi1047297cation
The contributions and novelties of the proposed methods are in
the capacity to (i) classify challenging ictal and interictal signals
irrespective of brain location patient type of seizure and hospital
setting with a certain range of artifact tolerance (ii) identify that
high frequency component of EEG signals has the distinguishing
features to classify ictal and interictal EEG signals (iii) formulate
a weighing vector for high frequency coef 1047297cients to provide
individual importance of different coef 1047297cients (iv) provide the
rationality and way of extracting high frequency components from
each transformationdecomposition technique (v) to analysis EEG
signals for seizure detection best of our knowledge we are the 1047297rst
one to use two dimensional SVD (vi) identify the distinguishing
features (provided pictorial comparison) for ictal and interictal
signals for classi1047297cation and (vii) provide comprehensive result
analysis using receiver operating characteristics (ROC) curve Whis-ker-box graphical view of classi1047297cation outcome and classi1047297ca-
tion performance in terms of sensitivity speci1047297city and accuracy
2 Existing methods for seizure detection
Existing feature extraction and classi1047297cation methods based on
wavelet [11ndash13] and Fourier transformation [14] were employed
for the detection of seizure in EEG signals Panda et al [11]
computed various features such as energy entropy and standard
deviation (STD) using discrete wavelet transformation (DWT) and
used support vector machine (SVM) as a classi1047297er Dastidar et al
[12] applied wavelet transformation to decompose the EEG signals
into different range of frequencies and then extracted three
features such as STD correlation dimension and the largest
Lyapunov exponent (quantifying the non-linear chaotic dynamics
of the signals) and applied different methods for classi1047297cation
Ocak [13] proposed a fourth level wavelet packet decomposition
method to decompose the normal and epileptic EEG epochs to
various frequency-bands and then used genetic algorithm to 1047297nd
optimal feature subsets which maximize the classi1047297cation perfor-
mance Polat et al [14] extracted feature using fast Fourier trans-
form (FFT) and then classi1047297ed using decision making classi1047297er The
techniques [111214] have used small dataset [19] (for detailed
information of the dataset see Section 21) Sample examples of
the dataset are shown in Fig 1(a) The classi1047297cation accuracies
were 9120 [11] 9670 [12] and 9872 [14] respectively Ocaks
[13] experimental dataset was of the same pattern with the
dataset in [19] and the classi1047297cation accuracy was 98
Other feature extraction and classi1047297cation methods [15ndash18]
based on empirical mode decomposition (EMD) principle component
analysis (PCA) and genetic algorithm (GA) have been proposed
Liang et al [15] extracted EEG features where the approximate
entropy(ApEn) analysis was evaluated for its ability to analyze
multiclass EEGs and the performance of ApEn analysis was
enhanced by incorporating the power of EEG sub-bands or auto-
regressive models PCA and GA were compared for their ability to
select features by applying various linear and nonlinear methods
for classi1047297cation The maximum accuracy was 9867 Pachori [16]
decomposed EEG signals into intrinsic mode function (IMF) using
EMD and then computed mean frequency for each IMF by Fourierndash
Bessel expansion [20] to differentiate seizure and non-seizure EEG
signals Zhang et al [21] proposed an approach for the pattern
recognition of four complicated hand activities such as grasping
a football a small bar a cylinder and a hard paper based on EEG
signal in which each piece of raw data sequence for EEG signal is
decomposed by wavelet transform After forming a matrix they
applied singular value decomposition (SVD) on the matrix to 1047297ndthe singular value feature vectors These features are used as input
to arti1047297cial neural network to discriminate four hand activities and
they got 8900 classi1047297cation accuracy
Recently Bajaj et al [18] extracted amplitude modulation and
frequency modulation bandwidth as features from IMF using EMD
Among the existing contemporary techniques Bajaj et als tech-
nique is the latest and the best in terms of performance They used
the LS-SVM [23] technique for the classi1047297cation of seizure and
non-seizure EEG signals using the small dataset [19] They
obtained 980 to 995 accuracy using radial basis function (RBF)
kernel and also obtained 995 to 100 accuracy using Morlet
kernel
We have observed that the existing methods [11ndash18] exhibit
almost perfect classi1047297cation accuracy for the small dataset of
Fig 2 Gamma frequencies which provide the maximum amplitude for ictal and interictal signals (from a standard dataset [22])
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200192
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 411
seizure and non-seizure EEG signals However their performance
is below the expected results using the large dataset [22] for ictal
and interictal EEG signal classi1047297cations For example when we
applied the Bajaj et als [18] feature extraction and classi1047297cation
method to differentiate ictal and interictal EEG signals for the 1047297rst
IMF in the large dataset from Epilepsy Center of the University
Hospital of Freiburg [22] the classi1047297cation accuracy is 7790 for
Frontal lobe and 8020 for Temporal lobe which is far below the
accuracy 9850 obtained using the small dataset [19] of seizureand non-seizure EEG signals
The paper is organized as follows the details of the proposed
four methods with dataset feature extractions classi1047297cation
techniques and computational time are described in Section 3
the detail experimental results are explained in Section 4 while
Section 5 concludes the paper
3 Proposed methods
In this paper we propose four methods based on the various new
features extracted from DCT DCTndashDWT SVD and IMF We use the
transformation and decompositions to separate high frequency
components and then apply innovative ways to extract different
statistical features When we extract features from the high fre-quency component (ie IMF and detail signals respective) using EMD
and DWT we take total length of the high frequency components
and we do not use any weighting for them as all the coef 1047297cients
represent the high frequency components As SVD and DCT provide
coef 1047297cients of all frequencies we need to select a subset of the
coef 1047297cients which can represent high frequency components We
investigate the strength of each transformation and decomposition
for the capability of classi1047297cation accuracy as well as the combined
strength through different feature extractions The dataset individual
method their corresponding rationality and detailed methodology of
feature extractions are described in the following sub-sections
31 Dataset
Small Dataset [19] The existing seizure and non-seizure feature
extraction and classi1047297cation schemes [11ndash18] except [13] used
small dataset [19] The dataset consists of 1047297ve subsets each subset
contains 100 single-channels recoding and each recoding has
236 s in duration captured by the international 10ndash20 electrode
placement scheme ie 32 electrodes [14] Among them two
subsets are taken from healthy volunteers two subsets are taken
from seizure free intervals and another subset is also taken during
seizure period [18] (see sample examples in Fig 1(a)) Ocak [13]
used different datasets with the same pattern of small dataset
where 300 were normal and 100 were epileptic patients
Large Dataset [22] Seizure and brain discharges are much more
complicated than what can be seen in the scalp Thus in the paper we
have used intracranial EEG dataset using grid strip and depth electrodes
rather than scalp EEG The dataset used in the paper was recorded at
Epilepsy Centre of the University Hospital of Freiburg Germany [22] The
database contains invasive EEG recordings of 21 patients suffering frommedically intractable focal epilepsy The data obtained by Neuro1047297le NT
digital video EEG system with 128 channels 256 Hz sampling rate and
16 bit analog-to-digital converter According to the dataset [22] epileptic
EEG signals can be classi1047297ed into (i) ictal (ie seizure) (ii) preictal (a
certain period before seizure onset) (iii) interictal (period between
seizures) and (iv) non-seizure (period without any seizure activity after
a certain time of interictal) Normally duration of ictal is varied from few
seconds to 2 min The ictal records contain epileptic seizures with at
least 50 min of preictal data preceding each seizure The median time
period between the last seizure and the interictal data set is 5 h and
18 min and the median time period between the interictal data set and
the 1047297rst following seizure is 9 h and 36 min [29] As suggested in the
dataset we have treated preictal and ictal period as an ictal signals in
our experiments We do not consider any non-seizure signals into the
interictal period as we carefully select the same duration of interictal
signal (with the ictal signal) from the above duration of interictal signal
32 Feature extraction using DCT
DCT is a transformation method for converting a time series
signal into basic frequency in such a way that the DCT coef 1047297cients
are arranged from low frequency to high frequency components
Low frequency components represent the coarse signals and high
frequency components represent the detail signals DCT is widely
used in the image processing and video coding areas to compress
image signals based on their frequencies [34] As the ictal and
interictal EEG signals have different amplitudes and frequencies
(not visually separable) thus the most distinguishable featuresshould be located in the high frequency components of DCT
coef 1047297cients see in Fig 3(a) As the EEG signal is non-linear and
non-stationary in nature thus for real time processing of EEG
signals DCT may not be correct to directly correspond to the
frequency analysis however if we segment the EEG signals in time
window and apply DCT on them to 1047297nd DCT coef 1047297cients we can
avoid non-linear and non-stationary nature of the signals In our
experiment we apply DCT on the benchmark dataset [22] and take
Fig 3 (a)High frequency DCT coef 1047297cients and their energy and entropy for the signals of Temporal lobe the 1047297rst two rows show the high frequency DCT coef 1047297cients (ie
25) of ictal and interictal signals while the last row shows the energy and entropy of using the high frequency DCT coef 1047297cients (ie 25) (b) High frequency DWT
coef 1047297cients and STDs of high frequency DCT and DWT coef 1047297cients for the EEG signals of Temporal lobe the 1047297rst two rows show that the high frequency DWT coef 1047297cients of
ictal and interictal signals while the last row shows the STD of the high frequency DCT and DWT coef 1047297cients of ictal and interictal EEG signals
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 193
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 511
the higher frequency component (ie the last quartile (Q4) ie last
25 of DCT coef 1047297cient) Then we extract statistical features such as
entropy and energy of the weighted higher frequency coef 1047297cients to
classify ictal and interictal signals
The extraction of the statistical features from the Q4 is not
straightforward We use heuristics to provide different weight to
the different coef 1047297cients and investigate the classi1047297cation accu-
racy Although we try different type of weighting approximations
the experimental results show that exponential weight performswell The weight w i for the i-th position coef 1047297cient is selected as
wi frac14 β e β xi where xi x j frac14 1
Number of Selected Coeff icients We also try
different β and test for the classi1047297cation results We 1047297nd that
β frac142 provides the best results for the dataset Fig 4 shows an
example of weights According to the proposed weight we provide
more weight to the relatively low frequency coef 1047297cients and less
weight to the relatively high frequency coef 1047297cients within the
higher frequency area of the overall signal as sometimes the high
frequency coef 1047297cients in the tail would be very small
We de1047297ne weighted entropy and energy as follows
Energy frac14 sumn
i frac14 1
W i X ietht THORN
2and Entropy frac14 sum
n
i frac14 1
W iY 2i etht THORNlnethY 2i etht THORNTHORN eth1THORN
where n is the length of signal X (t ) is the higher frequency DCTcoef 1047297cients and Y (t ) is the normalized value of X (t )
The representation of ictal and interictal data for different higher
frequency DCT coef 1047297cients and their entropy and energy are given in
Fig 3(a) It can be observed that the changing rate of frequencies is
higher for interictal signal than ictal signal from Temporal lobe It is
also interesting to observe that the amplitude of the higher frequency
DCT coef 1047297cients for ictal signals is larger compared to that of interictal
signals Note that although the original ictal and interictal EEG signals
are not visually separable the high frequency DCT coef 1047297cients of ictal
and interictal signals are visually separable The third row of Fig 3
(a) also demonstrates that the weighted energy and entropy extracted
from the higher frequency DCT components have different values for
ictal and interictal signals Thus weighted entropy and energy would
be effective features for ictal and interictal signals classi1047297
cation
33 Feature extraction using DCT ndashDWT
DWT is a transformation and widely used in the image and signal
processing research to decompose the imagesignals into different
frequency and bands DWT can be speci1047297ed by low pass (LP) and high
pass (HP) 1047297lters named as standard quadrature mirror 1047297lters [28] The
cut-off frequency of these 1047297lters is one-fourth of the sampling
frequency The bandwidth of the 1047297lter outputs is half the bandwidth
of the original signal which generates downsampled output signals
without losing any information according to the Nyquist theorem [13]
The down-sampled signals from LP and HP 1047297lters refer to 1047297rst-level
approximation and details of the original signal respectively [24] DWT
is sometimes more effective compared to DCT to separate signals into
low and high frequency components when the signal is longer Thus
we also investigate the effectiveness of the DWT to extract some other
statistical features (such as STD) to classify ictal and interictal signals
In our second approach we apply single-level one dimensional DWT
on the EEG signal to 1047297nd the high frequency components from the
signal Then we calculate the STD from the detail frequency (ie highfrequency) components of the DWT coef 1047297cients and use as a feature
We use the whole high frequency component for feature extraction as
DWT separates the signals into low and high frequency components
Another feature is extracted using the STD of the weighted high
frequency DCT coef 1047297cients (similar to Fig 4) Two STD features from
DCT and DWT higher frequency components are then used for the
classi1047297cation purpose Note that we use different primary functions of
the wavelet transformation including Daubechies (db) series We 1047297nd
that db16 is the best for our experiment
In the experiment we take only the high frequency DWT compo-
nents The 1047297rst two rows in Fig 3(b) show the higher change of
frequency and lower amplitude for interictal signal compared to ictal
signal from Temporal lobe The third row of Fig 3(b) represents the
STDs generated from the high frequency DCT coef 1047297cients (weighted)
and DWT for ictal which are higher than that of interictal signal Thus
the STDs extracted from the high frequency DCT (weighted) and DWT
coef 1047297cients can be good features for classi1047297cation
34 Feature extraction using SVD
SVD decomposes two dimensional signals into three matrixes in
such a way that the middle matrix (ie singular value) has unique
properties to separate signals The middle matrix containing the
square root of eigenvalues and diagonalized contains much less non-
zero coef 1047297cients compared with the coef 1047297cient matrices of other
transformations [35] The values of diagonal matrix are with des-
cending order from top-left to bottom-right As the original ictal and
interictal signals are almost of the same nature we do not expect asigni1047297cant different values of other two matrixes after SVD however
we expect some distinguishable values on the extreme part (ie
bottom-right) of diagonal matrix Thus any feature extracted from
the bottom-right singular values could be used to classify ictal and
interictal signals To the best of our knowledge we are the 1047297rst to use
two dimensional SVD to analyze EEG signals for seizure detection
To extract the features using SVD we reshape the EEG signals
into a two dimensional matrix The column vector of EEG signal is
rearranged to square matrix for the computation of singular value
components using SVD From the singular values we take the
weighted Q4 (weight is similar to Fig 4 but the number of selected
coef 1047297cients are different) of the non-zero diagonal values to
calculate the STDs and energy and treat as the two features to
classify ictal and interictal signals SVD allows transforming any
given matrix AAC mxn to diagonal form using unitary matrices ie
A frac14 U Σ V H where Amn is a matrix U mm and V nnare orthogonal
matrix and Σ mn matrix with non-negative diagonal entries
The columns of U are called left singular vectors and the rows
of V H are called right singular vectors If A is a rectangular m n
matrix of rank K then U will be m m and Σ will be
Σ frac14
λ1 0 ⋯ 0
0 λ2 ⋯ 0
⋮ ⋮ ⋱ ⋮
0 0 ⋯ λn
⋮ ⋮ ⋮
0 0 ⋯ 0
2666666664
3777777775
eth2THORN
where λ1Z λ 2Zhellip λn-1 Z λnFig 4 Weight for different coef 1047297cients of selected high frequency coef 1047297cients
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200194
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 611
The data distribution of non-zero singular values of the ictal
and interictal signals from the Frontal lobe is shown in Fig 5(a)
The top part of the 1047297gure clearly shows that the Q4 singular values
of interictal signals are larger than that of the ictal signals The
bottom part of Fig 5(a) shows the STD and energy extracted from
the weighted Q4 of non-zero singular values from ictal and
interictal signals respectively The 1047297gure exposes that the STDs
extracted from the ictal and interictal signals have distinguishable
values thus they can be used for ictal and interictal signal
classi1047297cation
35 Feature extraction of IMF using EMD
The main strength of EMD is its ability to analyze nonlinear and
non-stationary signals like EEG signals When we apply EMD on
a signal the signal is decomposed into a 1047297nite number of signals
into IMFs which are ordered from higher frequency component to
lower frequency component The number of IMFs in a signal
depends on the local characteristics of the signal rather than
pre-de1047297ned number of IMFs This feature makes each IMF crucial
to identify some characteristics of the original signal Since the
IMFs are the representations of different frequenciesamplitudes
of the original signal and the distinguishing features between ictal
and interictal EEG signals lie on the frequencyamplitude our
hypothesis is that any feature extracted from IMFs will be an
effective feature to classify ictal and interictal successfully We also
observe that the ictal and interictal signals are deferred in higherorder frequency components thus our other hypothesis is that
features extracted from the 1047297rst few IMFs should be enough to
classify ictal and interictal signals successfully Through the
decomposition we keep only few mode functions of the original
signals for the 1047297nal feature extraction however we are able to
keep the important mode functions for the classi1047297cation purpose
Experimental results reveal that our both hypothesis are correct as
extracted features from the 1047297rst IMF successfully classify ictal and
interictal signals (see Section 4) Thus IMFs can be used in
classi1047297cation applications where the original signals are non-
linear and non-stationary and distinguishing features exist on
the different frequenciesamplitudes
Each IMF satis1047297es two following conditions (i) the number of
extrema and the number of zero crossing are identical or differ at
most by one and (ii) the mean value between the upper and the
lower envelope is equal to zero at any time
The EMD algorithm can be summarized as follows
1 Extract the extrema (minima and maxima) of the signal xetht THORN
2 Interpolate between minima and maxima to obtain
ℓminetht THORN and ℓmaxetht THORN
3 Calculate local mean metht THORN frac14 ℓminetht THORN thornℓmaxetht THORNfrac12 =2
4 Extract the detail detht THORN frac14 xetht THORN metht THORN
5 Check detht THORN is an IMF according to the conditions which are
mentioned above If yes repeat the procedure from the step1 on the residual signal r etht THORN frac14 xetht THORN metht THORN If no replace xetht THORN with
detht THORN and repeat the procedure from step 1
The pictorial scenario of generating IMFs is shown in Fig 6 The
signal xetht THORNcan be represented as a combination of IMFs and
residual component
xetht THORN frac14 sumL
i frac14 1
dietht THORN thornr Letht THORN eth3THORN
where di and r L are the i-th IMF and L-th is residual signal
We observe that features extracted from the 1047297rst two IMFs
provide the best classi1047297cation results To reduce computational
time we consider only the 1047297rst IMF from the EMD and then
calculate STD of IMF to capture the variation of the signal Byignoring other IMFs we do not sacri1047297ce any signi1047297cant classi1047297cation
accuracy The strength of a signal is distributed in the frequency
domain relative to the strengths of other ambient signals is central
to the design of any 1047297lter intended to extract or suppress the signal
To 1047297nd the strength of the signal we apply power spectral density
(PSD) on the 1047297rst IMF Then determine the STD feature from
normalized PSD The calculation of PSD is as follows
P ethwiTHORN frac14 1
N DethwiTHORN 2
eth4THORN
where DethwiTHORNis the discrete Fourier transform of IMFdietht THORN Normalized
PSD can be de1047297ned as
pi frac14 P ethwiTHORN=sum
i
P ethwiTHORN eth5THORN
Fig 5 (a) Non-zero singular value of SVD for Frontal lobe The 1047297rst row shows that the non-zero singular value distribution for ictal and interictal signals from Frontal lobe
and the last row shows the STD and energy of the last 25 non-zero singular values of ictal and interictal EEG signals (b) The 1047297rst IMFs of ictal and interictal signals with the
extracted features the top two rows represent the 1047297rst IMF from the ictal and interictal signals of Temporal lobe bottom row represents STD of the 1047297rst IMF and STD of the
normalized PSD (power spectral density) of the same IMF of ictal and interictal EEG signals
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 195
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 711
In the experiment we extract two features such as STD of the
1047297rst IMF and STD of normalized PSD of the 1047297rst IMF From the 1047297rst
two rows in Fig 5(b) it can be observed that the 1047297rst IMF of ictal
signal carries the higher amplitude compared to the interictal
signal
The third row of the 1047297gure represents the STD of the 1047297rst IMF
and STD of PSD of the same IMF from using ictal and interictal
signals The STDs for both cases of interictal signal are signi1047297cantly
larger than that of ictal signal Thus the two new features
extracted by EMD method can be good features for classi1047297cation
Table 1 presents the summary of all corresponding features use
for ictal and interictal classi1047297cation by each method
36 Classi 1047297er
We extract a number of features from different transforma-
tionsdecompositions To classify ictal and interictal signals we
need to use a classi1047297er The goal of a classi1047297er is to 1047297nd patients
states such as ictal (class 1) and interictal (class 2) using machine
learning approaches with cross-validation The challenge is to 1047297nd
the mapping that generalizes from training sets to unseen test
sets For the cross-validation data are partitioned into training set
and test set In our experiment we use 4-fold cross validation
where each dataset was randomly divided into 4 splits where
three of them were used for training and the fourth one used for
testing
To classify the ictal and interictal signals we extract various
features from DCT DWT SVD and IMF of EMD For classi1047297cation we
use SVM-based classi1047297er as the SVM is one of the best classi1047297ers in
the EEG signal analysis SVM is a potential methodology for solving
problem in linear and nonlinear classi1047297cation function estimation
and kernel based learning methods [25] It can minimize the
operational error and maximize the margin hyperplane as a result
it will maximize the classi1047297cation performance [25] LS-SVM is the
extended version of SVM and it is closely related to regularization
networks and Gaussian process and it also has primal-dual inter-
pretations [23] The major drawback of SVM is in its higher
computational burden of the constrained optimization program-
ming However LS-SVM can solve this problem [26]
The classi1047297er is a LS-SVM which learns nonlinear mapping
from the training set features xifrac141hellipnT where nT is the number
of training features into the patients state ictal ( thorn1) and interictal( 1) For unbiased classi1047297cation results [27] we divide the whole
1000 trials into four subsets in each subset we randomly select
75 samples for the validation training set and the remaining 25
for the validation testing set Let f yigi frac14 C 12 designate the LS-SVM
validation test outputs mapping to class 1 or class 2 The equation
of LS-SVM can be de1047297ned in [18] as
f eth xTHORN frac14 sign sumN
i 1
α i yiK eth x xiTHORN thornc
eth6THORN
where K ( x xi) is a kernel function α i are the Lagrange multipliers c
is the bias term xi is the training input and yi is the training
output pairs Linear RBF and Morlet kernel are used in our
experiments and RBF and Morlet functions can be de1047297ned in [18]
Fig 6 Extracted IMFs by EMD from Temporal lobe (Patient 12 channel one 15th datablock for ictal and 35th datablock for interictal)
Table 1Extracted features of proposed and existing methods
Proposed methods Feature-1 Feature-2
Q4WHF_SVD (Quartile four weighted high frequency component using SVD ) STD on 25 weighted high frequency
singular values
Energy on 25 weighted high frequency
singular values
Q4WHF_DCT (Quartile four weighted high frequency component using DCT ) Energy on 25 weighted high
frequency DCT coef 1047297cients
Entropy on 25 weighted high frequency DCT
coef 1047297cients
FIMF_EMD (First IMF using EMD ) STD on the 1047297rst IMF (ie high
frequency components)
STD on normalized PSD
Q4WHF_DCT ndash HF_DWT (Quartile four weighted high frequency component
using DCT and high frequency from DWT)
STD on 25 weighted high frequency
DCT coef 1047297cients
STD on detail signals (ie high frequency
components) after apply DWT
Combined Combined all above features
Existing methods Feature-1 Feature-2
EMD[18] Amplitude modulation bandwidth Frequency modulation bandwidth
DCT [17] Features from 6 of quantized DCT coef 1047297cients at lower frequency
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200196
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 811
as
keth x xiTHORN frac14 expeth x xi2=2s
2THORN eth7THORN
where s controls the width of the RBF function
keth x xiTHORN frac14 prodd
k frac14 1
cos ω0etheth xk xki THORN=aTHORN
h iexpeth xk xk
i =2a2THORN eth8THORN
where d is the dimension a is the 1047298exible coef 1047297cient and xki is the
k-th component of i-th training data Matlab codes for LS-SVMclassi1047297er are available at httpwwwesatkuleuvenacbesista
lssvmlab
37 Computational time
Feature extractions using IMF through EMD take longer time
due to two-fold iterations to 1047297nd all IMFs In the 1047297rst iteration
EMD decomposes a signal and checks whether the signal makes an
IMF and in the second iteration it subtracts the current IMF from
the signal and continues for the next IMF Even if we stop the
iteration for the 1047297rst IMF the EMD-based technique takes a full
1047297rst iteration for calculating the IMF The experimental data
reveals that the proposed technique using DCT DWT and SVD
takes less computational time compared to the EMD-based tech-nique as DCT DWT and SVD transformation is faster compared to
the EMD decomposition Thus the proposed methods (except
FIMF_EMD method ie First IMF using EMD) take insigni1047297cant
time compared to the existing method [18]
4 Experimental results
Sometimes EEG signals have line noise and other kinds of
artifacts due to muscle and body movements Notch 1047297lter and
independent component analysis (ICA)-based method are recom-
mended to remove the line noise and artifacts respectively for
suf 1047297cient amount of EEG signals [30ndash32] However in this experi-
ment we do not use any explicit noiseartifacts removal technique
to see the tolerance of the proposed method against noise
artifacts
We propose four methods and the experimental results of the
proposed methods are compared with existing methods [1718]
and corresponding features are presented in Table 1Our ultimate
target is to classify ictal and interictal EEG signals using LS-SVM
with different kernels such as RBF and Morlet where the values of regularization parameter and kernel parameter are 10 and 6
respectively for both kernels We also use linear kernel In this
paper four methods are proposed based on the new features of
ictal and interictal EEG signals captured from Frontal and Temporal
lobes using DCT DWT SVD and IMF of EMD to see the strength of
individual transformationdecomposition techniques for seizure
detection Then we also test the classi1047297cation accuracy by combin-
ing all features In our experiment we use 200 signals from ictal
and 800 signals from interictal EEG signals After training and
testing of all features we calculate average value from the four
subsets to compute sensitivity speci1047297city and accuracy [18] The
sensitivity speci1047297city and accuracy are de1047297ned as
Sensitivity frac14 ethTP =ethTP thornFN THORNTHORN 100 eth9THORN
Specificity frac14 ethTN =ethTN thornFP THORNTHORN 100 eth10THORN
Accuracy frac14 ethethTP thornTN THORN=ethTP thornTN thornFP thorn FN THORNTHORN 100 eth11THORN
where TP and TN represent the total number of detected true
positive events and true negative events respectively The FP and
FN represent false positive and false negative respectively
The technique in [18] claims that the second IMF provides
better classi1047297cation results while they use frequency and ampli-
tude modulation bandwidth features for the dataset [19] We
conduct experiments using 1047297rst four IMFs however we provide
results using only the 1047297rst IMF presented in Table 2 as it carries the
Table 2
Sensitivity speci1047297city and accuracy for different features of ictal and interictal EEG signals from Frontal and Temporal lobes
Classi1047297er Brain location Criteria Existing method Proposed method
Kernel Lobe EMD[18] DCT[17] Q4WHF_SVD Q4WHF_DCT FIMF_EMD Q4WHF_DCT ndashHF_DWT Combined
RBF Frontal SEN 2941 870 6666 100 9583 7678 4914
SPE 8051 7973 8220 8130 8186 8337 8382
ACC 7790 7810 8150 8160 8220 8300 7980
Temporal SEN 6667 400 100 9824 9710 9747 9647
SPE 8028 9863 8351 8473 8571 8665 8710
ACC 8020 7970 8420 8550 8650 8750 8790
Average SEN 4804 635 8333 9912 9647 8713 7281
SPE 8040 8918 8286 8302 8379 8501 8546
ACC 7905 7890 8285 8355 8435 8525 8385
Morlet Frontal SEN 3382 000 6383 8889 9333 6000 2442
SPE 8101 7978 8216 8126 8112 8396 8042
ACC 7780 789 8130 8140 8130 8180 7560
Temporal SEN 8462 200 100 9841 9012 9368 9468
SPE 8354 9900 8430 8527 8618 8632 8775
ACC 8360 7960 8510 8610 8650 8690 8840
Average SEN 5922 100 8192 9365 9173 7684 5955
SPE 8228 8939 8323 8327 8365 8514 8409
ACC 8070 7925 8320 8375 8390 8435 8200
Linear Frontal SEN 2895 2728 8095 100 100 100 8750
SPE 8035 8061 8131 8089 8155 8081 8223
ACC 7840 766 8130 8110 8190 8100 8240
Temporal SEN 8421 600 9800 9722 9811 9800 9836
SPE 8528 9613 8411 8288 8437 8412 8509
ACC 8520 7810 8480 8340 8510 8480 8590
Average SEN 5658 1664 8948 9861 9906 9900 9293
SPE 8282 8837 8271 8189 8296 8247 8366
ACC 8180 7735 8305 8225 8350 8290 8415
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 197
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 911
best result compared to other IMFs For the RBF kernel function
the proposed methods outperform the state-of-the-art method
consistently and the average sensitivity using the proposed meth-
ods such as Quartile four weighted high frequency component using
SVD (Q4WHF_SVD) Quartile four weighted high frequency compo-
nent using DCT (Q4WHF_DCT) First IMF using EMD (FIMF_EMD)
and Quartile four weighted high frequency component using DCT and
high frequency from DWT (Q4WHF_DCTndashHF_DWT) are 8333
9912 9647 and 8713 for RBF kernel respectively whereas
the average sensitivity using the state-of-art-methods [17] and
[18] are 635 and 4804 respectively Using Morlet kernel the
average sensitivity of the proposed methods is higher compared to
the existing methods [17] and [18] Moreover we justify our result
applying linear kernel Table 2 also presents that in most of the
cases the proposed methods with linear kernel provide better
results compared to the proposed methods with Morlet and RBF
kernels in terms of sensitivity Table 2 also presents that the
proposed Q4WHF_DCT method is the best among all proposed
methods in terms of sensitivity It is also interesting to note that
the accuracy of the proposed Q4WHF_DCTndashHF_DWT method is
reasonably consistence in terms of three criteria for RBF and
Morlet kernel According to Table 2 all proposed methods always
outperform the existing methods [1718] in terms of sensitivity (i
e seizure and ictal detection rate) however in some cases the
existing methods [1718] outperform one or more proposed
methods in terms of speci1047297city (interictal detection rate) For
seizure detection sensitivity is more important compared to the
speci1047297city detection as the cost of failure detecting ictal is heavier
compared to interictal In overall classi1047297cation the proposed
methods are also better compared to the existing methods We
combine all features to classify ictal and interictal EEG signalsHowever combining is not good because LS-SVM may not prop-
erly combine all available features to make a good prediction
model The average sensitivity of the proposed methods is 9136
(average result of all kernel and lobe) whereas the sensitivities of
the state-of-the-arts methods are 5461 [18] and 799 [17]
respectively
The performance of the LS-SVM is evaluated by (ROC) plot
shown in Fig 7(a) ROC illustrates the performance of a binary
classi1047297er system where it is created by plotting the fraction of true
positives from the positives ie true positive rate (TPR) vs the
fraction of false positives from negatives ie false positive rate
(FPR) with various threshold settings TPR is known as sensitivity
and FPR is one minus the speci1047297city or true negative rate Fig 7(a)
shows that the proposed four methods are carrying good
classi1047297cation results than that of the state-of-the-art method
[18] using dataset from Frontal lobe of the third subset of training
and testing EEG signals
Fig 7(b) shows the distribution of normalized feature values
(ie energy amp entropy in the Q4WHF_DCT method STD in the
Q4WHF_DCTndashHF_DWT method STD amp energy in the Q4WHF_SVD
method and STD in the FIMF EMD method) of ictal and interictal
from the Frontal lobe dataset [22] The 1047297gure shows that the range
of feature values for ictal is different from that of the interictal
signals This phenomenon is the other evidence why we use those
features to classify ictal and interictal EEG signals in the proposed
methods
First column of Fig 8 shows classi1047297cation comparison using the
proposed methods (ie FIMF EMD and Q4WHF_DCTndashHF_DWT)
and the state-of-the-art method [18] Fig 8(b) and (c) shows the
LS-SVM classi1047297cation results using the proposed FIMF EMD and
Q4WHF_DCTndashHF_DWT against the state-of-the-art method (ie
EMD Fig 8(a)) We generate 1047297rst column of Fig 8 from Frontal lobe
for the second subset of training and testing and obtain the
classi1047297cation accuracy 9120 for the Q4WHF_DCTndashHF_DWT
method 8720 for the FIMF_EMD method and 8080 for the
EMD method It can be concluded from these 1047297gures that the
proposed method based on the DCTndashDWT and the IMF outperform
the state-of-the-art method (ie EMD) The same classi1047297cation
comparisons are also provided in second column of Fig 8 using the
state-of-the-art method based on the EMD and the proposed
method based on the SVD and DCT respectively We consider
Temporal lobe for the fourth subset of training and testing obtain
classi1047297cation accuracy 9120 for Q4WHF_SVD 8960 for
Q4WHF_DCT and 8520 for EMD [18]Thus it is concluded that
the proposed methods based on DCT and SVD also outperform theEMD method [18]
5 Conclusion
Unlike seizure and non-seizure EEG signals classi1047297cation of
ictal and interictal EEG signals is a challenging task due to non
abrupt phenomena (ie visually non-separable) between ictal and
interictal EEG signals Moreover the features of ictal and interictal
signals are not consistence in different brain locations patient
types of seizure and hospital setting during epileptic period of a
patient Thus the classi1047297cation performance for ictal and interictal
EEG signals using the existing features extraction and classi1047297cation
techniques is not at expected level Our initial observation
Fig 7 (a) The receiver operating characteristics (ROC) curves of the third subset of training and testing EEG signals by three proposed methods ( Q4WHF_SVD FIMF_EMD
and Q4WHF_DCTndashHF_DWT) against the EMD method using LS-SVM with RBF kernel from Frontal lobe and (b) Normalized distribution of feature values extracted from ictal
and interictal EEG signals using DCT DWT SVD and IMF based method where 200 samples are taken from each of ictal and interictal categories in Frontal lobe
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200198
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1011
indicates that high frequency components of EEG signals have
distinguished characteristics To improve the accuracy in this
paper we propose novel feature extractions strategies from
the high frequency components based on a number of estab-
lished transformation and decompositions The average sensi-
tivity (which is the measure for seizure) of the proposed
methods is 9136 (average result of all kernel and lobe)
whereas the sensitivity of the state-of-the-arts methods are
5461 [18] and 799 [17] respectively The experimental
results show that the proposed methods outperform the state-
of-the-art methods in terms of accuracy sensitivity and speci-
1047297city for the ictal and interictal EEG signals with greater
consistence from the large benchmark dataset in two different
brain locations
Fig 8 First column represents the classi1047297cation of ictal and interictal EEG signals from Frontal lobe for the second subset of training and testing using (a) the state-of-the-art
method against (b) the proposed IMF and (c) the DCT-DWT method Second column represents the classi1047297cation of ictal and interictal EEG signals from Temporal lobe for the
fourth subset of training and testing using (d) the state-of-the-art method against (e) the proposed SVD and (f) the DCT methods
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 199
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1111
References
[1] Neurology Applied How Science is Bringing Music Instruction Back toExpressive Development URL langhttpwwwthomasjwestmusiccomneurolo
gyappliedhtmrang (accessed 270812)[2] EEG Signal The science behind Bo-Tau URL langhttpwwwbo-taucoukindex
phpaction=the_science_behind_bo_tau rang (accessed 270812)[3] Epilepsy-SeizuresURL langhttpwwwhealthcommunitiescomepilepsy-sei
zuresabout-seizuresshtml rang (accessed 270812)[4] M Soleymani T Pun M Pantic Multi-modal emotion recognition in response
to videos IEEE Trans Affect Comput 3 (2) (2012) 211ndash223[5] S Scholler S Bosse MS Treder B Blankertz G Curio K-R Muumlller
T Wiegand Toward a direct measure of video quality perception using EEGIEEE Trans Image Process 21 (5) (2012) 2619ndash2629
[6] W Di C Zhihua F Ruifang L Guangyu and L Tian Study on human brainafter consuming alcohol based on EEG signal in Proceedings of the Interna-tional Conference on Computer Science and Information Technology (ICCSIT)5 406ndash409 2010
[7] E Estrada H Nazeran F Ebrahimi and M Mikaeili EEG signal features forcomputer-aided sleep stage detection in Proceedings of the IEEE EMBSConference on Neural Engineering 669ndash672 2009
[8] ZM Hana1047297ah KFM Yunos ZH Murat MN Taib S Lias EEG brainwavepattern for smoking behaviour after horizontal rotation treatment in Pro-ceedings of the IEEE Student Conference on Research and Development 2009
[9] ZH Murat R Shilawani SA Kadir RM Isa and MN Taib The effects of
mobile phone usage on human braingwave using EEG in Proceedings of the13th International Conference on Modelling and Simulation 2011
[10] K Lehnertz RG Andrzejak J Arnhold T Kreuz F Mormann C Rieke
G Widman CE Elger Nonliner EEG analysis in Epilepsy its possible use forinterictal focus localization seizure ancipation and prevention J ClinNeurophsysiol 18 (3) (2001) 209ndash222
[11] R Panda PS Khobragade PD Jambhule SN Jengthe PR Pal and TK GandhiClassi1047297cation of EEG signal using wavelet transform and support vectormachine for epileptic seizure diction in Proceedings of the InternationalConference on Systems in Medicine and Biology (ICSMB) 405ndash408 2010
[12] SG Dastidar H Adeli N Dadmehr Mixed-band wavelet chaos-neural net-
work methodology for epilepsy and epileptic seizure detection IEEE TransBiomeical Eng 54 (9) (2007) 1545ndash1551
[13] H Ocak Optimal classi1047297cation of epileptic seizures in EEG using waveletanalysis and genetic algorithm Signal Process 88 (2008) 1858ndash1867
[14] K Polat S Gunes Classi1047297cation of epileptiform EEG using a hybrid systembased on decision tree classi1047297er and fast Fourier transform Appl Math
Comput 187 (2) (2007) 1017ndash1026[15] SF Liang HC Wang WL Chang Combination of EEG complexity and spectral
analysis for epilepsy diagnosis and seizure detection EURASIP J Adv SignalProcess 853434 (2010) 1ndash15 (Article ID)
[16] RB Pachori Discrimination between ictal and seizure-free EEG signals usingempirical mode decomposition Res Lett Signal Process 293056 (2008) 1ndash5(Article ID)
[17] D Birvinskas V Jusas I Martisius and R Damasevicius EEG dataset reductionand feature extraction using discrete cosine transform in Proceedings of theEuropean symposium on computer modeling and simulation (EMS) 199-2042012
[18] V Bajaj RB Pachori Classi1047297cation of seizure and non-seizure EEG signalsusing empirical mode decomposition IEEE Trans Inf Technol Biomed 16 (6)(2012) 1135ndash1142
[19] Small dataset Epilepsy dataset (text format) langhttpepileptologie-bonndecmsfront_contentphpidcat=193amplang=3ampchangelang=3rang (accessed 280412)
[20] J Harrison Fast and accurate bessel function computation in Proceedings of the19th IEEE international symposium on computer arithmetic 104ndash113
2009[21] X Zhang W Diao Z Cheng Wavelet transform and singular value decom-
position of EEG signal for pattern recognition of complicated hand activitiesLect Notes Comput Sci Digit Hum Model 4561 (2007) 294ndash303
[22] Large dataset EEG Data set from Epilepsy Center of the University Hospital of
Freiburg langhttpepilepsyuni-freiburgdefreiburg-seizure-prediction-projecteeg-databaserang (accessed 100612)
[23] KD Brabanter P Karsmakers F Ojeda C Alzate and JD Brabanter KPelckmans B De Moor J Vandewalle JAK Suykens LS-SVMlab ToolboxUsers Guide version 18 Katholieke Universiteit Leuven 2011
[24] S Sanei JA Chambers EEG Signal Processing John Wiley amp Sons Ltd UK
2007 (Centre of Digital Signal Processing Cardiff University)[25] S Abe Support Vector Machine for Pattern Classi1047297cation Second Edition
Springer London Dordrecht Heidelberg New York 2010
[26] H Wang and D Hu Comparison of SVM and LS-SVM for regression inProceedings of the international conference on neural networks and brain 1279ndash283 2005
[27] H Xing M Ha B Hu D Tian Linear feature-weighted support vectormachine Fuzzy Inf Eng 1 (3) (2009) 289ndash305
[28] A Subasi EEG signal classi1047297cation using wavelet feature extraction and amixture of expert model Expert Syst Appl 32 (4) (2007) 1084ndash1093(Elsevier)
[29] JR Williamson DW Bliss DW Browne JT Narayanan Seizure predictionusing EEG spatiotemporal correlation structure Epilepsy Behav 25 (2) (2012)230ndash238
[30] TP Jung S Makeig C Humphries et al Removing electroencephalographicsartifacts by blind source separation Psychophysiology 37 (2000) 163ndash178
[31] W De Clercq A Vergult B Vanrumste WV Paesschen SV Huffel Canonicalcorrelation analysis applied to remove muscle artifacts from the electroence-phalogram IEEE Trans Biomeical Eng 53 (12) (2006) 2583ndash2587
[32] J Ma P Tao S Bayram V Svetnik Muscle artifacts in multichannel EEGcharacteristics and reduction ClinNeurophysiol 123 (8) (2012) 1676ndash1686
[33] GA Worrell L Parish SD Cranstoun R Jonas G Baltuch B Litt High-frequency oscillations and seizure generation in neocortical epilepsy Brain127 (2004) 1496ndash1506
[34] M Paul W Lin CT Lau BndashS Lee A long term reference frame for hierarchicalb-picture based video coding IEEE Trans Circuits Syst Video Technol (2014)httpdxdoiorg101109TCSVT20142302555
[35] Z Gu W Lin CT Lau BndashS Lee and M Paul Two dimensional singular valuedecomposition (2D-SVD) based video coding in Proceedings of the IEEEconference on image processing (IEEE ICIP-10) 181ndash184 2010
Mohammad Zavid Parvez received BScEng (hons)Degree in Computer Science and Engineering fromAsian University of Bangladesh in 2003 and MScDegree in Electrical Engineering with emphasis onSignal Processing from Blekinge Institute of TechnologySweden in 2010 Parvezs research interests includebiomedical signal processing and machine learning
He has more than three year industry experience asa software developer Moreover he was involved sev-eral software projects in Denmark Currently he is aPhD student and teaching staff in Charles Sturt Uni-versity
Parvez is a Graduate student member of IEEE He hasreceived full funded scholarship from the Faculty He has published several journaland conference papers in the area of EEG signal analysis and cognitive radio
Manoranjan Paul received BScEng (hons) Degree inComputer Science and Engineering from BangladeshUniversity of Engineering and Technology (BUET) Ban-gladesh in 1997 and PhD Degree from Monash Uni-versity Australia in 2005 He was an Assistant Professorin Ahsanullah University of Science and Technology Hewas a Post-Doctoral Research Fellow in the Universityof New South Wales in 2005ndash2006 Monash Universityin 2006ndash2009 and Nanyang Technological Universityin 2009ndash2011 He has joined in the School of Comput-ing and Mathematics Charles Sturt University (CSU) at2011 Currently he is a Senior Lecturer and AssociateDirector of the Center for Research in Complex Systems
(CRiCS) in CSUDr Paul is a senior member of the IEEE and ACS He is in editorial board of three
international journals including EURASIP Journal on Advances in Signal Processing(JASP) Dr Paul has served as a guest editor of Journal of Multimedia and Journal of Computers for 1047297ve special issues He has been in Technical Program Committees of more than 25 international conferences Dr Paul obtained more than A$1Mcompetitive grant money including Australian Research Council (ARC) DiscoveryProject He has published more than 65 refereed publications including journalsand conferences He organized a special session on ldquoNew video coding technolo-giesrdquo in IEEE ISCAS 2010 He was a keynote speaker on ldquoVision friendly videocodingrdquo in IEEE ICCIT 2010 and tutorial speaker on ldquoMultiview Video Codingrdquo inDICTA 2013 His major research interests are in the 1047297elds of video coding computervision and EEG Signals analysis Dr Paul received Vice Chancellor ResearchExcellence Award in Faculty level at 2013
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200200
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 311
established transformations and decompositions to get different
statistical features from the high frequencies component of the
signals to classify the ictal and interictal signals using the least
square support vector machine (LS-SVM) [23] Although we use
a number of established transformation and decompositions we
use weighted high frequency coef 1047297cients to extract the features In
the experiment we investigate the strength of each transformation
and decomposition for the capability of classi1047297cation accuracy as
well as the combined strength through different feature extrac-
tions We use a large dataset [22] consists of raw ictal and
interictal signals (no explicit artifact removal technique is applied)
from different brain locations different seizures and patients to
verify the performance of the proposed methods against two
state-of-the-art methods The experimental results demonstrate
that the proposed generic methods outperform the existing
methods in terms of accuracy (ie overall) sensitivity (ie ictal
or positive) and speci1047297city (ie interictal or negative) with greater
consistence for the challenging ictal and interictal EEG signals
classi1047297cation
The contributions and novelties of the proposed methods are in
the capacity to (i) classify challenging ictal and interictal signals
irrespective of brain location patient type of seizure and hospital
setting with a certain range of artifact tolerance (ii) identify that
high frequency component of EEG signals has the distinguishing
features to classify ictal and interictal EEG signals (iii) formulate
a weighing vector for high frequency coef 1047297cients to provide
individual importance of different coef 1047297cients (iv) provide the
rationality and way of extracting high frequency components from
each transformationdecomposition technique (v) to analysis EEG
signals for seizure detection best of our knowledge we are the 1047297rst
one to use two dimensional SVD (vi) identify the distinguishing
features (provided pictorial comparison) for ictal and interictal
signals for classi1047297cation and (vii) provide comprehensive result
analysis using receiver operating characteristics (ROC) curve Whis-ker-box graphical view of classi1047297cation outcome and classi1047297ca-
tion performance in terms of sensitivity speci1047297city and accuracy
2 Existing methods for seizure detection
Existing feature extraction and classi1047297cation methods based on
wavelet [11ndash13] and Fourier transformation [14] were employed
for the detection of seizure in EEG signals Panda et al [11]
computed various features such as energy entropy and standard
deviation (STD) using discrete wavelet transformation (DWT) and
used support vector machine (SVM) as a classi1047297er Dastidar et al
[12] applied wavelet transformation to decompose the EEG signals
into different range of frequencies and then extracted three
features such as STD correlation dimension and the largest
Lyapunov exponent (quantifying the non-linear chaotic dynamics
of the signals) and applied different methods for classi1047297cation
Ocak [13] proposed a fourth level wavelet packet decomposition
method to decompose the normal and epileptic EEG epochs to
various frequency-bands and then used genetic algorithm to 1047297nd
optimal feature subsets which maximize the classi1047297cation perfor-
mance Polat et al [14] extracted feature using fast Fourier trans-
form (FFT) and then classi1047297ed using decision making classi1047297er The
techniques [111214] have used small dataset [19] (for detailed
information of the dataset see Section 21) Sample examples of
the dataset are shown in Fig 1(a) The classi1047297cation accuracies
were 9120 [11] 9670 [12] and 9872 [14] respectively Ocaks
[13] experimental dataset was of the same pattern with the
dataset in [19] and the classi1047297cation accuracy was 98
Other feature extraction and classi1047297cation methods [15ndash18]
based on empirical mode decomposition (EMD) principle component
analysis (PCA) and genetic algorithm (GA) have been proposed
Liang et al [15] extracted EEG features where the approximate
entropy(ApEn) analysis was evaluated for its ability to analyze
multiclass EEGs and the performance of ApEn analysis was
enhanced by incorporating the power of EEG sub-bands or auto-
regressive models PCA and GA were compared for their ability to
select features by applying various linear and nonlinear methods
for classi1047297cation The maximum accuracy was 9867 Pachori [16]
decomposed EEG signals into intrinsic mode function (IMF) using
EMD and then computed mean frequency for each IMF by Fourierndash
Bessel expansion [20] to differentiate seizure and non-seizure EEG
signals Zhang et al [21] proposed an approach for the pattern
recognition of four complicated hand activities such as grasping
a football a small bar a cylinder and a hard paper based on EEG
signal in which each piece of raw data sequence for EEG signal is
decomposed by wavelet transform After forming a matrix they
applied singular value decomposition (SVD) on the matrix to 1047297ndthe singular value feature vectors These features are used as input
to arti1047297cial neural network to discriminate four hand activities and
they got 8900 classi1047297cation accuracy
Recently Bajaj et al [18] extracted amplitude modulation and
frequency modulation bandwidth as features from IMF using EMD
Among the existing contemporary techniques Bajaj et als tech-
nique is the latest and the best in terms of performance They used
the LS-SVM [23] technique for the classi1047297cation of seizure and
non-seizure EEG signals using the small dataset [19] They
obtained 980 to 995 accuracy using radial basis function (RBF)
kernel and also obtained 995 to 100 accuracy using Morlet
kernel
We have observed that the existing methods [11ndash18] exhibit
almost perfect classi1047297cation accuracy for the small dataset of
Fig 2 Gamma frequencies which provide the maximum amplitude for ictal and interictal signals (from a standard dataset [22])
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200192
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 411
seizure and non-seizure EEG signals However their performance
is below the expected results using the large dataset [22] for ictal
and interictal EEG signal classi1047297cations For example when we
applied the Bajaj et als [18] feature extraction and classi1047297cation
method to differentiate ictal and interictal EEG signals for the 1047297rst
IMF in the large dataset from Epilepsy Center of the University
Hospital of Freiburg [22] the classi1047297cation accuracy is 7790 for
Frontal lobe and 8020 for Temporal lobe which is far below the
accuracy 9850 obtained using the small dataset [19] of seizureand non-seizure EEG signals
The paper is organized as follows the details of the proposed
four methods with dataset feature extractions classi1047297cation
techniques and computational time are described in Section 3
the detail experimental results are explained in Section 4 while
Section 5 concludes the paper
3 Proposed methods
In this paper we propose four methods based on the various new
features extracted from DCT DCTndashDWT SVD and IMF We use the
transformation and decompositions to separate high frequency
components and then apply innovative ways to extract different
statistical features When we extract features from the high fre-quency component (ie IMF and detail signals respective) using EMD
and DWT we take total length of the high frequency components
and we do not use any weighting for them as all the coef 1047297cients
represent the high frequency components As SVD and DCT provide
coef 1047297cients of all frequencies we need to select a subset of the
coef 1047297cients which can represent high frequency components We
investigate the strength of each transformation and decomposition
for the capability of classi1047297cation accuracy as well as the combined
strength through different feature extractions The dataset individual
method their corresponding rationality and detailed methodology of
feature extractions are described in the following sub-sections
31 Dataset
Small Dataset [19] The existing seizure and non-seizure feature
extraction and classi1047297cation schemes [11ndash18] except [13] used
small dataset [19] The dataset consists of 1047297ve subsets each subset
contains 100 single-channels recoding and each recoding has
236 s in duration captured by the international 10ndash20 electrode
placement scheme ie 32 electrodes [14] Among them two
subsets are taken from healthy volunteers two subsets are taken
from seizure free intervals and another subset is also taken during
seizure period [18] (see sample examples in Fig 1(a)) Ocak [13]
used different datasets with the same pattern of small dataset
where 300 were normal and 100 were epileptic patients
Large Dataset [22] Seizure and brain discharges are much more
complicated than what can be seen in the scalp Thus in the paper we
have used intracranial EEG dataset using grid strip and depth electrodes
rather than scalp EEG The dataset used in the paper was recorded at
Epilepsy Centre of the University Hospital of Freiburg Germany [22] The
database contains invasive EEG recordings of 21 patients suffering frommedically intractable focal epilepsy The data obtained by Neuro1047297le NT
digital video EEG system with 128 channels 256 Hz sampling rate and
16 bit analog-to-digital converter According to the dataset [22] epileptic
EEG signals can be classi1047297ed into (i) ictal (ie seizure) (ii) preictal (a
certain period before seizure onset) (iii) interictal (period between
seizures) and (iv) non-seizure (period without any seizure activity after
a certain time of interictal) Normally duration of ictal is varied from few
seconds to 2 min The ictal records contain epileptic seizures with at
least 50 min of preictal data preceding each seizure The median time
period between the last seizure and the interictal data set is 5 h and
18 min and the median time period between the interictal data set and
the 1047297rst following seizure is 9 h and 36 min [29] As suggested in the
dataset we have treated preictal and ictal period as an ictal signals in
our experiments We do not consider any non-seizure signals into the
interictal period as we carefully select the same duration of interictal
signal (with the ictal signal) from the above duration of interictal signal
32 Feature extraction using DCT
DCT is a transformation method for converting a time series
signal into basic frequency in such a way that the DCT coef 1047297cients
are arranged from low frequency to high frequency components
Low frequency components represent the coarse signals and high
frequency components represent the detail signals DCT is widely
used in the image processing and video coding areas to compress
image signals based on their frequencies [34] As the ictal and
interictal EEG signals have different amplitudes and frequencies
(not visually separable) thus the most distinguishable featuresshould be located in the high frequency components of DCT
coef 1047297cients see in Fig 3(a) As the EEG signal is non-linear and
non-stationary in nature thus for real time processing of EEG
signals DCT may not be correct to directly correspond to the
frequency analysis however if we segment the EEG signals in time
window and apply DCT on them to 1047297nd DCT coef 1047297cients we can
avoid non-linear and non-stationary nature of the signals In our
experiment we apply DCT on the benchmark dataset [22] and take
Fig 3 (a)High frequency DCT coef 1047297cients and their energy and entropy for the signals of Temporal lobe the 1047297rst two rows show the high frequency DCT coef 1047297cients (ie
25) of ictal and interictal signals while the last row shows the energy and entropy of using the high frequency DCT coef 1047297cients (ie 25) (b) High frequency DWT
coef 1047297cients and STDs of high frequency DCT and DWT coef 1047297cients for the EEG signals of Temporal lobe the 1047297rst two rows show that the high frequency DWT coef 1047297cients of
ictal and interictal signals while the last row shows the STD of the high frequency DCT and DWT coef 1047297cients of ictal and interictal EEG signals
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 193
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 511
the higher frequency component (ie the last quartile (Q4) ie last
25 of DCT coef 1047297cient) Then we extract statistical features such as
entropy and energy of the weighted higher frequency coef 1047297cients to
classify ictal and interictal signals
The extraction of the statistical features from the Q4 is not
straightforward We use heuristics to provide different weight to
the different coef 1047297cients and investigate the classi1047297cation accu-
racy Although we try different type of weighting approximations
the experimental results show that exponential weight performswell The weight w i for the i-th position coef 1047297cient is selected as
wi frac14 β e β xi where xi x j frac14 1
Number of Selected Coeff icients We also try
different β and test for the classi1047297cation results We 1047297nd that
β frac142 provides the best results for the dataset Fig 4 shows an
example of weights According to the proposed weight we provide
more weight to the relatively low frequency coef 1047297cients and less
weight to the relatively high frequency coef 1047297cients within the
higher frequency area of the overall signal as sometimes the high
frequency coef 1047297cients in the tail would be very small
We de1047297ne weighted entropy and energy as follows
Energy frac14 sumn
i frac14 1
W i X ietht THORN
2and Entropy frac14 sum
n
i frac14 1
W iY 2i etht THORNlnethY 2i etht THORNTHORN eth1THORN
where n is the length of signal X (t ) is the higher frequency DCTcoef 1047297cients and Y (t ) is the normalized value of X (t )
The representation of ictal and interictal data for different higher
frequency DCT coef 1047297cients and their entropy and energy are given in
Fig 3(a) It can be observed that the changing rate of frequencies is
higher for interictal signal than ictal signal from Temporal lobe It is
also interesting to observe that the amplitude of the higher frequency
DCT coef 1047297cients for ictal signals is larger compared to that of interictal
signals Note that although the original ictal and interictal EEG signals
are not visually separable the high frequency DCT coef 1047297cients of ictal
and interictal signals are visually separable The third row of Fig 3
(a) also demonstrates that the weighted energy and entropy extracted
from the higher frequency DCT components have different values for
ictal and interictal signals Thus weighted entropy and energy would
be effective features for ictal and interictal signals classi1047297
cation
33 Feature extraction using DCT ndashDWT
DWT is a transformation and widely used in the image and signal
processing research to decompose the imagesignals into different
frequency and bands DWT can be speci1047297ed by low pass (LP) and high
pass (HP) 1047297lters named as standard quadrature mirror 1047297lters [28] The
cut-off frequency of these 1047297lters is one-fourth of the sampling
frequency The bandwidth of the 1047297lter outputs is half the bandwidth
of the original signal which generates downsampled output signals
without losing any information according to the Nyquist theorem [13]
The down-sampled signals from LP and HP 1047297lters refer to 1047297rst-level
approximation and details of the original signal respectively [24] DWT
is sometimes more effective compared to DCT to separate signals into
low and high frequency components when the signal is longer Thus
we also investigate the effectiveness of the DWT to extract some other
statistical features (such as STD) to classify ictal and interictal signals
In our second approach we apply single-level one dimensional DWT
on the EEG signal to 1047297nd the high frequency components from the
signal Then we calculate the STD from the detail frequency (ie highfrequency) components of the DWT coef 1047297cients and use as a feature
We use the whole high frequency component for feature extraction as
DWT separates the signals into low and high frequency components
Another feature is extracted using the STD of the weighted high
frequency DCT coef 1047297cients (similar to Fig 4) Two STD features from
DCT and DWT higher frequency components are then used for the
classi1047297cation purpose Note that we use different primary functions of
the wavelet transformation including Daubechies (db) series We 1047297nd
that db16 is the best for our experiment
In the experiment we take only the high frequency DWT compo-
nents The 1047297rst two rows in Fig 3(b) show the higher change of
frequency and lower amplitude for interictal signal compared to ictal
signal from Temporal lobe The third row of Fig 3(b) represents the
STDs generated from the high frequency DCT coef 1047297cients (weighted)
and DWT for ictal which are higher than that of interictal signal Thus
the STDs extracted from the high frequency DCT (weighted) and DWT
coef 1047297cients can be good features for classi1047297cation
34 Feature extraction using SVD
SVD decomposes two dimensional signals into three matrixes in
such a way that the middle matrix (ie singular value) has unique
properties to separate signals The middle matrix containing the
square root of eigenvalues and diagonalized contains much less non-
zero coef 1047297cients compared with the coef 1047297cient matrices of other
transformations [35] The values of diagonal matrix are with des-
cending order from top-left to bottom-right As the original ictal and
interictal signals are almost of the same nature we do not expect asigni1047297cant different values of other two matrixes after SVD however
we expect some distinguishable values on the extreme part (ie
bottom-right) of diagonal matrix Thus any feature extracted from
the bottom-right singular values could be used to classify ictal and
interictal signals To the best of our knowledge we are the 1047297rst to use
two dimensional SVD to analyze EEG signals for seizure detection
To extract the features using SVD we reshape the EEG signals
into a two dimensional matrix The column vector of EEG signal is
rearranged to square matrix for the computation of singular value
components using SVD From the singular values we take the
weighted Q4 (weight is similar to Fig 4 but the number of selected
coef 1047297cients are different) of the non-zero diagonal values to
calculate the STDs and energy and treat as the two features to
classify ictal and interictal signals SVD allows transforming any
given matrix AAC mxn to diagonal form using unitary matrices ie
A frac14 U Σ V H where Amn is a matrix U mm and V nnare orthogonal
matrix and Σ mn matrix with non-negative diagonal entries
The columns of U are called left singular vectors and the rows
of V H are called right singular vectors If A is a rectangular m n
matrix of rank K then U will be m m and Σ will be
Σ frac14
λ1 0 ⋯ 0
0 λ2 ⋯ 0
⋮ ⋮ ⋱ ⋮
0 0 ⋯ λn
⋮ ⋮ ⋮
0 0 ⋯ 0
2666666664
3777777775
eth2THORN
where λ1Z λ 2Zhellip λn-1 Z λnFig 4 Weight for different coef 1047297cients of selected high frequency coef 1047297cients
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200194
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 611
The data distribution of non-zero singular values of the ictal
and interictal signals from the Frontal lobe is shown in Fig 5(a)
The top part of the 1047297gure clearly shows that the Q4 singular values
of interictal signals are larger than that of the ictal signals The
bottom part of Fig 5(a) shows the STD and energy extracted from
the weighted Q4 of non-zero singular values from ictal and
interictal signals respectively The 1047297gure exposes that the STDs
extracted from the ictal and interictal signals have distinguishable
values thus they can be used for ictal and interictal signal
classi1047297cation
35 Feature extraction of IMF using EMD
The main strength of EMD is its ability to analyze nonlinear and
non-stationary signals like EEG signals When we apply EMD on
a signal the signal is decomposed into a 1047297nite number of signals
into IMFs which are ordered from higher frequency component to
lower frequency component The number of IMFs in a signal
depends on the local characteristics of the signal rather than
pre-de1047297ned number of IMFs This feature makes each IMF crucial
to identify some characteristics of the original signal Since the
IMFs are the representations of different frequenciesamplitudes
of the original signal and the distinguishing features between ictal
and interictal EEG signals lie on the frequencyamplitude our
hypothesis is that any feature extracted from IMFs will be an
effective feature to classify ictal and interictal successfully We also
observe that the ictal and interictal signals are deferred in higherorder frequency components thus our other hypothesis is that
features extracted from the 1047297rst few IMFs should be enough to
classify ictal and interictal signals successfully Through the
decomposition we keep only few mode functions of the original
signals for the 1047297nal feature extraction however we are able to
keep the important mode functions for the classi1047297cation purpose
Experimental results reveal that our both hypothesis are correct as
extracted features from the 1047297rst IMF successfully classify ictal and
interictal signals (see Section 4) Thus IMFs can be used in
classi1047297cation applications where the original signals are non-
linear and non-stationary and distinguishing features exist on
the different frequenciesamplitudes
Each IMF satis1047297es two following conditions (i) the number of
extrema and the number of zero crossing are identical or differ at
most by one and (ii) the mean value between the upper and the
lower envelope is equal to zero at any time
The EMD algorithm can be summarized as follows
1 Extract the extrema (minima and maxima) of the signal xetht THORN
2 Interpolate between minima and maxima to obtain
ℓminetht THORN and ℓmaxetht THORN
3 Calculate local mean metht THORN frac14 ℓminetht THORN thornℓmaxetht THORNfrac12 =2
4 Extract the detail detht THORN frac14 xetht THORN metht THORN
5 Check detht THORN is an IMF according to the conditions which are
mentioned above If yes repeat the procedure from the step1 on the residual signal r etht THORN frac14 xetht THORN metht THORN If no replace xetht THORN with
detht THORN and repeat the procedure from step 1
The pictorial scenario of generating IMFs is shown in Fig 6 The
signal xetht THORNcan be represented as a combination of IMFs and
residual component
xetht THORN frac14 sumL
i frac14 1
dietht THORN thornr Letht THORN eth3THORN
where di and r L are the i-th IMF and L-th is residual signal
We observe that features extracted from the 1047297rst two IMFs
provide the best classi1047297cation results To reduce computational
time we consider only the 1047297rst IMF from the EMD and then
calculate STD of IMF to capture the variation of the signal Byignoring other IMFs we do not sacri1047297ce any signi1047297cant classi1047297cation
accuracy The strength of a signal is distributed in the frequency
domain relative to the strengths of other ambient signals is central
to the design of any 1047297lter intended to extract or suppress the signal
To 1047297nd the strength of the signal we apply power spectral density
(PSD) on the 1047297rst IMF Then determine the STD feature from
normalized PSD The calculation of PSD is as follows
P ethwiTHORN frac14 1
N DethwiTHORN 2
eth4THORN
where DethwiTHORNis the discrete Fourier transform of IMFdietht THORN Normalized
PSD can be de1047297ned as
pi frac14 P ethwiTHORN=sum
i
P ethwiTHORN eth5THORN
Fig 5 (a) Non-zero singular value of SVD for Frontal lobe The 1047297rst row shows that the non-zero singular value distribution for ictal and interictal signals from Frontal lobe
and the last row shows the STD and energy of the last 25 non-zero singular values of ictal and interictal EEG signals (b) The 1047297rst IMFs of ictal and interictal signals with the
extracted features the top two rows represent the 1047297rst IMF from the ictal and interictal signals of Temporal lobe bottom row represents STD of the 1047297rst IMF and STD of the
normalized PSD (power spectral density) of the same IMF of ictal and interictal EEG signals
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 195
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 711
In the experiment we extract two features such as STD of the
1047297rst IMF and STD of normalized PSD of the 1047297rst IMF From the 1047297rst
two rows in Fig 5(b) it can be observed that the 1047297rst IMF of ictal
signal carries the higher amplitude compared to the interictal
signal
The third row of the 1047297gure represents the STD of the 1047297rst IMF
and STD of PSD of the same IMF from using ictal and interictal
signals The STDs for both cases of interictal signal are signi1047297cantly
larger than that of ictal signal Thus the two new features
extracted by EMD method can be good features for classi1047297cation
Table 1 presents the summary of all corresponding features use
for ictal and interictal classi1047297cation by each method
36 Classi 1047297er
We extract a number of features from different transforma-
tionsdecompositions To classify ictal and interictal signals we
need to use a classi1047297er The goal of a classi1047297er is to 1047297nd patients
states such as ictal (class 1) and interictal (class 2) using machine
learning approaches with cross-validation The challenge is to 1047297nd
the mapping that generalizes from training sets to unseen test
sets For the cross-validation data are partitioned into training set
and test set In our experiment we use 4-fold cross validation
where each dataset was randomly divided into 4 splits where
three of them were used for training and the fourth one used for
testing
To classify the ictal and interictal signals we extract various
features from DCT DWT SVD and IMF of EMD For classi1047297cation we
use SVM-based classi1047297er as the SVM is one of the best classi1047297ers in
the EEG signal analysis SVM is a potential methodology for solving
problem in linear and nonlinear classi1047297cation function estimation
and kernel based learning methods [25] It can minimize the
operational error and maximize the margin hyperplane as a result
it will maximize the classi1047297cation performance [25] LS-SVM is the
extended version of SVM and it is closely related to regularization
networks and Gaussian process and it also has primal-dual inter-
pretations [23] The major drawback of SVM is in its higher
computational burden of the constrained optimization program-
ming However LS-SVM can solve this problem [26]
The classi1047297er is a LS-SVM which learns nonlinear mapping
from the training set features xifrac141hellipnT where nT is the number
of training features into the patients state ictal ( thorn1) and interictal( 1) For unbiased classi1047297cation results [27] we divide the whole
1000 trials into four subsets in each subset we randomly select
75 samples for the validation training set and the remaining 25
for the validation testing set Let f yigi frac14 C 12 designate the LS-SVM
validation test outputs mapping to class 1 or class 2 The equation
of LS-SVM can be de1047297ned in [18] as
f eth xTHORN frac14 sign sumN
i 1
α i yiK eth x xiTHORN thornc
eth6THORN
where K ( x xi) is a kernel function α i are the Lagrange multipliers c
is the bias term xi is the training input and yi is the training
output pairs Linear RBF and Morlet kernel are used in our
experiments and RBF and Morlet functions can be de1047297ned in [18]
Fig 6 Extracted IMFs by EMD from Temporal lobe (Patient 12 channel one 15th datablock for ictal and 35th datablock for interictal)
Table 1Extracted features of proposed and existing methods
Proposed methods Feature-1 Feature-2
Q4WHF_SVD (Quartile four weighted high frequency component using SVD ) STD on 25 weighted high frequency
singular values
Energy on 25 weighted high frequency
singular values
Q4WHF_DCT (Quartile four weighted high frequency component using DCT ) Energy on 25 weighted high
frequency DCT coef 1047297cients
Entropy on 25 weighted high frequency DCT
coef 1047297cients
FIMF_EMD (First IMF using EMD ) STD on the 1047297rst IMF (ie high
frequency components)
STD on normalized PSD
Q4WHF_DCT ndash HF_DWT (Quartile four weighted high frequency component
using DCT and high frequency from DWT)
STD on 25 weighted high frequency
DCT coef 1047297cients
STD on detail signals (ie high frequency
components) after apply DWT
Combined Combined all above features
Existing methods Feature-1 Feature-2
EMD[18] Amplitude modulation bandwidth Frequency modulation bandwidth
DCT [17] Features from 6 of quantized DCT coef 1047297cients at lower frequency
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200196
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 811
as
keth x xiTHORN frac14 expeth x xi2=2s
2THORN eth7THORN
where s controls the width of the RBF function
keth x xiTHORN frac14 prodd
k frac14 1
cos ω0etheth xk xki THORN=aTHORN
h iexpeth xk xk
i =2a2THORN eth8THORN
where d is the dimension a is the 1047298exible coef 1047297cient and xki is the
k-th component of i-th training data Matlab codes for LS-SVMclassi1047297er are available at httpwwwesatkuleuvenacbesista
lssvmlab
37 Computational time
Feature extractions using IMF through EMD take longer time
due to two-fold iterations to 1047297nd all IMFs In the 1047297rst iteration
EMD decomposes a signal and checks whether the signal makes an
IMF and in the second iteration it subtracts the current IMF from
the signal and continues for the next IMF Even if we stop the
iteration for the 1047297rst IMF the EMD-based technique takes a full
1047297rst iteration for calculating the IMF The experimental data
reveals that the proposed technique using DCT DWT and SVD
takes less computational time compared to the EMD-based tech-nique as DCT DWT and SVD transformation is faster compared to
the EMD decomposition Thus the proposed methods (except
FIMF_EMD method ie First IMF using EMD) take insigni1047297cant
time compared to the existing method [18]
4 Experimental results
Sometimes EEG signals have line noise and other kinds of
artifacts due to muscle and body movements Notch 1047297lter and
independent component analysis (ICA)-based method are recom-
mended to remove the line noise and artifacts respectively for
suf 1047297cient amount of EEG signals [30ndash32] However in this experi-
ment we do not use any explicit noiseartifacts removal technique
to see the tolerance of the proposed method against noise
artifacts
We propose four methods and the experimental results of the
proposed methods are compared with existing methods [1718]
and corresponding features are presented in Table 1Our ultimate
target is to classify ictal and interictal EEG signals using LS-SVM
with different kernels such as RBF and Morlet where the values of regularization parameter and kernel parameter are 10 and 6
respectively for both kernels We also use linear kernel In this
paper four methods are proposed based on the new features of
ictal and interictal EEG signals captured from Frontal and Temporal
lobes using DCT DWT SVD and IMF of EMD to see the strength of
individual transformationdecomposition techniques for seizure
detection Then we also test the classi1047297cation accuracy by combin-
ing all features In our experiment we use 200 signals from ictal
and 800 signals from interictal EEG signals After training and
testing of all features we calculate average value from the four
subsets to compute sensitivity speci1047297city and accuracy [18] The
sensitivity speci1047297city and accuracy are de1047297ned as
Sensitivity frac14 ethTP =ethTP thornFN THORNTHORN 100 eth9THORN
Specificity frac14 ethTN =ethTN thornFP THORNTHORN 100 eth10THORN
Accuracy frac14 ethethTP thornTN THORN=ethTP thornTN thornFP thorn FN THORNTHORN 100 eth11THORN
where TP and TN represent the total number of detected true
positive events and true negative events respectively The FP and
FN represent false positive and false negative respectively
The technique in [18] claims that the second IMF provides
better classi1047297cation results while they use frequency and ampli-
tude modulation bandwidth features for the dataset [19] We
conduct experiments using 1047297rst four IMFs however we provide
results using only the 1047297rst IMF presented in Table 2 as it carries the
Table 2
Sensitivity speci1047297city and accuracy for different features of ictal and interictal EEG signals from Frontal and Temporal lobes
Classi1047297er Brain location Criteria Existing method Proposed method
Kernel Lobe EMD[18] DCT[17] Q4WHF_SVD Q4WHF_DCT FIMF_EMD Q4WHF_DCT ndashHF_DWT Combined
RBF Frontal SEN 2941 870 6666 100 9583 7678 4914
SPE 8051 7973 8220 8130 8186 8337 8382
ACC 7790 7810 8150 8160 8220 8300 7980
Temporal SEN 6667 400 100 9824 9710 9747 9647
SPE 8028 9863 8351 8473 8571 8665 8710
ACC 8020 7970 8420 8550 8650 8750 8790
Average SEN 4804 635 8333 9912 9647 8713 7281
SPE 8040 8918 8286 8302 8379 8501 8546
ACC 7905 7890 8285 8355 8435 8525 8385
Morlet Frontal SEN 3382 000 6383 8889 9333 6000 2442
SPE 8101 7978 8216 8126 8112 8396 8042
ACC 7780 789 8130 8140 8130 8180 7560
Temporal SEN 8462 200 100 9841 9012 9368 9468
SPE 8354 9900 8430 8527 8618 8632 8775
ACC 8360 7960 8510 8610 8650 8690 8840
Average SEN 5922 100 8192 9365 9173 7684 5955
SPE 8228 8939 8323 8327 8365 8514 8409
ACC 8070 7925 8320 8375 8390 8435 8200
Linear Frontal SEN 2895 2728 8095 100 100 100 8750
SPE 8035 8061 8131 8089 8155 8081 8223
ACC 7840 766 8130 8110 8190 8100 8240
Temporal SEN 8421 600 9800 9722 9811 9800 9836
SPE 8528 9613 8411 8288 8437 8412 8509
ACC 8520 7810 8480 8340 8510 8480 8590
Average SEN 5658 1664 8948 9861 9906 9900 9293
SPE 8282 8837 8271 8189 8296 8247 8366
ACC 8180 7735 8305 8225 8350 8290 8415
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 197
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 911
best result compared to other IMFs For the RBF kernel function
the proposed methods outperform the state-of-the-art method
consistently and the average sensitivity using the proposed meth-
ods such as Quartile four weighted high frequency component using
SVD (Q4WHF_SVD) Quartile four weighted high frequency compo-
nent using DCT (Q4WHF_DCT) First IMF using EMD (FIMF_EMD)
and Quartile four weighted high frequency component using DCT and
high frequency from DWT (Q4WHF_DCTndashHF_DWT) are 8333
9912 9647 and 8713 for RBF kernel respectively whereas
the average sensitivity using the state-of-art-methods [17] and
[18] are 635 and 4804 respectively Using Morlet kernel the
average sensitivity of the proposed methods is higher compared to
the existing methods [17] and [18] Moreover we justify our result
applying linear kernel Table 2 also presents that in most of the
cases the proposed methods with linear kernel provide better
results compared to the proposed methods with Morlet and RBF
kernels in terms of sensitivity Table 2 also presents that the
proposed Q4WHF_DCT method is the best among all proposed
methods in terms of sensitivity It is also interesting to note that
the accuracy of the proposed Q4WHF_DCTndashHF_DWT method is
reasonably consistence in terms of three criteria for RBF and
Morlet kernel According to Table 2 all proposed methods always
outperform the existing methods [1718] in terms of sensitivity (i
e seizure and ictal detection rate) however in some cases the
existing methods [1718] outperform one or more proposed
methods in terms of speci1047297city (interictal detection rate) For
seizure detection sensitivity is more important compared to the
speci1047297city detection as the cost of failure detecting ictal is heavier
compared to interictal In overall classi1047297cation the proposed
methods are also better compared to the existing methods We
combine all features to classify ictal and interictal EEG signalsHowever combining is not good because LS-SVM may not prop-
erly combine all available features to make a good prediction
model The average sensitivity of the proposed methods is 9136
(average result of all kernel and lobe) whereas the sensitivities of
the state-of-the-arts methods are 5461 [18] and 799 [17]
respectively
The performance of the LS-SVM is evaluated by (ROC) plot
shown in Fig 7(a) ROC illustrates the performance of a binary
classi1047297er system where it is created by plotting the fraction of true
positives from the positives ie true positive rate (TPR) vs the
fraction of false positives from negatives ie false positive rate
(FPR) with various threshold settings TPR is known as sensitivity
and FPR is one minus the speci1047297city or true negative rate Fig 7(a)
shows that the proposed four methods are carrying good
classi1047297cation results than that of the state-of-the-art method
[18] using dataset from Frontal lobe of the third subset of training
and testing EEG signals
Fig 7(b) shows the distribution of normalized feature values
(ie energy amp entropy in the Q4WHF_DCT method STD in the
Q4WHF_DCTndashHF_DWT method STD amp energy in the Q4WHF_SVD
method and STD in the FIMF EMD method) of ictal and interictal
from the Frontal lobe dataset [22] The 1047297gure shows that the range
of feature values for ictal is different from that of the interictal
signals This phenomenon is the other evidence why we use those
features to classify ictal and interictal EEG signals in the proposed
methods
First column of Fig 8 shows classi1047297cation comparison using the
proposed methods (ie FIMF EMD and Q4WHF_DCTndashHF_DWT)
and the state-of-the-art method [18] Fig 8(b) and (c) shows the
LS-SVM classi1047297cation results using the proposed FIMF EMD and
Q4WHF_DCTndashHF_DWT against the state-of-the-art method (ie
EMD Fig 8(a)) We generate 1047297rst column of Fig 8 from Frontal lobe
for the second subset of training and testing and obtain the
classi1047297cation accuracy 9120 for the Q4WHF_DCTndashHF_DWT
method 8720 for the FIMF_EMD method and 8080 for the
EMD method It can be concluded from these 1047297gures that the
proposed method based on the DCTndashDWT and the IMF outperform
the state-of-the-art method (ie EMD) The same classi1047297cation
comparisons are also provided in second column of Fig 8 using the
state-of-the-art method based on the EMD and the proposed
method based on the SVD and DCT respectively We consider
Temporal lobe for the fourth subset of training and testing obtain
classi1047297cation accuracy 9120 for Q4WHF_SVD 8960 for
Q4WHF_DCT and 8520 for EMD [18]Thus it is concluded that
the proposed methods based on DCT and SVD also outperform theEMD method [18]
5 Conclusion
Unlike seizure and non-seizure EEG signals classi1047297cation of
ictal and interictal EEG signals is a challenging task due to non
abrupt phenomena (ie visually non-separable) between ictal and
interictal EEG signals Moreover the features of ictal and interictal
signals are not consistence in different brain locations patient
types of seizure and hospital setting during epileptic period of a
patient Thus the classi1047297cation performance for ictal and interictal
EEG signals using the existing features extraction and classi1047297cation
techniques is not at expected level Our initial observation
Fig 7 (a) The receiver operating characteristics (ROC) curves of the third subset of training and testing EEG signals by three proposed methods ( Q4WHF_SVD FIMF_EMD
and Q4WHF_DCTndashHF_DWT) against the EMD method using LS-SVM with RBF kernel from Frontal lobe and (b) Normalized distribution of feature values extracted from ictal
and interictal EEG signals using DCT DWT SVD and IMF based method where 200 samples are taken from each of ictal and interictal categories in Frontal lobe
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200198
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1011
indicates that high frequency components of EEG signals have
distinguished characteristics To improve the accuracy in this
paper we propose novel feature extractions strategies from
the high frequency components based on a number of estab-
lished transformation and decompositions The average sensi-
tivity (which is the measure for seizure) of the proposed
methods is 9136 (average result of all kernel and lobe)
whereas the sensitivity of the state-of-the-arts methods are
5461 [18] and 799 [17] respectively The experimental
results show that the proposed methods outperform the state-
of-the-art methods in terms of accuracy sensitivity and speci-
1047297city for the ictal and interictal EEG signals with greater
consistence from the large benchmark dataset in two different
brain locations
Fig 8 First column represents the classi1047297cation of ictal and interictal EEG signals from Frontal lobe for the second subset of training and testing using (a) the state-of-the-art
method against (b) the proposed IMF and (c) the DCT-DWT method Second column represents the classi1047297cation of ictal and interictal EEG signals from Temporal lobe for the
fourth subset of training and testing using (d) the state-of-the-art method against (e) the proposed SVD and (f) the DCT methods
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 199
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1111
References
[1] Neurology Applied How Science is Bringing Music Instruction Back toExpressive Development URL langhttpwwwthomasjwestmusiccomneurolo
gyappliedhtmrang (accessed 270812)[2] EEG Signal The science behind Bo-Tau URL langhttpwwwbo-taucoukindex
phpaction=the_science_behind_bo_tau rang (accessed 270812)[3] Epilepsy-SeizuresURL langhttpwwwhealthcommunitiescomepilepsy-sei
zuresabout-seizuresshtml rang (accessed 270812)[4] M Soleymani T Pun M Pantic Multi-modal emotion recognition in response
to videos IEEE Trans Affect Comput 3 (2) (2012) 211ndash223[5] S Scholler S Bosse MS Treder B Blankertz G Curio K-R Muumlller
T Wiegand Toward a direct measure of video quality perception using EEGIEEE Trans Image Process 21 (5) (2012) 2619ndash2629
[6] W Di C Zhihua F Ruifang L Guangyu and L Tian Study on human brainafter consuming alcohol based on EEG signal in Proceedings of the Interna-tional Conference on Computer Science and Information Technology (ICCSIT)5 406ndash409 2010
[7] E Estrada H Nazeran F Ebrahimi and M Mikaeili EEG signal features forcomputer-aided sleep stage detection in Proceedings of the IEEE EMBSConference on Neural Engineering 669ndash672 2009
[8] ZM Hana1047297ah KFM Yunos ZH Murat MN Taib S Lias EEG brainwavepattern for smoking behaviour after horizontal rotation treatment in Pro-ceedings of the IEEE Student Conference on Research and Development 2009
[9] ZH Murat R Shilawani SA Kadir RM Isa and MN Taib The effects of
mobile phone usage on human braingwave using EEG in Proceedings of the13th International Conference on Modelling and Simulation 2011
[10] K Lehnertz RG Andrzejak J Arnhold T Kreuz F Mormann C Rieke
G Widman CE Elger Nonliner EEG analysis in Epilepsy its possible use forinterictal focus localization seizure ancipation and prevention J ClinNeurophsysiol 18 (3) (2001) 209ndash222
[11] R Panda PS Khobragade PD Jambhule SN Jengthe PR Pal and TK GandhiClassi1047297cation of EEG signal using wavelet transform and support vectormachine for epileptic seizure diction in Proceedings of the InternationalConference on Systems in Medicine and Biology (ICSMB) 405ndash408 2010
[12] SG Dastidar H Adeli N Dadmehr Mixed-band wavelet chaos-neural net-
work methodology for epilepsy and epileptic seizure detection IEEE TransBiomeical Eng 54 (9) (2007) 1545ndash1551
[13] H Ocak Optimal classi1047297cation of epileptic seizures in EEG using waveletanalysis and genetic algorithm Signal Process 88 (2008) 1858ndash1867
[14] K Polat S Gunes Classi1047297cation of epileptiform EEG using a hybrid systembased on decision tree classi1047297er and fast Fourier transform Appl Math
Comput 187 (2) (2007) 1017ndash1026[15] SF Liang HC Wang WL Chang Combination of EEG complexity and spectral
analysis for epilepsy diagnosis and seizure detection EURASIP J Adv SignalProcess 853434 (2010) 1ndash15 (Article ID)
[16] RB Pachori Discrimination between ictal and seizure-free EEG signals usingempirical mode decomposition Res Lett Signal Process 293056 (2008) 1ndash5(Article ID)
[17] D Birvinskas V Jusas I Martisius and R Damasevicius EEG dataset reductionand feature extraction using discrete cosine transform in Proceedings of theEuropean symposium on computer modeling and simulation (EMS) 199-2042012
[18] V Bajaj RB Pachori Classi1047297cation of seizure and non-seizure EEG signalsusing empirical mode decomposition IEEE Trans Inf Technol Biomed 16 (6)(2012) 1135ndash1142
[19] Small dataset Epilepsy dataset (text format) langhttpepileptologie-bonndecmsfront_contentphpidcat=193amplang=3ampchangelang=3rang (accessed 280412)
[20] J Harrison Fast and accurate bessel function computation in Proceedings of the19th IEEE international symposium on computer arithmetic 104ndash113
2009[21] X Zhang W Diao Z Cheng Wavelet transform and singular value decom-
position of EEG signal for pattern recognition of complicated hand activitiesLect Notes Comput Sci Digit Hum Model 4561 (2007) 294ndash303
[22] Large dataset EEG Data set from Epilepsy Center of the University Hospital of
Freiburg langhttpepilepsyuni-freiburgdefreiburg-seizure-prediction-projecteeg-databaserang (accessed 100612)
[23] KD Brabanter P Karsmakers F Ojeda C Alzate and JD Brabanter KPelckmans B De Moor J Vandewalle JAK Suykens LS-SVMlab ToolboxUsers Guide version 18 Katholieke Universiteit Leuven 2011
[24] S Sanei JA Chambers EEG Signal Processing John Wiley amp Sons Ltd UK
2007 (Centre of Digital Signal Processing Cardiff University)[25] S Abe Support Vector Machine for Pattern Classi1047297cation Second Edition
Springer London Dordrecht Heidelberg New York 2010
[26] H Wang and D Hu Comparison of SVM and LS-SVM for regression inProceedings of the international conference on neural networks and brain 1279ndash283 2005
[27] H Xing M Ha B Hu D Tian Linear feature-weighted support vectormachine Fuzzy Inf Eng 1 (3) (2009) 289ndash305
[28] A Subasi EEG signal classi1047297cation using wavelet feature extraction and amixture of expert model Expert Syst Appl 32 (4) (2007) 1084ndash1093(Elsevier)
[29] JR Williamson DW Bliss DW Browne JT Narayanan Seizure predictionusing EEG spatiotemporal correlation structure Epilepsy Behav 25 (2) (2012)230ndash238
[30] TP Jung S Makeig C Humphries et al Removing electroencephalographicsartifacts by blind source separation Psychophysiology 37 (2000) 163ndash178
[31] W De Clercq A Vergult B Vanrumste WV Paesschen SV Huffel Canonicalcorrelation analysis applied to remove muscle artifacts from the electroence-phalogram IEEE Trans Biomeical Eng 53 (12) (2006) 2583ndash2587
[32] J Ma P Tao S Bayram V Svetnik Muscle artifacts in multichannel EEGcharacteristics and reduction ClinNeurophysiol 123 (8) (2012) 1676ndash1686
[33] GA Worrell L Parish SD Cranstoun R Jonas G Baltuch B Litt High-frequency oscillations and seizure generation in neocortical epilepsy Brain127 (2004) 1496ndash1506
[34] M Paul W Lin CT Lau BndashS Lee A long term reference frame for hierarchicalb-picture based video coding IEEE Trans Circuits Syst Video Technol (2014)httpdxdoiorg101109TCSVT20142302555
[35] Z Gu W Lin CT Lau BndashS Lee and M Paul Two dimensional singular valuedecomposition (2D-SVD) based video coding in Proceedings of the IEEEconference on image processing (IEEE ICIP-10) 181ndash184 2010
Mohammad Zavid Parvez received BScEng (hons)Degree in Computer Science and Engineering fromAsian University of Bangladesh in 2003 and MScDegree in Electrical Engineering with emphasis onSignal Processing from Blekinge Institute of TechnologySweden in 2010 Parvezs research interests includebiomedical signal processing and machine learning
He has more than three year industry experience asa software developer Moreover he was involved sev-eral software projects in Denmark Currently he is aPhD student and teaching staff in Charles Sturt Uni-versity
Parvez is a Graduate student member of IEEE He hasreceived full funded scholarship from the Faculty He has published several journaland conference papers in the area of EEG signal analysis and cognitive radio
Manoranjan Paul received BScEng (hons) Degree inComputer Science and Engineering from BangladeshUniversity of Engineering and Technology (BUET) Ban-gladesh in 1997 and PhD Degree from Monash Uni-versity Australia in 2005 He was an Assistant Professorin Ahsanullah University of Science and Technology Hewas a Post-Doctoral Research Fellow in the Universityof New South Wales in 2005ndash2006 Monash Universityin 2006ndash2009 and Nanyang Technological Universityin 2009ndash2011 He has joined in the School of Comput-ing and Mathematics Charles Sturt University (CSU) at2011 Currently he is a Senior Lecturer and AssociateDirector of the Center for Research in Complex Systems
(CRiCS) in CSUDr Paul is a senior member of the IEEE and ACS He is in editorial board of three
international journals including EURASIP Journal on Advances in Signal Processing(JASP) Dr Paul has served as a guest editor of Journal of Multimedia and Journal of Computers for 1047297ve special issues He has been in Technical Program Committees of more than 25 international conferences Dr Paul obtained more than A$1Mcompetitive grant money including Australian Research Council (ARC) DiscoveryProject He has published more than 65 refereed publications including journalsand conferences He organized a special session on ldquoNew video coding technolo-giesrdquo in IEEE ISCAS 2010 He was a keynote speaker on ldquoVision friendly videocodingrdquo in IEEE ICCIT 2010 and tutorial speaker on ldquoMultiview Video Codingrdquo inDICTA 2013 His major research interests are in the 1047297elds of video coding computervision and EEG Signals analysis Dr Paul received Vice Chancellor ResearchExcellence Award in Faculty level at 2013
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200200
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 411
seizure and non-seizure EEG signals However their performance
is below the expected results using the large dataset [22] for ictal
and interictal EEG signal classi1047297cations For example when we
applied the Bajaj et als [18] feature extraction and classi1047297cation
method to differentiate ictal and interictal EEG signals for the 1047297rst
IMF in the large dataset from Epilepsy Center of the University
Hospital of Freiburg [22] the classi1047297cation accuracy is 7790 for
Frontal lobe and 8020 for Temporal lobe which is far below the
accuracy 9850 obtained using the small dataset [19] of seizureand non-seizure EEG signals
The paper is organized as follows the details of the proposed
four methods with dataset feature extractions classi1047297cation
techniques and computational time are described in Section 3
the detail experimental results are explained in Section 4 while
Section 5 concludes the paper
3 Proposed methods
In this paper we propose four methods based on the various new
features extracted from DCT DCTndashDWT SVD and IMF We use the
transformation and decompositions to separate high frequency
components and then apply innovative ways to extract different
statistical features When we extract features from the high fre-quency component (ie IMF and detail signals respective) using EMD
and DWT we take total length of the high frequency components
and we do not use any weighting for them as all the coef 1047297cients
represent the high frequency components As SVD and DCT provide
coef 1047297cients of all frequencies we need to select a subset of the
coef 1047297cients which can represent high frequency components We
investigate the strength of each transformation and decomposition
for the capability of classi1047297cation accuracy as well as the combined
strength through different feature extractions The dataset individual
method their corresponding rationality and detailed methodology of
feature extractions are described in the following sub-sections
31 Dataset
Small Dataset [19] The existing seizure and non-seizure feature
extraction and classi1047297cation schemes [11ndash18] except [13] used
small dataset [19] The dataset consists of 1047297ve subsets each subset
contains 100 single-channels recoding and each recoding has
236 s in duration captured by the international 10ndash20 electrode
placement scheme ie 32 electrodes [14] Among them two
subsets are taken from healthy volunteers two subsets are taken
from seizure free intervals and another subset is also taken during
seizure period [18] (see sample examples in Fig 1(a)) Ocak [13]
used different datasets with the same pattern of small dataset
where 300 were normal and 100 were epileptic patients
Large Dataset [22] Seizure and brain discharges are much more
complicated than what can be seen in the scalp Thus in the paper we
have used intracranial EEG dataset using grid strip and depth electrodes
rather than scalp EEG The dataset used in the paper was recorded at
Epilepsy Centre of the University Hospital of Freiburg Germany [22] The
database contains invasive EEG recordings of 21 patients suffering frommedically intractable focal epilepsy The data obtained by Neuro1047297le NT
digital video EEG system with 128 channels 256 Hz sampling rate and
16 bit analog-to-digital converter According to the dataset [22] epileptic
EEG signals can be classi1047297ed into (i) ictal (ie seizure) (ii) preictal (a
certain period before seizure onset) (iii) interictal (period between
seizures) and (iv) non-seizure (period without any seizure activity after
a certain time of interictal) Normally duration of ictal is varied from few
seconds to 2 min The ictal records contain epileptic seizures with at
least 50 min of preictal data preceding each seizure The median time
period between the last seizure and the interictal data set is 5 h and
18 min and the median time period between the interictal data set and
the 1047297rst following seizure is 9 h and 36 min [29] As suggested in the
dataset we have treated preictal and ictal period as an ictal signals in
our experiments We do not consider any non-seizure signals into the
interictal period as we carefully select the same duration of interictal
signal (with the ictal signal) from the above duration of interictal signal
32 Feature extraction using DCT
DCT is a transformation method for converting a time series
signal into basic frequency in such a way that the DCT coef 1047297cients
are arranged from low frequency to high frequency components
Low frequency components represent the coarse signals and high
frequency components represent the detail signals DCT is widely
used in the image processing and video coding areas to compress
image signals based on their frequencies [34] As the ictal and
interictal EEG signals have different amplitudes and frequencies
(not visually separable) thus the most distinguishable featuresshould be located in the high frequency components of DCT
coef 1047297cients see in Fig 3(a) As the EEG signal is non-linear and
non-stationary in nature thus for real time processing of EEG
signals DCT may not be correct to directly correspond to the
frequency analysis however if we segment the EEG signals in time
window and apply DCT on them to 1047297nd DCT coef 1047297cients we can
avoid non-linear and non-stationary nature of the signals In our
experiment we apply DCT on the benchmark dataset [22] and take
Fig 3 (a)High frequency DCT coef 1047297cients and their energy and entropy for the signals of Temporal lobe the 1047297rst two rows show the high frequency DCT coef 1047297cients (ie
25) of ictal and interictal signals while the last row shows the energy and entropy of using the high frequency DCT coef 1047297cients (ie 25) (b) High frequency DWT
coef 1047297cients and STDs of high frequency DCT and DWT coef 1047297cients for the EEG signals of Temporal lobe the 1047297rst two rows show that the high frequency DWT coef 1047297cients of
ictal and interictal signals while the last row shows the STD of the high frequency DCT and DWT coef 1047297cients of ictal and interictal EEG signals
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 193
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 511
the higher frequency component (ie the last quartile (Q4) ie last
25 of DCT coef 1047297cient) Then we extract statistical features such as
entropy and energy of the weighted higher frequency coef 1047297cients to
classify ictal and interictal signals
The extraction of the statistical features from the Q4 is not
straightforward We use heuristics to provide different weight to
the different coef 1047297cients and investigate the classi1047297cation accu-
racy Although we try different type of weighting approximations
the experimental results show that exponential weight performswell The weight w i for the i-th position coef 1047297cient is selected as
wi frac14 β e β xi where xi x j frac14 1
Number of Selected Coeff icients We also try
different β and test for the classi1047297cation results We 1047297nd that
β frac142 provides the best results for the dataset Fig 4 shows an
example of weights According to the proposed weight we provide
more weight to the relatively low frequency coef 1047297cients and less
weight to the relatively high frequency coef 1047297cients within the
higher frequency area of the overall signal as sometimes the high
frequency coef 1047297cients in the tail would be very small
We de1047297ne weighted entropy and energy as follows
Energy frac14 sumn
i frac14 1
W i X ietht THORN
2and Entropy frac14 sum
n
i frac14 1
W iY 2i etht THORNlnethY 2i etht THORNTHORN eth1THORN
where n is the length of signal X (t ) is the higher frequency DCTcoef 1047297cients and Y (t ) is the normalized value of X (t )
The representation of ictal and interictal data for different higher
frequency DCT coef 1047297cients and their entropy and energy are given in
Fig 3(a) It can be observed that the changing rate of frequencies is
higher for interictal signal than ictal signal from Temporal lobe It is
also interesting to observe that the amplitude of the higher frequency
DCT coef 1047297cients for ictal signals is larger compared to that of interictal
signals Note that although the original ictal and interictal EEG signals
are not visually separable the high frequency DCT coef 1047297cients of ictal
and interictal signals are visually separable The third row of Fig 3
(a) also demonstrates that the weighted energy and entropy extracted
from the higher frequency DCT components have different values for
ictal and interictal signals Thus weighted entropy and energy would
be effective features for ictal and interictal signals classi1047297
cation
33 Feature extraction using DCT ndashDWT
DWT is a transformation and widely used in the image and signal
processing research to decompose the imagesignals into different
frequency and bands DWT can be speci1047297ed by low pass (LP) and high
pass (HP) 1047297lters named as standard quadrature mirror 1047297lters [28] The
cut-off frequency of these 1047297lters is one-fourth of the sampling
frequency The bandwidth of the 1047297lter outputs is half the bandwidth
of the original signal which generates downsampled output signals
without losing any information according to the Nyquist theorem [13]
The down-sampled signals from LP and HP 1047297lters refer to 1047297rst-level
approximation and details of the original signal respectively [24] DWT
is sometimes more effective compared to DCT to separate signals into
low and high frequency components when the signal is longer Thus
we also investigate the effectiveness of the DWT to extract some other
statistical features (such as STD) to classify ictal and interictal signals
In our second approach we apply single-level one dimensional DWT
on the EEG signal to 1047297nd the high frequency components from the
signal Then we calculate the STD from the detail frequency (ie highfrequency) components of the DWT coef 1047297cients and use as a feature
We use the whole high frequency component for feature extraction as
DWT separates the signals into low and high frequency components
Another feature is extracted using the STD of the weighted high
frequency DCT coef 1047297cients (similar to Fig 4) Two STD features from
DCT and DWT higher frequency components are then used for the
classi1047297cation purpose Note that we use different primary functions of
the wavelet transformation including Daubechies (db) series We 1047297nd
that db16 is the best for our experiment
In the experiment we take only the high frequency DWT compo-
nents The 1047297rst two rows in Fig 3(b) show the higher change of
frequency and lower amplitude for interictal signal compared to ictal
signal from Temporal lobe The third row of Fig 3(b) represents the
STDs generated from the high frequency DCT coef 1047297cients (weighted)
and DWT for ictal which are higher than that of interictal signal Thus
the STDs extracted from the high frequency DCT (weighted) and DWT
coef 1047297cients can be good features for classi1047297cation
34 Feature extraction using SVD
SVD decomposes two dimensional signals into three matrixes in
such a way that the middle matrix (ie singular value) has unique
properties to separate signals The middle matrix containing the
square root of eigenvalues and diagonalized contains much less non-
zero coef 1047297cients compared with the coef 1047297cient matrices of other
transformations [35] The values of diagonal matrix are with des-
cending order from top-left to bottom-right As the original ictal and
interictal signals are almost of the same nature we do not expect asigni1047297cant different values of other two matrixes after SVD however
we expect some distinguishable values on the extreme part (ie
bottom-right) of diagonal matrix Thus any feature extracted from
the bottom-right singular values could be used to classify ictal and
interictal signals To the best of our knowledge we are the 1047297rst to use
two dimensional SVD to analyze EEG signals for seizure detection
To extract the features using SVD we reshape the EEG signals
into a two dimensional matrix The column vector of EEG signal is
rearranged to square matrix for the computation of singular value
components using SVD From the singular values we take the
weighted Q4 (weight is similar to Fig 4 but the number of selected
coef 1047297cients are different) of the non-zero diagonal values to
calculate the STDs and energy and treat as the two features to
classify ictal and interictal signals SVD allows transforming any
given matrix AAC mxn to diagonal form using unitary matrices ie
A frac14 U Σ V H where Amn is a matrix U mm and V nnare orthogonal
matrix and Σ mn matrix with non-negative diagonal entries
The columns of U are called left singular vectors and the rows
of V H are called right singular vectors If A is a rectangular m n
matrix of rank K then U will be m m and Σ will be
Σ frac14
λ1 0 ⋯ 0
0 λ2 ⋯ 0
⋮ ⋮ ⋱ ⋮
0 0 ⋯ λn
⋮ ⋮ ⋮
0 0 ⋯ 0
2666666664
3777777775
eth2THORN
where λ1Z λ 2Zhellip λn-1 Z λnFig 4 Weight for different coef 1047297cients of selected high frequency coef 1047297cients
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200194
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 611
The data distribution of non-zero singular values of the ictal
and interictal signals from the Frontal lobe is shown in Fig 5(a)
The top part of the 1047297gure clearly shows that the Q4 singular values
of interictal signals are larger than that of the ictal signals The
bottom part of Fig 5(a) shows the STD and energy extracted from
the weighted Q4 of non-zero singular values from ictal and
interictal signals respectively The 1047297gure exposes that the STDs
extracted from the ictal and interictal signals have distinguishable
values thus they can be used for ictal and interictal signal
classi1047297cation
35 Feature extraction of IMF using EMD
The main strength of EMD is its ability to analyze nonlinear and
non-stationary signals like EEG signals When we apply EMD on
a signal the signal is decomposed into a 1047297nite number of signals
into IMFs which are ordered from higher frequency component to
lower frequency component The number of IMFs in a signal
depends on the local characteristics of the signal rather than
pre-de1047297ned number of IMFs This feature makes each IMF crucial
to identify some characteristics of the original signal Since the
IMFs are the representations of different frequenciesamplitudes
of the original signal and the distinguishing features between ictal
and interictal EEG signals lie on the frequencyamplitude our
hypothesis is that any feature extracted from IMFs will be an
effective feature to classify ictal and interictal successfully We also
observe that the ictal and interictal signals are deferred in higherorder frequency components thus our other hypothesis is that
features extracted from the 1047297rst few IMFs should be enough to
classify ictal and interictal signals successfully Through the
decomposition we keep only few mode functions of the original
signals for the 1047297nal feature extraction however we are able to
keep the important mode functions for the classi1047297cation purpose
Experimental results reveal that our both hypothesis are correct as
extracted features from the 1047297rst IMF successfully classify ictal and
interictal signals (see Section 4) Thus IMFs can be used in
classi1047297cation applications where the original signals are non-
linear and non-stationary and distinguishing features exist on
the different frequenciesamplitudes
Each IMF satis1047297es two following conditions (i) the number of
extrema and the number of zero crossing are identical or differ at
most by one and (ii) the mean value between the upper and the
lower envelope is equal to zero at any time
The EMD algorithm can be summarized as follows
1 Extract the extrema (minima and maxima) of the signal xetht THORN
2 Interpolate between minima and maxima to obtain
ℓminetht THORN and ℓmaxetht THORN
3 Calculate local mean metht THORN frac14 ℓminetht THORN thornℓmaxetht THORNfrac12 =2
4 Extract the detail detht THORN frac14 xetht THORN metht THORN
5 Check detht THORN is an IMF according to the conditions which are
mentioned above If yes repeat the procedure from the step1 on the residual signal r etht THORN frac14 xetht THORN metht THORN If no replace xetht THORN with
detht THORN and repeat the procedure from step 1
The pictorial scenario of generating IMFs is shown in Fig 6 The
signal xetht THORNcan be represented as a combination of IMFs and
residual component
xetht THORN frac14 sumL
i frac14 1
dietht THORN thornr Letht THORN eth3THORN
where di and r L are the i-th IMF and L-th is residual signal
We observe that features extracted from the 1047297rst two IMFs
provide the best classi1047297cation results To reduce computational
time we consider only the 1047297rst IMF from the EMD and then
calculate STD of IMF to capture the variation of the signal Byignoring other IMFs we do not sacri1047297ce any signi1047297cant classi1047297cation
accuracy The strength of a signal is distributed in the frequency
domain relative to the strengths of other ambient signals is central
to the design of any 1047297lter intended to extract or suppress the signal
To 1047297nd the strength of the signal we apply power spectral density
(PSD) on the 1047297rst IMF Then determine the STD feature from
normalized PSD The calculation of PSD is as follows
P ethwiTHORN frac14 1
N DethwiTHORN 2
eth4THORN
where DethwiTHORNis the discrete Fourier transform of IMFdietht THORN Normalized
PSD can be de1047297ned as
pi frac14 P ethwiTHORN=sum
i
P ethwiTHORN eth5THORN
Fig 5 (a) Non-zero singular value of SVD for Frontal lobe The 1047297rst row shows that the non-zero singular value distribution for ictal and interictal signals from Frontal lobe
and the last row shows the STD and energy of the last 25 non-zero singular values of ictal and interictal EEG signals (b) The 1047297rst IMFs of ictal and interictal signals with the
extracted features the top two rows represent the 1047297rst IMF from the ictal and interictal signals of Temporal lobe bottom row represents STD of the 1047297rst IMF and STD of the
normalized PSD (power spectral density) of the same IMF of ictal and interictal EEG signals
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 195
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 711
In the experiment we extract two features such as STD of the
1047297rst IMF and STD of normalized PSD of the 1047297rst IMF From the 1047297rst
two rows in Fig 5(b) it can be observed that the 1047297rst IMF of ictal
signal carries the higher amplitude compared to the interictal
signal
The third row of the 1047297gure represents the STD of the 1047297rst IMF
and STD of PSD of the same IMF from using ictal and interictal
signals The STDs for both cases of interictal signal are signi1047297cantly
larger than that of ictal signal Thus the two new features
extracted by EMD method can be good features for classi1047297cation
Table 1 presents the summary of all corresponding features use
for ictal and interictal classi1047297cation by each method
36 Classi 1047297er
We extract a number of features from different transforma-
tionsdecompositions To classify ictal and interictal signals we
need to use a classi1047297er The goal of a classi1047297er is to 1047297nd patients
states such as ictal (class 1) and interictal (class 2) using machine
learning approaches with cross-validation The challenge is to 1047297nd
the mapping that generalizes from training sets to unseen test
sets For the cross-validation data are partitioned into training set
and test set In our experiment we use 4-fold cross validation
where each dataset was randomly divided into 4 splits where
three of them were used for training and the fourth one used for
testing
To classify the ictal and interictal signals we extract various
features from DCT DWT SVD and IMF of EMD For classi1047297cation we
use SVM-based classi1047297er as the SVM is one of the best classi1047297ers in
the EEG signal analysis SVM is a potential methodology for solving
problem in linear and nonlinear classi1047297cation function estimation
and kernel based learning methods [25] It can minimize the
operational error and maximize the margin hyperplane as a result
it will maximize the classi1047297cation performance [25] LS-SVM is the
extended version of SVM and it is closely related to regularization
networks and Gaussian process and it also has primal-dual inter-
pretations [23] The major drawback of SVM is in its higher
computational burden of the constrained optimization program-
ming However LS-SVM can solve this problem [26]
The classi1047297er is a LS-SVM which learns nonlinear mapping
from the training set features xifrac141hellipnT where nT is the number
of training features into the patients state ictal ( thorn1) and interictal( 1) For unbiased classi1047297cation results [27] we divide the whole
1000 trials into four subsets in each subset we randomly select
75 samples for the validation training set and the remaining 25
for the validation testing set Let f yigi frac14 C 12 designate the LS-SVM
validation test outputs mapping to class 1 or class 2 The equation
of LS-SVM can be de1047297ned in [18] as
f eth xTHORN frac14 sign sumN
i 1
α i yiK eth x xiTHORN thornc
eth6THORN
where K ( x xi) is a kernel function α i are the Lagrange multipliers c
is the bias term xi is the training input and yi is the training
output pairs Linear RBF and Morlet kernel are used in our
experiments and RBF and Morlet functions can be de1047297ned in [18]
Fig 6 Extracted IMFs by EMD from Temporal lobe (Patient 12 channel one 15th datablock for ictal and 35th datablock for interictal)
Table 1Extracted features of proposed and existing methods
Proposed methods Feature-1 Feature-2
Q4WHF_SVD (Quartile four weighted high frequency component using SVD ) STD on 25 weighted high frequency
singular values
Energy on 25 weighted high frequency
singular values
Q4WHF_DCT (Quartile four weighted high frequency component using DCT ) Energy on 25 weighted high
frequency DCT coef 1047297cients
Entropy on 25 weighted high frequency DCT
coef 1047297cients
FIMF_EMD (First IMF using EMD ) STD on the 1047297rst IMF (ie high
frequency components)
STD on normalized PSD
Q4WHF_DCT ndash HF_DWT (Quartile four weighted high frequency component
using DCT and high frequency from DWT)
STD on 25 weighted high frequency
DCT coef 1047297cients
STD on detail signals (ie high frequency
components) after apply DWT
Combined Combined all above features
Existing methods Feature-1 Feature-2
EMD[18] Amplitude modulation bandwidth Frequency modulation bandwidth
DCT [17] Features from 6 of quantized DCT coef 1047297cients at lower frequency
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200196
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 811
as
keth x xiTHORN frac14 expeth x xi2=2s
2THORN eth7THORN
where s controls the width of the RBF function
keth x xiTHORN frac14 prodd
k frac14 1
cos ω0etheth xk xki THORN=aTHORN
h iexpeth xk xk
i =2a2THORN eth8THORN
where d is the dimension a is the 1047298exible coef 1047297cient and xki is the
k-th component of i-th training data Matlab codes for LS-SVMclassi1047297er are available at httpwwwesatkuleuvenacbesista
lssvmlab
37 Computational time
Feature extractions using IMF through EMD take longer time
due to two-fold iterations to 1047297nd all IMFs In the 1047297rst iteration
EMD decomposes a signal and checks whether the signal makes an
IMF and in the second iteration it subtracts the current IMF from
the signal and continues for the next IMF Even if we stop the
iteration for the 1047297rst IMF the EMD-based technique takes a full
1047297rst iteration for calculating the IMF The experimental data
reveals that the proposed technique using DCT DWT and SVD
takes less computational time compared to the EMD-based tech-nique as DCT DWT and SVD transformation is faster compared to
the EMD decomposition Thus the proposed methods (except
FIMF_EMD method ie First IMF using EMD) take insigni1047297cant
time compared to the existing method [18]
4 Experimental results
Sometimes EEG signals have line noise and other kinds of
artifacts due to muscle and body movements Notch 1047297lter and
independent component analysis (ICA)-based method are recom-
mended to remove the line noise and artifacts respectively for
suf 1047297cient amount of EEG signals [30ndash32] However in this experi-
ment we do not use any explicit noiseartifacts removal technique
to see the tolerance of the proposed method against noise
artifacts
We propose four methods and the experimental results of the
proposed methods are compared with existing methods [1718]
and corresponding features are presented in Table 1Our ultimate
target is to classify ictal and interictal EEG signals using LS-SVM
with different kernels such as RBF and Morlet where the values of regularization parameter and kernel parameter are 10 and 6
respectively for both kernels We also use linear kernel In this
paper four methods are proposed based on the new features of
ictal and interictal EEG signals captured from Frontal and Temporal
lobes using DCT DWT SVD and IMF of EMD to see the strength of
individual transformationdecomposition techniques for seizure
detection Then we also test the classi1047297cation accuracy by combin-
ing all features In our experiment we use 200 signals from ictal
and 800 signals from interictal EEG signals After training and
testing of all features we calculate average value from the four
subsets to compute sensitivity speci1047297city and accuracy [18] The
sensitivity speci1047297city and accuracy are de1047297ned as
Sensitivity frac14 ethTP =ethTP thornFN THORNTHORN 100 eth9THORN
Specificity frac14 ethTN =ethTN thornFP THORNTHORN 100 eth10THORN
Accuracy frac14 ethethTP thornTN THORN=ethTP thornTN thornFP thorn FN THORNTHORN 100 eth11THORN
where TP and TN represent the total number of detected true
positive events and true negative events respectively The FP and
FN represent false positive and false negative respectively
The technique in [18] claims that the second IMF provides
better classi1047297cation results while they use frequency and ampli-
tude modulation bandwidth features for the dataset [19] We
conduct experiments using 1047297rst four IMFs however we provide
results using only the 1047297rst IMF presented in Table 2 as it carries the
Table 2
Sensitivity speci1047297city and accuracy for different features of ictal and interictal EEG signals from Frontal and Temporal lobes
Classi1047297er Brain location Criteria Existing method Proposed method
Kernel Lobe EMD[18] DCT[17] Q4WHF_SVD Q4WHF_DCT FIMF_EMD Q4WHF_DCT ndashHF_DWT Combined
RBF Frontal SEN 2941 870 6666 100 9583 7678 4914
SPE 8051 7973 8220 8130 8186 8337 8382
ACC 7790 7810 8150 8160 8220 8300 7980
Temporal SEN 6667 400 100 9824 9710 9747 9647
SPE 8028 9863 8351 8473 8571 8665 8710
ACC 8020 7970 8420 8550 8650 8750 8790
Average SEN 4804 635 8333 9912 9647 8713 7281
SPE 8040 8918 8286 8302 8379 8501 8546
ACC 7905 7890 8285 8355 8435 8525 8385
Morlet Frontal SEN 3382 000 6383 8889 9333 6000 2442
SPE 8101 7978 8216 8126 8112 8396 8042
ACC 7780 789 8130 8140 8130 8180 7560
Temporal SEN 8462 200 100 9841 9012 9368 9468
SPE 8354 9900 8430 8527 8618 8632 8775
ACC 8360 7960 8510 8610 8650 8690 8840
Average SEN 5922 100 8192 9365 9173 7684 5955
SPE 8228 8939 8323 8327 8365 8514 8409
ACC 8070 7925 8320 8375 8390 8435 8200
Linear Frontal SEN 2895 2728 8095 100 100 100 8750
SPE 8035 8061 8131 8089 8155 8081 8223
ACC 7840 766 8130 8110 8190 8100 8240
Temporal SEN 8421 600 9800 9722 9811 9800 9836
SPE 8528 9613 8411 8288 8437 8412 8509
ACC 8520 7810 8480 8340 8510 8480 8590
Average SEN 5658 1664 8948 9861 9906 9900 9293
SPE 8282 8837 8271 8189 8296 8247 8366
ACC 8180 7735 8305 8225 8350 8290 8415
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 197
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 911
best result compared to other IMFs For the RBF kernel function
the proposed methods outperform the state-of-the-art method
consistently and the average sensitivity using the proposed meth-
ods such as Quartile four weighted high frequency component using
SVD (Q4WHF_SVD) Quartile four weighted high frequency compo-
nent using DCT (Q4WHF_DCT) First IMF using EMD (FIMF_EMD)
and Quartile four weighted high frequency component using DCT and
high frequency from DWT (Q4WHF_DCTndashHF_DWT) are 8333
9912 9647 and 8713 for RBF kernel respectively whereas
the average sensitivity using the state-of-art-methods [17] and
[18] are 635 and 4804 respectively Using Morlet kernel the
average sensitivity of the proposed methods is higher compared to
the existing methods [17] and [18] Moreover we justify our result
applying linear kernel Table 2 also presents that in most of the
cases the proposed methods with linear kernel provide better
results compared to the proposed methods with Morlet and RBF
kernels in terms of sensitivity Table 2 also presents that the
proposed Q4WHF_DCT method is the best among all proposed
methods in terms of sensitivity It is also interesting to note that
the accuracy of the proposed Q4WHF_DCTndashHF_DWT method is
reasonably consistence in terms of three criteria for RBF and
Morlet kernel According to Table 2 all proposed methods always
outperform the existing methods [1718] in terms of sensitivity (i
e seizure and ictal detection rate) however in some cases the
existing methods [1718] outperform one or more proposed
methods in terms of speci1047297city (interictal detection rate) For
seizure detection sensitivity is more important compared to the
speci1047297city detection as the cost of failure detecting ictal is heavier
compared to interictal In overall classi1047297cation the proposed
methods are also better compared to the existing methods We
combine all features to classify ictal and interictal EEG signalsHowever combining is not good because LS-SVM may not prop-
erly combine all available features to make a good prediction
model The average sensitivity of the proposed methods is 9136
(average result of all kernel and lobe) whereas the sensitivities of
the state-of-the-arts methods are 5461 [18] and 799 [17]
respectively
The performance of the LS-SVM is evaluated by (ROC) plot
shown in Fig 7(a) ROC illustrates the performance of a binary
classi1047297er system where it is created by plotting the fraction of true
positives from the positives ie true positive rate (TPR) vs the
fraction of false positives from negatives ie false positive rate
(FPR) with various threshold settings TPR is known as sensitivity
and FPR is one minus the speci1047297city or true negative rate Fig 7(a)
shows that the proposed four methods are carrying good
classi1047297cation results than that of the state-of-the-art method
[18] using dataset from Frontal lobe of the third subset of training
and testing EEG signals
Fig 7(b) shows the distribution of normalized feature values
(ie energy amp entropy in the Q4WHF_DCT method STD in the
Q4WHF_DCTndashHF_DWT method STD amp energy in the Q4WHF_SVD
method and STD in the FIMF EMD method) of ictal and interictal
from the Frontal lobe dataset [22] The 1047297gure shows that the range
of feature values for ictal is different from that of the interictal
signals This phenomenon is the other evidence why we use those
features to classify ictal and interictal EEG signals in the proposed
methods
First column of Fig 8 shows classi1047297cation comparison using the
proposed methods (ie FIMF EMD and Q4WHF_DCTndashHF_DWT)
and the state-of-the-art method [18] Fig 8(b) and (c) shows the
LS-SVM classi1047297cation results using the proposed FIMF EMD and
Q4WHF_DCTndashHF_DWT against the state-of-the-art method (ie
EMD Fig 8(a)) We generate 1047297rst column of Fig 8 from Frontal lobe
for the second subset of training and testing and obtain the
classi1047297cation accuracy 9120 for the Q4WHF_DCTndashHF_DWT
method 8720 for the FIMF_EMD method and 8080 for the
EMD method It can be concluded from these 1047297gures that the
proposed method based on the DCTndashDWT and the IMF outperform
the state-of-the-art method (ie EMD) The same classi1047297cation
comparisons are also provided in second column of Fig 8 using the
state-of-the-art method based on the EMD and the proposed
method based on the SVD and DCT respectively We consider
Temporal lobe for the fourth subset of training and testing obtain
classi1047297cation accuracy 9120 for Q4WHF_SVD 8960 for
Q4WHF_DCT and 8520 for EMD [18]Thus it is concluded that
the proposed methods based on DCT and SVD also outperform theEMD method [18]
5 Conclusion
Unlike seizure and non-seizure EEG signals classi1047297cation of
ictal and interictal EEG signals is a challenging task due to non
abrupt phenomena (ie visually non-separable) between ictal and
interictal EEG signals Moreover the features of ictal and interictal
signals are not consistence in different brain locations patient
types of seizure and hospital setting during epileptic period of a
patient Thus the classi1047297cation performance for ictal and interictal
EEG signals using the existing features extraction and classi1047297cation
techniques is not at expected level Our initial observation
Fig 7 (a) The receiver operating characteristics (ROC) curves of the third subset of training and testing EEG signals by three proposed methods ( Q4WHF_SVD FIMF_EMD
and Q4WHF_DCTndashHF_DWT) against the EMD method using LS-SVM with RBF kernel from Frontal lobe and (b) Normalized distribution of feature values extracted from ictal
and interictal EEG signals using DCT DWT SVD and IMF based method where 200 samples are taken from each of ictal and interictal categories in Frontal lobe
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200198
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1011
indicates that high frequency components of EEG signals have
distinguished characteristics To improve the accuracy in this
paper we propose novel feature extractions strategies from
the high frequency components based on a number of estab-
lished transformation and decompositions The average sensi-
tivity (which is the measure for seizure) of the proposed
methods is 9136 (average result of all kernel and lobe)
whereas the sensitivity of the state-of-the-arts methods are
5461 [18] and 799 [17] respectively The experimental
results show that the proposed methods outperform the state-
of-the-art methods in terms of accuracy sensitivity and speci-
1047297city for the ictal and interictal EEG signals with greater
consistence from the large benchmark dataset in two different
brain locations
Fig 8 First column represents the classi1047297cation of ictal and interictal EEG signals from Frontal lobe for the second subset of training and testing using (a) the state-of-the-art
method against (b) the proposed IMF and (c) the DCT-DWT method Second column represents the classi1047297cation of ictal and interictal EEG signals from Temporal lobe for the
fourth subset of training and testing using (d) the state-of-the-art method against (e) the proposed SVD and (f) the DCT methods
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 199
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1111
References
[1] Neurology Applied How Science is Bringing Music Instruction Back toExpressive Development URL langhttpwwwthomasjwestmusiccomneurolo
gyappliedhtmrang (accessed 270812)[2] EEG Signal The science behind Bo-Tau URL langhttpwwwbo-taucoukindex
phpaction=the_science_behind_bo_tau rang (accessed 270812)[3] Epilepsy-SeizuresURL langhttpwwwhealthcommunitiescomepilepsy-sei
zuresabout-seizuresshtml rang (accessed 270812)[4] M Soleymani T Pun M Pantic Multi-modal emotion recognition in response
to videos IEEE Trans Affect Comput 3 (2) (2012) 211ndash223[5] S Scholler S Bosse MS Treder B Blankertz G Curio K-R Muumlller
T Wiegand Toward a direct measure of video quality perception using EEGIEEE Trans Image Process 21 (5) (2012) 2619ndash2629
[6] W Di C Zhihua F Ruifang L Guangyu and L Tian Study on human brainafter consuming alcohol based on EEG signal in Proceedings of the Interna-tional Conference on Computer Science and Information Technology (ICCSIT)5 406ndash409 2010
[7] E Estrada H Nazeran F Ebrahimi and M Mikaeili EEG signal features forcomputer-aided sleep stage detection in Proceedings of the IEEE EMBSConference on Neural Engineering 669ndash672 2009
[8] ZM Hana1047297ah KFM Yunos ZH Murat MN Taib S Lias EEG brainwavepattern for smoking behaviour after horizontal rotation treatment in Pro-ceedings of the IEEE Student Conference on Research and Development 2009
[9] ZH Murat R Shilawani SA Kadir RM Isa and MN Taib The effects of
mobile phone usage on human braingwave using EEG in Proceedings of the13th International Conference on Modelling and Simulation 2011
[10] K Lehnertz RG Andrzejak J Arnhold T Kreuz F Mormann C Rieke
G Widman CE Elger Nonliner EEG analysis in Epilepsy its possible use forinterictal focus localization seizure ancipation and prevention J ClinNeurophsysiol 18 (3) (2001) 209ndash222
[11] R Panda PS Khobragade PD Jambhule SN Jengthe PR Pal and TK GandhiClassi1047297cation of EEG signal using wavelet transform and support vectormachine for epileptic seizure diction in Proceedings of the InternationalConference on Systems in Medicine and Biology (ICSMB) 405ndash408 2010
[12] SG Dastidar H Adeli N Dadmehr Mixed-band wavelet chaos-neural net-
work methodology for epilepsy and epileptic seizure detection IEEE TransBiomeical Eng 54 (9) (2007) 1545ndash1551
[13] H Ocak Optimal classi1047297cation of epileptic seizures in EEG using waveletanalysis and genetic algorithm Signal Process 88 (2008) 1858ndash1867
[14] K Polat S Gunes Classi1047297cation of epileptiform EEG using a hybrid systembased on decision tree classi1047297er and fast Fourier transform Appl Math
Comput 187 (2) (2007) 1017ndash1026[15] SF Liang HC Wang WL Chang Combination of EEG complexity and spectral
analysis for epilepsy diagnosis and seizure detection EURASIP J Adv SignalProcess 853434 (2010) 1ndash15 (Article ID)
[16] RB Pachori Discrimination between ictal and seizure-free EEG signals usingempirical mode decomposition Res Lett Signal Process 293056 (2008) 1ndash5(Article ID)
[17] D Birvinskas V Jusas I Martisius and R Damasevicius EEG dataset reductionand feature extraction using discrete cosine transform in Proceedings of theEuropean symposium on computer modeling and simulation (EMS) 199-2042012
[18] V Bajaj RB Pachori Classi1047297cation of seizure and non-seizure EEG signalsusing empirical mode decomposition IEEE Trans Inf Technol Biomed 16 (6)(2012) 1135ndash1142
[19] Small dataset Epilepsy dataset (text format) langhttpepileptologie-bonndecmsfront_contentphpidcat=193amplang=3ampchangelang=3rang (accessed 280412)
[20] J Harrison Fast and accurate bessel function computation in Proceedings of the19th IEEE international symposium on computer arithmetic 104ndash113
2009[21] X Zhang W Diao Z Cheng Wavelet transform and singular value decom-
position of EEG signal for pattern recognition of complicated hand activitiesLect Notes Comput Sci Digit Hum Model 4561 (2007) 294ndash303
[22] Large dataset EEG Data set from Epilepsy Center of the University Hospital of
Freiburg langhttpepilepsyuni-freiburgdefreiburg-seizure-prediction-projecteeg-databaserang (accessed 100612)
[23] KD Brabanter P Karsmakers F Ojeda C Alzate and JD Brabanter KPelckmans B De Moor J Vandewalle JAK Suykens LS-SVMlab ToolboxUsers Guide version 18 Katholieke Universiteit Leuven 2011
[24] S Sanei JA Chambers EEG Signal Processing John Wiley amp Sons Ltd UK
2007 (Centre of Digital Signal Processing Cardiff University)[25] S Abe Support Vector Machine for Pattern Classi1047297cation Second Edition
Springer London Dordrecht Heidelberg New York 2010
[26] H Wang and D Hu Comparison of SVM and LS-SVM for regression inProceedings of the international conference on neural networks and brain 1279ndash283 2005
[27] H Xing M Ha B Hu D Tian Linear feature-weighted support vectormachine Fuzzy Inf Eng 1 (3) (2009) 289ndash305
[28] A Subasi EEG signal classi1047297cation using wavelet feature extraction and amixture of expert model Expert Syst Appl 32 (4) (2007) 1084ndash1093(Elsevier)
[29] JR Williamson DW Bliss DW Browne JT Narayanan Seizure predictionusing EEG spatiotemporal correlation structure Epilepsy Behav 25 (2) (2012)230ndash238
[30] TP Jung S Makeig C Humphries et al Removing electroencephalographicsartifacts by blind source separation Psychophysiology 37 (2000) 163ndash178
[31] W De Clercq A Vergult B Vanrumste WV Paesschen SV Huffel Canonicalcorrelation analysis applied to remove muscle artifacts from the electroence-phalogram IEEE Trans Biomeical Eng 53 (12) (2006) 2583ndash2587
[32] J Ma P Tao S Bayram V Svetnik Muscle artifacts in multichannel EEGcharacteristics and reduction ClinNeurophysiol 123 (8) (2012) 1676ndash1686
[33] GA Worrell L Parish SD Cranstoun R Jonas G Baltuch B Litt High-frequency oscillations and seizure generation in neocortical epilepsy Brain127 (2004) 1496ndash1506
[34] M Paul W Lin CT Lau BndashS Lee A long term reference frame for hierarchicalb-picture based video coding IEEE Trans Circuits Syst Video Technol (2014)httpdxdoiorg101109TCSVT20142302555
[35] Z Gu W Lin CT Lau BndashS Lee and M Paul Two dimensional singular valuedecomposition (2D-SVD) based video coding in Proceedings of the IEEEconference on image processing (IEEE ICIP-10) 181ndash184 2010
Mohammad Zavid Parvez received BScEng (hons)Degree in Computer Science and Engineering fromAsian University of Bangladesh in 2003 and MScDegree in Electrical Engineering with emphasis onSignal Processing from Blekinge Institute of TechnologySweden in 2010 Parvezs research interests includebiomedical signal processing and machine learning
He has more than three year industry experience asa software developer Moreover he was involved sev-eral software projects in Denmark Currently he is aPhD student and teaching staff in Charles Sturt Uni-versity
Parvez is a Graduate student member of IEEE He hasreceived full funded scholarship from the Faculty He has published several journaland conference papers in the area of EEG signal analysis and cognitive radio
Manoranjan Paul received BScEng (hons) Degree inComputer Science and Engineering from BangladeshUniversity of Engineering and Technology (BUET) Ban-gladesh in 1997 and PhD Degree from Monash Uni-versity Australia in 2005 He was an Assistant Professorin Ahsanullah University of Science and Technology Hewas a Post-Doctoral Research Fellow in the Universityof New South Wales in 2005ndash2006 Monash Universityin 2006ndash2009 and Nanyang Technological Universityin 2009ndash2011 He has joined in the School of Comput-ing and Mathematics Charles Sturt University (CSU) at2011 Currently he is a Senior Lecturer and AssociateDirector of the Center for Research in Complex Systems
(CRiCS) in CSUDr Paul is a senior member of the IEEE and ACS He is in editorial board of three
international journals including EURASIP Journal on Advances in Signal Processing(JASP) Dr Paul has served as a guest editor of Journal of Multimedia and Journal of Computers for 1047297ve special issues He has been in Technical Program Committees of more than 25 international conferences Dr Paul obtained more than A$1Mcompetitive grant money including Australian Research Council (ARC) DiscoveryProject He has published more than 65 refereed publications including journalsand conferences He organized a special session on ldquoNew video coding technolo-giesrdquo in IEEE ISCAS 2010 He was a keynote speaker on ldquoVision friendly videocodingrdquo in IEEE ICCIT 2010 and tutorial speaker on ldquoMultiview Video Codingrdquo inDICTA 2013 His major research interests are in the 1047297elds of video coding computervision and EEG Signals analysis Dr Paul received Vice Chancellor ResearchExcellence Award in Faculty level at 2013
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200200
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 511
the higher frequency component (ie the last quartile (Q4) ie last
25 of DCT coef 1047297cient) Then we extract statistical features such as
entropy and energy of the weighted higher frequency coef 1047297cients to
classify ictal and interictal signals
The extraction of the statistical features from the Q4 is not
straightforward We use heuristics to provide different weight to
the different coef 1047297cients and investigate the classi1047297cation accu-
racy Although we try different type of weighting approximations
the experimental results show that exponential weight performswell The weight w i for the i-th position coef 1047297cient is selected as
wi frac14 β e β xi where xi x j frac14 1
Number of Selected Coeff icients We also try
different β and test for the classi1047297cation results We 1047297nd that
β frac142 provides the best results for the dataset Fig 4 shows an
example of weights According to the proposed weight we provide
more weight to the relatively low frequency coef 1047297cients and less
weight to the relatively high frequency coef 1047297cients within the
higher frequency area of the overall signal as sometimes the high
frequency coef 1047297cients in the tail would be very small
We de1047297ne weighted entropy and energy as follows
Energy frac14 sumn
i frac14 1
W i X ietht THORN
2and Entropy frac14 sum
n
i frac14 1
W iY 2i etht THORNlnethY 2i etht THORNTHORN eth1THORN
where n is the length of signal X (t ) is the higher frequency DCTcoef 1047297cients and Y (t ) is the normalized value of X (t )
The representation of ictal and interictal data for different higher
frequency DCT coef 1047297cients and their entropy and energy are given in
Fig 3(a) It can be observed that the changing rate of frequencies is
higher for interictal signal than ictal signal from Temporal lobe It is
also interesting to observe that the amplitude of the higher frequency
DCT coef 1047297cients for ictal signals is larger compared to that of interictal
signals Note that although the original ictal and interictal EEG signals
are not visually separable the high frequency DCT coef 1047297cients of ictal
and interictal signals are visually separable The third row of Fig 3
(a) also demonstrates that the weighted energy and entropy extracted
from the higher frequency DCT components have different values for
ictal and interictal signals Thus weighted entropy and energy would
be effective features for ictal and interictal signals classi1047297
cation
33 Feature extraction using DCT ndashDWT
DWT is a transformation and widely used in the image and signal
processing research to decompose the imagesignals into different
frequency and bands DWT can be speci1047297ed by low pass (LP) and high
pass (HP) 1047297lters named as standard quadrature mirror 1047297lters [28] The
cut-off frequency of these 1047297lters is one-fourth of the sampling
frequency The bandwidth of the 1047297lter outputs is half the bandwidth
of the original signal which generates downsampled output signals
without losing any information according to the Nyquist theorem [13]
The down-sampled signals from LP and HP 1047297lters refer to 1047297rst-level
approximation and details of the original signal respectively [24] DWT
is sometimes more effective compared to DCT to separate signals into
low and high frequency components when the signal is longer Thus
we also investigate the effectiveness of the DWT to extract some other
statistical features (such as STD) to classify ictal and interictal signals
In our second approach we apply single-level one dimensional DWT
on the EEG signal to 1047297nd the high frequency components from the
signal Then we calculate the STD from the detail frequency (ie highfrequency) components of the DWT coef 1047297cients and use as a feature
We use the whole high frequency component for feature extraction as
DWT separates the signals into low and high frequency components
Another feature is extracted using the STD of the weighted high
frequency DCT coef 1047297cients (similar to Fig 4) Two STD features from
DCT and DWT higher frequency components are then used for the
classi1047297cation purpose Note that we use different primary functions of
the wavelet transformation including Daubechies (db) series We 1047297nd
that db16 is the best for our experiment
In the experiment we take only the high frequency DWT compo-
nents The 1047297rst two rows in Fig 3(b) show the higher change of
frequency and lower amplitude for interictal signal compared to ictal
signal from Temporal lobe The third row of Fig 3(b) represents the
STDs generated from the high frequency DCT coef 1047297cients (weighted)
and DWT for ictal which are higher than that of interictal signal Thus
the STDs extracted from the high frequency DCT (weighted) and DWT
coef 1047297cients can be good features for classi1047297cation
34 Feature extraction using SVD
SVD decomposes two dimensional signals into three matrixes in
such a way that the middle matrix (ie singular value) has unique
properties to separate signals The middle matrix containing the
square root of eigenvalues and diagonalized contains much less non-
zero coef 1047297cients compared with the coef 1047297cient matrices of other
transformations [35] The values of diagonal matrix are with des-
cending order from top-left to bottom-right As the original ictal and
interictal signals are almost of the same nature we do not expect asigni1047297cant different values of other two matrixes after SVD however
we expect some distinguishable values on the extreme part (ie
bottom-right) of diagonal matrix Thus any feature extracted from
the bottom-right singular values could be used to classify ictal and
interictal signals To the best of our knowledge we are the 1047297rst to use
two dimensional SVD to analyze EEG signals for seizure detection
To extract the features using SVD we reshape the EEG signals
into a two dimensional matrix The column vector of EEG signal is
rearranged to square matrix for the computation of singular value
components using SVD From the singular values we take the
weighted Q4 (weight is similar to Fig 4 but the number of selected
coef 1047297cients are different) of the non-zero diagonal values to
calculate the STDs and energy and treat as the two features to
classify ictal and interictal signals SVD allows transforming any
given matrix AAC mxn to diagonal form using unitary matrices ie
A frac14 U Σ V H where Amn is a matrix U mm and V nnare orthogonal
matrix and Σ mn matrix with non-negative diagonal entries
The columns of U are called left singular vectors and the rows
of V H are called right singular vectors If A is a rectangular m n
matrix of rank K then U will be m m and Σ will be
Σ frac14
λ1 0 ⋯ 0
0 λ2 ⋯ 0
⋮ ⋮ ⋱ ⋮
0 0 ⋯ λn
⋮ ⋮ ⋮
0 0 ⋯ 0
2666666664
3777777775
eth2THORN
where λ1Z λ 2Zhellip λn-1 Z λnFig 4 Weight for different coef 1047297cients of selected high frequency coef 1047297cients
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200194
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 611
The data distribution of non-zero singular values of the ictal
and interictal signals from the Frontal lobe is shown in Fig 5(a)
The top part of the 1047297gure clearly shows that the Q4 singular values
of interictal signals are larger than that of the ictal signals The
bottom part of Fig 5(a) shows the STD and energy extracted from
the weighted Q4 of non-zero singular values from ictal and
interictal signals respectively The 1047297gure exposes that the STDs
extracted from the ictal and interictal signals have distinguishable
values thus they can be used for ictal and interictal signal
classi1047297cation
35 Feature extraction of IMF using EMD
The main strength of EMD is its ability to analyze nonlinear and
non-stationary signals like EEG signals When we apply EMD on
a signal the signal is decomposed into a 1047297nite number of signals
into IMFs which are ordered from higher frequency component to
lower frequency component The number of IMFs in a signal
depends on the local characteristics of the signal rather than
pre-de1047297ned number of IMFs This feature makes each IMF crucial
to identify some characteristics of the original signal Since the
IMFs are the representations of different frequenciesamplitudes
of the original signal and the distinguishing features between ictal
and interictal EEG signals lie on the frequencyamplitude our
hypothesis is that any feature extracted from IMFs will be an
effective feature to classify ictal and interictal successfully We also
observe that the ictal and interictal signals are deferred in higherorder frequency components thus our other hypothesis is that
features extracted from the 1047297rst few IMFs should be enough to
classify ictal and interictal signals successfully Through the
decomposition we keep only few mode functions of the original
signals for the 1047297nal feature extraction however we are able to
keep the important mode functions for the classi1047297cation purpose
Experimental results reveal that our both hypothesis are correct as
extracted features from the 1047297rst IMF successfully classify ictal and
interictal signals (see Section 4) Thus IMFs can be used in
classi1047297cation applications where the original signals are non-
linear and non-stationary and distinguishing features exist on
the different frequenciesamplitudes
Each IMF satis1047297es two following conditions (i) the number of
extrema and the number of zero crossing are identical or differ at
most by one and (ii) the mean value between the upper and the
lower envelope is equal to zero at any time
The EMD algorithm can be summarized as follows
1 Extract the extrema (minima and maxima) of the signal xetht THORN
2 Interpolate between minima and maxima to obtain
ℓminetht THORN and ℓmaxetht THORN
3 Calculate local mean metht THORN frac14 ℓminetht THORN thornℓmaxetht THORNfrac12 =2
4 Extract the detail detht THORN frac14 xetht THORN metht THORN
5 Check detht THORN is an IMF according to the conditions which are
mentioned above If yes repeat the procedure from the step1 on the residual signal r etht THORN frac14 xetht THORN metht THORN If no replace xetht THORN with
detht THORN and repeat the procedure from step 1
The pictorial scenario of generating IMFs is shown in Fig 6 The
signal xetht THORNcan be represented as a combination of IMFs and
residual component
xetht THORN frac14 sumL
i frac14 1
dietht THORN thornr Letht THORN eth3THORN
where di and r L are the i-th IMF and L-th is residual signal
We observe that features extracted from the 1047297rst two IMFs
provide the best classi1047297cation results To reduce computational
time we consider only the 1047297rst IMF from the EMD and then
calculate STD of IMF to capture the variation of the signal Byignoring other IMFs we do not sacri1047297ce any signi1047297cant classi1047297cation
accuracy The strength of a signal is distributed in the frequency
domain relative to the strengths of other ambient signals is central
to the design of any 1047297lter intended to extract or suppress the signal
To 1047297nd the strength of the signal we apply power spectral density
(PSD) on the 1047297rst IMF Then determine the STD feature from
normalized PSD The calculation of PSD is as follows
P ethwiTHORN frac14 1
N DethwiTHORN 2
eth4THORN
where DethwiTHORNis the discrete Fourier transform of IMFdietht THORN Normalized
PSD can be de1047297ned as
pi frac14 P ethwiTHORN=sum
i
P ethwiTHORN eth5THORN
Fig 5 (a) Non-zero singular value of SVD for Frontal lobe The 1047297rst row shows that the non-zero singular value distribution for ictal and interictal signals from Frontal lobe
and the last row shows the STD and energy of the last 25 non-zero singular values of ictal and interictal EEG signals (b) The 1047297rst IMFs of ictal and interictal signals with the
extracted features the top two rows represent the 1047297rst IMF from the ictal and interictal signals of Temporal lobe bottom row represents STD of the 1047297rst IMF and STD of the
normalized PSD (power spectral density) of the same IMF of ictal and interictal EEG signals
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 195
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 711
In the experiment we extract two features such as STD of the
1047297rst IMF and STD of normalized PSD of the 1047297rst IMF From the 1047297rst
two rows in Fig 5(b) it can be observed that the 1047297rst IMF of ictal
signal carries the higher amplitude compared to the interictal
signal
The third row of the 1047297gure represents the STD of the 1047297rst IMF
and STD of PSD of the same IMF from using ictal and interictal
signals The STDs for both cases of interictal signal are signi1047297cantly
larger than that of ictal signal Thus the two new features
extracted by EMD method can be good features for classi1047297cation
Table 1 presents the summary of all corresponding features use
for ictal and interictal classi1047297cation by each method
36 Classi 1047297er
We extract a number of features from different transforma-
tionsdecompositions To classify ictal and interictal signals we
need to use a classi1047297er The goal of a classi1047297er is to 1047297nd patients
states such as ictal (class 1) and interictal (class 2) using machine
learning approaches with cross-validation The challenge is to 1047297nd
the mapping that generalizes from training sets to unseen test
sets For the cross-validation data are partitioned into training set
and test set In our experiment we use 4-fold cross validation
where each dataset was randomly divided into 4 splits where
three of them were used for training and the fourth one used for
testing
To classify the ictal and interictal signals we extract various
features from DCT DWT SVD and IMF of EMD For classi1047297cation we
use SVM-based classi1047297er as the SVM is one of the best classi1047297ers in
the EEG signal analysis SVM is a potential methodology for solving
problem in linear and nonlinear classi1047297cation function estimation
and kernel based learning methods [25] It can minimize the
operational error and maximize the margin hyperplane as a result
it will maximize the classi1047297cation performance [25] LS-SVM is the
extended version of SVM and it is closely related to regularization
networks and Gaussian process and it also has primal-dual inter-
pretations [23] The major drawback of SVM is in its higher
computational burden of the constrained optimization program-
ming However LS-SVM can solve this problem [26]
The classi1047297er is a LS-SVM which learns nonlinear mapping
from the training set features xifrac141hellipnT where nT is the number
of training features into the patients state ictal ( thorn1) and interictal( 1) For unbiased classi1047297cation results [27] we divide the whole
1000 trials into four subsets in each subset we randomly select
75 samples for the validation training set and the remaining 25
for the validation testing set Let f yigi frac14 C 12 designate the LS-SVM
validation test outputs mapping to class 1 or class 2 The equation
of LS-SVM can be de1047297ned in [18] as
f eth xTHORN frac14 sign sumN
i 1
α i yiK eth x xiTHORN thornc
eth6THORN
where K ( x xi) is a kernel function α i are the Lagrange multipliers c
is the bias term xi is the training input and yi is the training
output pairs Linear RBF and Morlet kernel are used in our
experiments and RBF and Morlet functions can be de1047297ned in [18]
Fig 6 Extracted IMFs by EMD from Temporal lobe (Patient 12 channel one 15th datablock for ictal and 35th datablock for interictal)
Table 1Extracted features of proposed and existing methods
Proposed methods Feature-1 Feature-2
Q4WHF_SVD (Quartile four weighted high frequency component using SVD ) STD on 25 weighted high frequency
singular values
Energy on 25 weighted high frequency
singular values
Q4WHF_DCT (Quartile four weighted high frequency component using DCT ) Energy on 25 weighted high
frequency DCT coef 1047297cients
Entropy on 25 weighted high frequency DCT
coef 1047297cients
FIMF_EMD (First IMF using EMD ) STD on the 1047297rst IMF (ie high
frequency components)
STD on normalized PSD
Q4WHF_DCT ndash HF_DWT (Quartile four weighted high frequency component
using DCT and high frequency from DWT)
STD on 25 weighted high frequency
DCT coef 1047297cients
STD on detail signals (ie high frequency
components) after apply DWT
Combined Combined all above features
Existing methods Feature-1 Feature-2
EMD[18] Amplitude modulation bandwidth Frequency modulation bandwidth
DCT [17] Features from 6 of quantized DCT coef 1047297cients at lower frequency
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200196
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 811
as
keth x xiTHORN frac14 expeth x xi2=2s
2THORN eth7THORN
where s controls the width of the RBF function
keth x xiTHORN frac14 prodd
k frac14 1
cos ω0etheth xk xki THORN=aTHORN
h iexpeth xk xk
i =2a2THORN eth8THORN
where d is the dimension a is the 1047298exible coef 1047297cient and xki is the
k-th component of i-th training data Matlab codes for LS-SVMclassi1047297er are available at httpwwwesatkuleuvenacbesista
lssvmlab
37 Computational time
Feature extractions using IMF through EMD take longer time
due to two-fold iterations to 1047297nd all IMFs In the 1047297rst iteration
EMD decomposes a signal and checks whether the signal makes an
IMF and in the second iteration it subtracts the current IMF from
the signal and continues for the next IMF Even if we stop the
iteration for the 1047297rst IMF the EMD-based technique takes a full
1047297rst iteration for calculating the IMF The experimental data
reveals that the proposed technique using DCT DWT and SVD
takes less computational time compared to the EMD-based tech-nique as DCT DWT and SVD transformation is faster compared to
the EMD decomposition Thus the proposed methods (except
FIMF_EMD method ie First IMF using EMD) take insigni1047297cant
time compared to the existing method [18]
4 Experimental results
Sometimes EEG signals have line noise and other kinds of
artifacts due to muscle and body movements Notch 1047297lter and
independent component analysis (ICA)-based method are recom-
mended to remove the line noise and artifacts respectively for
suf 1047297cient amount of EEG signals [30ndash32] However in this experi-
ment we do not use any explicit noiseartifacts removal technique
to see the tolerance of the proposed method against noise
artifacts
We propose four methods and the experimental results of the
proposed methods are compared with existing methods [1718]
and corresponding features are presented in Table 1Our ultimate
target is to classify ictal and interictal EEG signals using LS-SVM
with different kernels such as RBF and Morlet where the values of regularization parameter and kernel parameter are 10 and 6
respectively for both kernels We also use linear kernel In this
paper four methods are proposed based on the new features of
ictal and interictal EEG signals captured from Frontal and Temporal
lobes using DCT DWT SVD and IMF of EMD to see the strength of
individual transformationdecomposition techniques for seizure
detection Then we also test the classi1047297cation accuracy by combin-
ing all features In our experiment we use 200 signals from ictal
and 800 signals from interictal EEG signals After training and
testing of all features we calculate average value from the four
subsets to compute sensitivity speci1047297city and accuracy [18] The
sensitivity speci1047297city and accuracy are de1047297ned as
Sensitivity frac14 ethTP =ethTP thornFN THORNTHORN 100 eth9THORN
Specificity frac14 ethTN =ethTN thornFP THORNTHORN 100 eth10THORN
Accuracy frac14 ethethTP thornTN THORN=ethTP thornTN thornFP thorn FN THORNTHORN 100 eth11THORN
where TP and TN represent the total number of detected true
positive events and true negative events respectively The FP and
FN represent false positive and false negative respectively
The technique in [18] claims that the second IMF provides
better classi1047297cation results while they use frequency and ampli-
tude modulation bandwidth features for the dataset [19] We
conduct experiments using 1047297rst four IMFs however we provide
results using only the 1047297rst IMF presented in Table 2 as it carries the
Table 2
Sensitivity speci1047297city and accuracy for different features of ictal and interictal EEG signals from Frontal and Temporal lobes
Classi1047297er Brain location Criteria Existing method Proposed method
Kernel Lobe EMD[18] DCT[17] Q4WHF_SVD Q4WHF_DCT FIMF_EMD Q4WHF_DCT ndashHF_DWT Combined
RBF Frontal SEN 2941 870 6666 100 9583 7678 4914
SPE 8051 7973 8220 8130 8186 8337 8382
ACC 7790 7810 8150 8160 8220 8300 7980
Temporal SEN 6667 400 100 9824 9710 9747 9647
SPE 8028 9863 8351 8473 8571 8665 8710
ACC 8020 7970 8420 8550 8650 8750 8790
Average SEN 4804 635 8333 9912 9647 8713 7281
SPE 8040 8918 8286 8302 8379 8501 8546
ACC 7905 7890 8285 8355 8435 8525 8385
Morlet Frontal SEN 3382 000 6383 8889 9333 6000 2442
SPE 8101 7978 8216 8126 8112 8396 8042
ACC 7780 789 8130 8140 8130 8180 7560
Temporal SEN 8462 200 100 9841 9012 9368 9468
SPE 8354 9900 8430 8527 8618 8632 8775
ACC 8360 7960 8510 8610 8650 8690 8840
Average SEN 5922 100 8192 9365 9173 7684 5955
SPE 8228 8939 8323 8327 8365 8514 8409
ACC 8070 7925 8320 8375 8390 8435 8200
Linear Frontal SEN 2895 2728 8095 100 100 100 8750
SPE 8035 8061 8131 8089 8155 8081 8223
ACC 7840 766 8130 8110 8190 8100 8240
Temporal SEN 8421 600 9800 9722 9811 9800 9836
SPE 8528 9613 8411 8288 8437 8412 8509
ACC 8520 7810 8480 8340 8510 8480 8590
Average SEN 5658 1664 8948 9861 9906 9900 9293
SPE 8282 8837 8271 8189 8296 8247 8366
ACC 8180 7735 8305 8225 8350 8290 8415
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 197
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 911
best result compared to other IMFs For the RBF kernel function
the proposed methods outperform the state-of-the-art method
consistently and the average sensitivity using the proposed meth-
ods such as Quartile four weighted high frequency component using
SVD (Q4WHF_SVD) Quartile four weighted high frequency compo-
nent using DCT (Q4WHF_DCT) First IMF using EMD (FIMF_EMD)
and Quartile four weighted high frequency component using DCT and
high frequency from DWT (Q4WHF_DCTndashHF_DWT) are 8333
9912 9647 and 8713 for RBF kernel respectively whereas
the average sensitivity using the state-of-art-methods [17] and
[18] are 635 and 4804 respectively Using Morlet kernel the
average sensitivity of the proposed methods is higher compared to
the existing methods [17] and [18] Moreover we justify our result
applying linear kernel Table 2 also presents that in most of the
cases the proposed methods with linear kernel provide better
results compared to the proposed methods with Morlet and RBF
kernels in terms of sensitivity Table 2 also presents that the
proposed Q4WHF_DCT method is the best among all proposed
methods in terms of sensitivity It is also interesting to note that
the accuracy of the proposed Q4WHF_DCTndashHF_DWT method is
reasonably consistence in terms of three criteria for RBF and
Morlet kernel According to Table 2 all proposed methods always
outperform the existing methods [1718] in terms of sensitivity (i
e seizure and ictal detection rate) however in some cases the
existing methods [1718] outperform one or more proposed
methods in terms of speci1047297city (interictal detection rate) For
seizure detection sensitivity is more important compared to the
speci1047297city detection as the cost of failure detecting ictal is heavier
compared to interictal In overall classi1047297cation the proposed
methods are also better compared to the existing methods We
combine all features to classify ictal and interictal EEG signalsHowever combining is not good because LS-SVM may not prop-
erly combine all available features to make a good prediction
model The average sensitivity of the proposed methods is 9136
(average result of all kernel and lobe) whereas the sensitivities of
the state-of-the-arts methods are 5461 [18] and 799 [17]
respectively
The performance of the LS-SVM is evaluated by (ROC) plot
shown in Fig 7(a) ROC illustrates the performance of a binary
classi1047297er system where it is created by plotting the fraction of true
positives from the positives ie true positive rate (TPR) vs the
fraction of false positives from negatives ie false positive rate
(FPR) with various threshold settings TPR is known as sensitivity
and FPR is one minus the speci1047297city or true negative rate Fig 7(a)
shows that the proposed four methods are carrying good
classi1047297cation results than that of the state-of-the-art method
[18] using dataset from Frontal lobe of the third subset of training
and testing EEG signals
Fig 7(b) shows the distribution of normalized feature values
(ie energy amp entropy in the Q4WHF_DCT method STD in the
Q4WHF_DCTndashHF_DWT method STD amp energy in the Q4WHF_SVD
method and STD in the FIMF EMD method) of ictal and interictal
from the Frontal lobe dataset [22] The 1047297gure shows that the range
of feature values for ictal is different from that of the interictal
signals This phenomenon is the other evidence why we use those
features to classify ictal and interictal EEG signals in the proposed
methods
First column of Fig 8 shows classi1047297cation comparison using the
proposed methods (ie FIMF EMD and Q4WHF_DCTndashHF_DWT)
and the state-of-the-art method [18] Fig 8(b) and (c) shows the
LS-SVM classi1047297cation results using the proposed FIMF EMD and
Q4WHF_DCTndashHF_DWT against the state-of-the-art method (ie
EMD Fig 8(a)) We generate 1047297rst column of Fig 8 from Frontal lobe
for the second subset of training and testing and obtain the
classi1047297cation accuracy 9120 for the Q4WHF_DCTndashHF_DWT
method 8720 for the FIMF_EMD method and 8080 for the
EMD method It can be concluded from these 1047297gures that the
proposed method based on the DCTndashDWT and the IMF outperform
the state-of-the-art method (ie EMD) The same classi1047297cation
comparisons are also provided in second column of Fig 8 using the
state-of-the-art method based on the EMD and the proposed
method based on the SVD and DCT respectively We consider
Temporal lobe for the fourth subset of training and testing obtain
classi1047297cation accuracy 9120 for Q4WHF_SVD 8960 for
Q4WHF_DCT and 8520 for EMD [18]Thus it is concluded that
the proposed methods based on DCT and SVD also outperform theEMD method [18]
5 Conclusion
Unlike seizure and non-seizure EEG signals classi1047297cation of
ictal and interictal EEG signals is a challenging task due to non
abrupt phenomena (ie visually non-separable) between ictal and
interictal EEG signals Moreover the features of ictal and interictal
signals are not consistence in different brain locations patient
types of seizure and hospital setting during epileptic period of a
patient Thus the classi1047297cation performance for ictal and interictal
EEG signals using the existing features extraction and classi1047297cation
techniques is not at expected level Our initial observation
Fig 7 (a) The receiver operating characteristics (ROC) curves of the third subset of training and testing EEG signals by three proposed methods ( Q4WHF_SVD FIMF_EMD
and Q4WHF_DCTndashHF_DWT) against the EMD method using LS-SVM with RBF kernel from Frontal lobe and (b) Normalized distribution of feature values extracted from ictal
and interictal EEG signals using DCT DWT SVD and IMF based method where 200 samples are taken from each of ictal and interictal categories in Frontal lobe
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200198
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1011
indicates that high frequency components of EEG signals have
distinguished characteristics To improve the accuracy in this
paper we propose novel feature extractions strategies from
the high frequency components based on a number of estab-
lished transformation and decompositions The average sensi-
tivity (which is the measure for seizure) of the proposed
methods is 9136 (average result of all kernel and lobe)
whereas the sensitivity of the state-of-the-arts methods are
5461 [18] and 799 [17] respectively The experimental
results show that the proposed methods outperform the state-
of-the-art methods in terms of accuracy sensitivity and speci-
1047297city for the ictal and interictal EEG signals with greater
consistence from the large benchmark dataset in two different
brain locations
Fig 8 First column represents the classi1047297cation of ictal and interictal EEG signals from Frontal lobe for the second subset of training and testing using (a) the state-of-the-art
method against (b) the proposed IMF and (c) the DCT-DWT method Second column represents the classi1047297cation of ictal and interictal EEG signals from Temporal lobe for the
fourth subset of training and testing using (d) the state-of-the-art method against (e) the proposed SVD and (f) the DCT methods
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 199
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1111
References
[1] Neurology Applied How Science is Bringing Music Instruction Back toExpressive Development URL langhttpwwwthomasjwestmusiccomneurolo
gyappliedhtmrang (accessed 270812)[2] EEG Signal The science behind Bo-Tau URL langhttpwwwbo-taucoukindex
phpaction=the_science_behind_bo_tau rang (accessed 270812)[3] Epilepsy-SeizuresURL langhttpwwwhealthcommunitiescomepilepsy-sei
zuresabout-seizuresshtml rang (accessed 270812)[4] M Soleymani T Pun M Pantic Multi-modal emotion recognition in response
to videos IEEE Trans Affect Comput 3 (2) (2012) 211ndash223[5] S Scholler S Bosse MS Treder B Blankertz G Curio K-R Muumlller
T Wiegand Toward a direct measure of video quality perception using EEGIEEE Trans Image Process 21 (5) (2012) 2619ndash2629
[6] W Di C Zhihua F Ruifang L Guangyu and L Tian Study on human brainafter consuming alcohol based on EEG signal in Proceedings of the Interna-tional Conference on Computer Science and Information Technology (ICCSIT)5 406ndash409 2010
[7] E Estrada H Nazeran F Ebrahimi and M Mikaeili EEG signal features forcomputer-aided sleep stage detection in Proceedings of the IEEE EMBSConference on Neural Engineering 669ndash672 2009
[8] ZM Hana1047297ah KFM Yunos ZH Murat MN Taib S Lias EEG brainwavepattern for smoking behaviour after horizontal rotation treatment in Pro-ceedings of the IEEE Student Conference on Research and Development 2009
[9] ZH Murat R Shilawani SA Kadir RM Isa and MN Taib The effects of
mobile phone usage on human braingwave using EEG in Proceedings of the13th International Conference on Modelling and Simulation 2011
[10] K Lehnertz RG Andrzejak J Arnhold T Kreuz F Mormann C Rieke
G Widman CE Elger Nonliner EEG analysis in Epilepsy its possible use forinterictal focus localization seizure ancipation and prevention J ClinNeurophsysiol 18 (3) (2001) 209ndash222
[11] R Panda PS Khobragade PD Jambhule SN Jengthe PR Pal and TK GandhiClassi1047297cation of EEG signal using wavelet transform and support vectormachine for epileptic seizure diction in Proceedings of the InternationalConference on Systems in Medicine and Biology (ICSMB) 405ndash408 2010
[12] SG Dastidar H Adeli N Dadmehr Mixed-band wavelet chaos-neural net-
work methodology for epilepsy and epileptic seizure detection IEEE TransBiomeical Eng 54 (9) (2007) 1545ndash1551
[13] H Ocak Optimal classi1047297cation of epileptic seizures in EEG using waveletanalysis and genetic algorithm Signal Process 88 (2008) 1858ndash1867
[14] K Polat S Gunes Classi1047297cation of epileptiform EEG using a hybrid systembased on decision tree classi1047297er and fast Fourier transform Appl Math
Comput 187 (2) (2007) 1017ndash1026[15] SF Liang HC Wang WL Chang Combination of EEG complexity and spectral
analysis for epilepsy diagnosis and seizure detection EURASIP J Adv SignalProcess 853434 (2010) 1ndash15 (Article ID)
[16] RB Pachori Discrimination between ictal and seizure-free EEG signals usingempirical mode decomposition Res Lett Signal Process 293056 (2008) 1ndash5(Article ID)
[17] D Birvinskas V Jusas I Martisius and R Damasevicius EEG dataset reductionand feature extraction using discrete cosine transform in Proceedings of theEuropean symposium on computer modeling and simulation (EMS) 199-2042012
[18] V Bajaj RB Pachori Classi1047297cation of seizure and non-seizure EEG signalsusing empirical mode decomposition IEEE Trans Inf Technol Biomed 16 (6)(2012) 1135ndash1142
[19] Small dataset Epilepsy dataset (text format) langhttpepileptologie-bonndecmsfront_contentphpidcat=193amplang=3ampchangelang=3rang (accessed 280412)
[20] J Harrison Fast and accurate bessel function computation in Proceedings of the19th IEEE international symposium on computer arithmetic 104ndash113
2009[21] X Zhang W Diao Z Cheng Wavelet transform and singular value decom-
position of EEG signal for pattern recognition of complicated hand activitiesLect Notes Comput Sci Digit Hum Model 4561 (2007) 294ndash303
[22] Large dataset EEG Data set from Epilepsy Center of the University Hospital of
Freiburg langhttpepilepsyuni-freiburgdefreiburg-seizure-prediction-projecteeg-databaserang (accessed 100612)
[23] KD Brabanter P Karsmakers F Ojeda C Alzate and JD Brabanter KPelckmans B De Moor J Vandewalle JAK Suykens LS-SVMlab ToolboxUsers Guide version 18 Katholieke Universiteit Leuven 2011
[24] S Sanei JA Chambers EEG Signal Processing John Wiley amp Sons Ltd UK
2007 (Centre of Digital Signal Processing Cardiff University)[25] S Abe Support Vector Machine for Pattern Classi1047297cation Second Edition
Springer London Dordrecht Heidelberg New York 2010
[26] H Wang and D Hu Comparison of SVM and LS-SVM for regression inProceedings of the international conference on neural networks and brain 1279ndash283 2005
[27] H Xing M Ha B Hu D Tian Linear feature-weighted support vectormachine Fuzzy Inf Eng 1 (3) (2009) 289ndash305
[28] A Subasi EEG signal classi1047297cation using wavelet feature extraction and amixture of expert model Expert Syst Appl 32 (4) (2007) 1084ndash1093(Elsevier)
[29] JR Williamson DW Bliss DW Browne JT Narayanan Seizure predictionusing EEG spatiotemporal correlation structure Epilepsy Behav 25 (2) (2012)230ndash238
[30] TP Jung S Makeig C Humphries et al Removing electroencephalographicsartifacts by blind source separation Psychophysiology 37 (2000) 163ndash178
[31] W De Clercq A Vergult B Vanrumste WV Paesschen SV Huffel Canonicalcorrelation analysis applied to remove muscle artifacts from the electroence-phalogram IEEE Trans Biomeical Eng 53 (12) (2006) 2583ndash2587
[32] J Ma P Tao S Bayram V Svetnik Muscle artifacts in multichannel EEGcharacteristics and reduction ClinNeurophysiol 123 (8) (2012) 1676ndash1686
[33] GA Worrell L Parish SD Cranstoun R Jonas G Baltuch B Litt High-frequency oscillations and seizure generation in neocortical epilepsy Brain127 (2004) 1496ndash1506
[34] M Paul W Lin CT Lau BndashS Lee A long term reference frame for hierarchicalb-picture based video coding IEEE Trans Circuits Syst Video Technol (2014)httpdxdoiorg101109TCSVT20142302555
[35] Z Gu W Lin CT Lau BndashS Lee and M Paul Two dimensional singular valuedecomposition (2D-SVD) based video coding in Proceedings of the IEEEconference on image processing (IEEE ICIP-10) 181ndash184 2010
Mohammad Zavid Parvez received BScEng (hons)Degree in Computer Science and Engineering fromAsian University of Bangladesh in 2003 and MScDegree in Electrical Engineering with emphasis onSignal Processing from Blekinge Institute of TechnologySweden in 2010 Parvezs research interests includebiomedical signal processing and machine learning
He has more than three year industry experience asa software developer Moreover he was involved sev-eral software projects in Denmark Currently he is aPhD student and teaching staff in Charles Sturt Uni-versity
Parvez is a Graduate student member of IEEE He hasreceived full funded scholarship from the Faculty He has published several journaland conference papers in the area of EEG signal analysis and cognitive radio
Manoranjan Paul received BScEng (hons) Degree inComputer Science and Engineering from BangladeshUniversity of Engineering and Technology (BUET) Ban-gladesh in 1997 and PhD Degree from Monash Uni-versity Australia in 2005 He was an Assistant Professorin Ahsanullah University of Science and Technology Hewas a Post-Doctoral Research Fellow in the Universityof New South Wales in 2005ndash2006 Monash Universityin 2006ndash2009 and Nanyang Technological Universityin 2009ndash2011 He has joined in the School of Comput-ing and Mathematics Charles Sturt University (CSU) at2011 Currently he is a Senior Lecturer and AssociateDirector of the Center for Research in Complex Systems
(CRiCS) in CSUDr Paul is a senior member of the IEEE and ACS He is in editorial board of three
international journals including EURASIP Journal on Advances in Signal Processing(JASP) Dr Paul has served as a guest editor of Journal of Multimedia and Journal of Computers for 1047297ve special issues He has been in Technical Program Committees of more than 25 international conferences Dr Paul obtained more than A$1Mcompetitive grant money including Australian Research Council (ARC) DiscoveryProject He has published more than 65 refereed publications including journalsand conferences He organized a special session on ldquoNew video coding technolo-giesrdquo in IEEE ISCAS 2010 He was a keynote speaker on ldquoVision friendly videocodingrdquo in IEEE ICCIT 2010 and tutorial speaker on ldquoMultiview Video Codingrdquo inDICTA 2013 His major research interests are in the 1047297elds of video coding computervision and EEG Signals analysis Dr Paul received Vice Chancellor ResearchExcellence Award in Faculty level at 2013
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200200
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 611
The data distribution of non-zero singular values of the ictal
and interictal signals from the Frontal lobe is shown in Fig 5(a)
The top part of the 1047297gure clearly shows that the Q4 singular values
of interictal signals are larger than that of the ictal signals The
bottom part of Fig 5(a) shows the STD and energy extracted from
the weighted Q4 of non-zero singular values from ictal and
interictal signals respectively The 1047297gure exposes that the STDs
extracted from the ictal and interictal signals have distinguishable
values thus they can be used for ictal and interictal signal
classi1047297cation
35 Feature extraction of IMF using EMD
The main strength of EMD is its ability to analyze nonlinear and
non-stationary signals like EEG signals When we apply EMD on
a signal the signal is decomposed into a 1047297nite number of signals
into IMFs which are ordered from higher frequency component to
lower frequency component The number of IMFs in a signal
depends on the local characteristics of the signal rather than
pre-de1047297ned number of IMFs This feature makes each IMF crucial
to identify some characteristics of the original signal Since the
IMFs are the representations of different frequenciesamplitudes
of the original signal and the distinguishing features between ictal
and interictal EEG signals lie on the frequencyamplitude our
hypothesis is that any feature extracted from IMFs will be an
effective feature to classify ictal and interictal successfully We also
observe that the ictal and interictal signals are deferred in higherorder frequency components thus our other hypothesis is that
features extracted from the 1047297rst few IMFs should be enough to
classify ictal and interictal signals successfully Through the
decomposition we keep only few mode functions of the original
signals for the 1047297nal feature extraction however we are able to
keep the important mode functions for the classi1047297cation purpose
Experimental results reveal that our both hypothesis are correct as
extracted features from the 1047297rst IMF successfully classify ictal and
interictal signals (see Section 4) Thus IMFs can be used in
classi1047297cation applications where the original signals are non-
linear and non-stationary and distinguishing features exist on
the different frequenciesamplitudes
Each IMF satis1047297es two following conditions (i) the number of
extrema and the number of zero crossing are identical or differ at
most by one and (ii) the mean value between the upper and the
lower envelope is equal to zero at any time
The EMD algorithm can be summarized as follows
1 Extract the extrema (minima and maxima) of the signal xetht THORN
2 Interpolate between minima and maxima to obtain
ℓminetht THORN and ℓmaxetht THORN
3 Calculate local mean metht THORN frac14 ℓminetht THORN thornℓmaxetht THORNfrac12 =2
4 Extract the detail detht THORN frac14 xetht THORN metht THORN
5 Check detht THORN is an IMF according to the conditions which are
mentioned above If yes repeat the procedure from the step1 on the residual signal r etht THORN frac14 xetht THORN metht THORN If no replace xetht THORN with
detht THORN and repeat the procedure from step 1
The pictorial scenario of generating IMFs is shown in Fig 6 The
signal xetht THORNcan be represented as a combination of IMFs and
residual component
xetht THORN frac14 sumL
i frac14 1
dietht THORN thornr Letht THORN eth3THORN
where di and r L are the i-th IMF and L-th is residual signal
We observe that features extracted from the 1047297rst two IMFs
provide the best classi1047297cation results To reduce computational
time we consider only the 1047297rst IMF from the EMD and then
calculate STD of IMF to capture the variation of the signal Byignoring other IMFs we do not sacri1047297ce any signi1047297cant classi1047297cation
accuracy The strength of a signal is distributed in the frequency
domain relative to the strengths of other ambient signals is central
to the design of any 1047297lter intended to extract or suppress the signal
To 1047297nd the strength of the signal we apply power spectral density
(PSD) on the 1047297rst IMF Then determine the STD feature from
normalized PSD The calculation of PSD is as follows
P ethwiTHORN frac14 1
N DethwiTHORN 2
eth4THORN
where DethwiTHORNis the discrete Fourier transform of IMFdietht THORN Normalized
PSD can be de1047297ned as
pi frac14 P ethwiTHORN=sum
i
P ethwiTHORN eth5THORN
Fig 5 (a) Non-zero singular value of SVD for Frontal lobe The 1047297rst row shows that the non-zero singular value distribution for ictal and interictal signals from Frontal lobe
and the last row shows the STD and energy of the last 25 non-zero singular values of ictal and interictal EEG signals (b) The 1047297rst IMFs of ictal and interictal signals with the
extracted features the top two rows represent the 1047297rst IMF from the ictal and interictal signals of Temporal lobe bottom row represents STD of the 1047297rst IMF and STD of the
normalized PSD (power spectral density) of the same IMF of ictal and interictal EEG signals
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 195
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 711
In the experiment we extract two features such as STD of the
1047297rst IMF and STD of normalized PSD of the 1047297rst IMF From the 1047297rst
two rows in Fig 5(b) it can be observed that the 1047297rst IMF of ictal
signal carries the higher amplitude compared to the interictal
signal
The third row of the 1047297gure represents the STD of the 1047297rst IMF
and STD of PSD of the same IMF from using ictal and interictal
signals The STDs for both cases of interictal signal are signi1047297cantly
larger than that of ictal signal Thus the two new features
extracted by EMD method can be good features for classi1047297cation
Table 1 presents the summary of all corresponding features use
for ictal and interictal classi1047297cation by each method
36 Classi 1047297er
We extract a number of features from different transforma-
tionsdecompositions To classify ictal and interictal signals we
need to use a classi1047297er The goal of a classi1047297er is to 1047297nd patients
states such as ictal (class 1) and interictal (class 2) using machine
learning approaches with cross-validation The challenge is to 1047297nd
the mapping that generalizes from training sets to unseen test
sets For the cross-validation data are partitioned into training set
and test set In our experiment we use 4-fold cross validation
where each dataset was randomly divided into 4 splits where
three of them were used for training and the fourth one used for
testing
To classify the ictal and interictal signals we extract various
features from DCT DWT SVD and IMF of EMD For classi1047297cation we
use SVM-based classi1047297er as the SVM is one of the best classi1047297ers in
the EEG signal analysis SVM is a potential methodology for solving
problem in linear and nonlinear classi1047297cation function estimation
and kernel based learning methods [25] It can minimize the
operational error and maximize the margin hyperplane as a result
it will maximize the classi1047297cation performance [25] LS-SVM is the
extended version of SVM and it is closely related to regularization
networks and Gaussian process and it also has primal-dual inter-
pretations [23] The major drawback of SVM is in its higher
computational burden of the constrained optimization program-
ming However LS-SVM can solve this problem [26]
The classi1047297er is a LS-SVM which learns nonlinear mapping
from the training set features xifrac141hellipnT where nT is the number
of training features into the patients state ictal ( thorn1) and interictal( 1) For unbiased classi1047297cation results [27] we divide the whole
1000 trials into four subsets in each subset we randomly select
75 samples for the validation training set and the remaining 25
for the validation testing set Let f yigi frac14 C 12 designate the LS-SVM
validation test outputs mapping to class 1 or class 2 The equation
of LS-SVM can be de1047297ned in [18] as
f eth xTHORN frac14 sign sumN
i 1
α i yiK eth x xiTHORN thornc
eth6THORN
where K ( x xi) is a kernel function α i are the Lagrange multipliers c
is the bias term xi is the training input and yi is the training
output pairs Linear RBF and Morlet kernel are used in our
experiments and RBF and Morlet functions can be de1047297ned in [18]
Fig 6 Extracted IMFs by EMD from Temporal lobe (Patient 12 channel one 15th datablock for ictal and 35th datablock for interictal)
Table 1Extracted features of proposed and existing methods
Proposed methods Feature-1 Feature-2
Q4WHF_SVD (Quartile four weighted high frequency component using SVD ) STD on 25 weighted high frequency
singular values
Energy on 25 weighted high frequency
singular values
Q4WHF_DCT (Quartile four weighted high frequency component using DCT ) Energy on 25 weighted high
frequency DCT coef 1047297cients
Entropy on 25 weighted high frequency DCT
coef 1047297cients
FIMF_EMD (First IMF using EMD ) STD on the 1047297rst IMF (ie high
frequency components)
STD on normalized PSD
Q4WHF_DCT ndash HF_DWT (Quartile four weighted high frequency component
using DCT and high frequency from DWT)
STD on 25 weighted high frequency
DCT coef 1047297cients
STD on detail signals (ie high frequency
components) after apply DWT
Combined Combined all above features
Existing methods Feature-1 Feature-2
EMD[18] Amplitude modulation bandwidth Frequency modulation bandwidth
DCT [17] Features from 6 of quantized DCT coef 1047297cients at lower frequency
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200196
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 811
as
keth x xiTHORN frac14 expeth x xi2=2s
2THORN eth7THORN
where s controls the width of the RBF function
keth x xiTHORN frac14 prodd
k frac14 1
cos ω0etheth xk xki THORN=aTHORN
h iexpeth xk xk
i =2a2THORN eth8THORN
where d is the dimension a is the 1047298exible coef 1047297cient and xki is the
k-th component of i-th training data Matlab codes for LS-SVMclassi1047297er are available at httpwwwesatkuleuvenacbesista
lssvmlab
37 Computational time
Feature extractions using IMF through EMD take longer time
due to two-fold iterations to 1047297nd all IMFs In the 1047297rst iteration
EMD decomposes a signal and checks whether the signal makes an
IMF and in the second iteration it subtracts the current IMF from
the signal and continues for the next IMF Even if we stop the
iteration for the 1047297rst IMF the EMD-based technique takes a full
1047297rst iteration for calculating the IMF The experimental data
reveals that the proposed technique using DCT DWT and SVD
takes less computational time compared to the EMD-based tech-nique as DCT DWT and SVD transformation is faster compared to
the EMD decomposition Thus the proposed methods (except
FIMF_EMD method ie First IMF using EMD) take insigni1047297cant
time compared to the existing method [18]
4 Experimental results
Sometimes EEG signals have line noise and other kinds of
artifacts due to muscle and body movements Notch 1047297lter and
independent component analysis (ICA)-based method are recom-
mended to remove the line noise and artifacts respectively for
suf 1047297cient amount of EEG signals [30ndash32] However in this experi-
ment we do not use any explicit noiseartifacts removal technique
to see the tolerance of the proposed method against noise
artifacts
We propose four methods and the experimental results of the
proposed methods are compared with existing methods [1718]
and corresponding features are presented in Table 1Our ultimate
target is to classify ictal and interictal EEG signals using LS-SVM
with different kernels such as RBF and Morlet where the values of regularization parameter and kernel parameter are 10 and 6
respectively for both kernels We also use linear kernel In this
paper four methods are proposed based on the new features of
ictal and interictal EEG signals captured from Frontal and Temporal
lobes using DCT DWT SVD and IMF of EMD to see the strength of
individual transformationdecomposition techniques for seizure
detection Then we also test the classi1047297cation accuracy by combin-
ing all features In our experiment we use 200 signals from ictal
and 800 signals from interictal EEG signals After training and
testing of all features we calculate average value from the four
subsets to compute sensitivity speci1047297city and accuracy [18] The
sensitivity speci1047297city and accuracy are de1047297ned as
Sensitivity frac14 ethTP =ethTP thornFN THORNTHORN 100 eth9THORN
Specificity frac14 ethTN =ethTN thornFP THORNTHORN 100 eth10THORN
Accuracy frac14 ethethTP thornTN THORN=ethTP thornTN thornFP thorn FN THORNTHORN 100 eth11THORN
where TP and TN represent the total number of detected true
positive events and true negative events respectively The FP and
FN represent false positive and false negative respectively
The technique in [18] claims that the second IMF provides
better classi1047297cation results while they use frequency and ampli-
tude modulation bandwidth features for the dataset [19] We
conduct experiments using 1047297rst four IMFs however we provide
results using only the 1047297rst IMF presented in Table 2 as it carries the
Table 2
Sensitivity speci1047297city and accuracy for different features of ictal and interictal EEG signals from Frontal and Temporal lobes
Classi1047297er Brain location Criteria Existing method Proposed method
Kernel Lobe EMD[18] DCT[17] Q4WHF_SVD Q4WHF_DCT FIMF_EMD Q4WHF_DCT ndashHF_DWT Combined
RBF Frontal SEN 2941 870 6666 100 9583 7678 4914
SPE 8051 7973 8220 8130 8186 8337 8382
ACC 7790 7810 8150 8160 8220 8300 7980
Temporal SEN 6667 400 100 9824 9710 9747 9647
SPE 8028 9863 8351 8473 8571 8665 8710
ACC 8020 7970 8420 8550 8650 8750 8790
Average SEN 4804 635 8333 9912 9647 8713 7281
SPE 8040 8918 8286 8302 8379 8501 8546
ACC 7905 7890 8285 8355 8435 8525 8385
Morlet Frontal SEN 3382 000 6383 8889 9333 6000 2442
SPE 8101 7978 8216 8126 8112 8396 8042
ACC 7780 789 8130 8140 8130 8180 7560
Temporal SEN 8462 200 100 9841 9012 9368 9468
SPE 8354 9900 8430 8527 8618 8632 8775
ACC 8360 7960 8510 8610 8650 8690 8840
Average SEN 5922 100 8192 9365 9173 7684 5955
SPE 8228 8939 8323 8327 8365 8514 8409
ACC 8070 7925 8320 8375 8390 8435 8200
Linear Frontal SEN 2895 2728 8095 100 100 100 8750
SPE 8035 8061 8131 8089 8155 8081 8223
ACC 7840 766 8130 8110 8190 8100 8240
Temporal SEN 8421 600 9800 9722 9811 9800 9836
SPE 8528 9613 8411 8288 8437 8412 8509
ACC 8520 7810 8480 8340 8510 8480 8590
Average SEN 5658 1664 8948 9861 9906 9900 9293
SPE 8282 8837 8271 8189 8296 8247 8366
ACC 8180 7735 8305 8225 8350 8290 8415
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 197
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 911
best result compared to other IMFs For the RBF kernel function
the proposed methods outperform the state-of-the-art method
consistently and the average sensitivity using the proposed meth-
ods such as Quartile four weighted high frequency component using
SVD (Q4WHF_SVD) Quartile four weighted high frequency compo-
nent using DCT (Q4WHF_DCT) First IMF using EMD (FIMF_EMD)
and Quartile four weighted high frequency component using DCT and
high frequency from DWT (Q4WHF_DCTndashHF_DWT) are 8333
9912 9647 and 8713 for RBF kernel respectively whereas
the average sensitivity using the state-of-art-methods [17] and
[18] are 635 and 4804 respectively Using Morlet kernel the
average sensitivity of the proposed methods is higher compared to
the existing methods [17] and [18] Moreover we justify our result
applying linear kernel Table 2 also presents that in most of the
cases the proposed methods with linear kernel provide better
results compared to the proposed methods with Morlet and RBF
kernels in terms of sensitivity Table 2 also presents that the
proposed Q4WHF_DCT method is the best among all proposed
methods in terms of sensitivity It is also interesting to note that
the accuracy of the proposed Q4WHF_DCTndashHF_DWT method is
reasonably consistence in terms of three criteria for RBF and
Morlet kernel According to Table 2 all proposed methods always
outperform the existing methods [1718] in terms of sensitivity (i
e seizure and ictal detection rate) however in some cases the
existing methods [1718] outperform one or more proposed
methods in terms of speci1047297city (interictal detection rate) For
seizure detection sensitivity is more important compared to the
speci1047297city detection as the cost of failure detecting ictal is heavier
compared to interictal In overall classi1047297cation the proposed
methods are also better compared to the existing methods We
combine all features to classify ictal and interictal EEG signalsHowever combining is not good because LS-SVM may not prop-
erly combine all available features to make a good prediction
model The average sensitivity of the proposed methods is 9136
(average result of all kernel and lobe) whereas the sensitivities of
the state-of-the-arts methods are 5461 [18] and 799 [17]
respectively
The performance of the LS-SVM is evaluated by (ROC) plot
shown in Fig 7(a) ROC illustrates the performance of a binary
classi1047297er system where it is created by plotting the fraction of true
positives from the positives ie true positive rate (TPR) vs the
fraction of false positives from negatives ie false positive rate
(FPR) with various threshold settings TPR is known as sensitivity
and FPR is one minus the speci1047297city or true negative rate Fig 7(a)
shows that the proposed four methods are carrying good
classi1047297cation results than that of the state-of-the-art method
[18] using dataset from Frontal lobe of the third subset of training
and testing EEG signals
Fig 7(b) shows the distribution of normalized feature values
(ie energy amp entropy in the Q4WHF_DCT method STD in the
Q4WHF_DCTndashHF_DWT method STD amp energy in the Q4WHF_SVD
method and STD in the FIMF EMD method) of ictal and interictal
from the Frontal lobe dataset [22] The 1047297gure shows that the range
of feature values for ictal is different from that of the interictal
signals This phenomenon is the other evidence why we use those
features to classify ictal and interictal EEG signals in the proposed
methods
First column of Fig 8 shows classi1047297cation comparison using the
proposed methods (ie FIMF EMD and Q4WHF_DCTndashHF_DWT)
and the state-of-the-art method [18] Fig 8(b) and (c) shows the
LS-SVM classi1047297cation results using the proposed FIMF EMD and
Q4WHF_DCTndashHF_DWT against the state-of-the-art method (ie
EMD Fig 8(a)) We generate 1047297rst column of Fig 8 from Frontal lobe
for the second subset of training and testing and obtain the
classi1047297cation accuracy 9120 for the Q4WHF_DCTndashHF_DWT
method 8720 for the FIMF_EMD method and 8080 for the
EMD method It can be concluded from these 1047297gures that the
proposed method based on the DCTndashDWT and the IMF outperform
the state-of-the-art method (ie EMD) The same classi1047297cation
comparisons are also provided in second column of Fig 8 using the
state-of-the-art method based on the EMD and the proposed
method based on the SVD and DCT respectively We consider
Temporal lobe for the fourth subset of training and testing obtain
classi1047297cation accuracy 9120 for Q4WHF_SVD 8960 for
Q4WHF_DCT and 8520 for EMD [18]Thus it is concluded that
the proposed methods based on DCT and SVD also outperform theEMD method [18]
5 Conclusion
Unlike seizure and non-seizure EEG signals classi1047297cation of
ictal and interictal EEG signals is a challenging task due to non
abrupt phenomena (ie visually non-separable) between ictal and
interictal EEG signals Moreover the features of ictal and interictal
signals are not consistence in different brain locations patient
types of seizure and hospital setting during epileptic period of a
patient Thus the classi1047297cation performance for ictal and interictal
EEG signals using the existing features extraction and classi1047297cation
techniques is not at expected level Our initial observation
Fig 7 (a) The receiver operating characteristics (ROC) curves of the third subset of training and testing EEG signals by three proposed methods ( Q4WHF_SVD FIMF_EMD
and Q4WHF_DCTndashHF_DWT) against the EMD method using LS-SVM with RBF kernel from Frontal lobe and (b) Normalized distribution of feature values extracted from ictal
and interictal EEG signals using DCT DWT SVD and IMF based method where 200 samples are taken from each of ictal and interictal categories in Frontal lobe
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200198
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1011
indicates that high frequency components of EEG signals have
distinguished characteristics To improve the accuracy in this
paper we propose novel feature extractions strategies from
the high frequency components based on a number of estab-
lished transformation and decompositions The average sensi-
tivity (which is the measure for seizure) of the proposed
methods is 9136 (average result of all kernel and lobe)
whereas the sensitivity of the state-of-the-arts methods are
5461 [18] and 799 [17] respectively The experimental
results show that the proposed methods outperform the state-
of-the-art methods in terms of accuracy sensitivity and speci-
1047297city for the ictal and interictal EEG signals with greater
consistence from the large benchmark dataset in two different
brain locations
Fig 8 First column represents the classi1047297cation of ictal and interictal EEG signals from Frontal lobe for the second subset of training and testing using (a) the state-of-the-art
method against (b) the proposed IMF and (c) the DCT-DWT method Second column represents the classi1047297cation of ictal and interictal EEG signals from Temporal lobe for the
fourth subset of training and testing using (d) the state-of-the-art method against (e) the proposed SVD and (f) the DCT methods
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 199
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1111
References
[1] Neurology Applied How Science is Bringing Music Instruction Back toExpressive Development URL langhttpwwwthomasjwestmusiccomneurolo
gyappliedhtmrang (accessed 270812)[2] EEG Signal The science behind Bo-Tau URL langhttpwwwbo-taucoukindex
phpaction=the_science_behind_bo_tau rang (accessed 270812)[3] Epilepsy-SeizuresURL langhttpwwwhealthcommunitiescomepilepsy-sei
zuresabout-seizuresshtml rang (accessed 270812)[4] M Soleymani T Pun M Pantic Multi-modal emotion recognition in response
to videos IEEE Trans Affect Comput 3 (2) (2012) 211ndash223[5] S Scholler S Bosse MS Treder B Blankertz G Curio K-R Muumlller
T Wiegand Toward a direct measure of video quality perception using EEGIEEE Trans Image Process 21 (5) (2012) 2619ndash2629
[6] W Di C Zhihua F Ruifang L Guangyu and L Tian Study on human brainafter consuming alcohol based on EEG signal in Proceedings of the Interna-tional Conference on Computer Science and Information Technology (ICCSIT)5 406ndash409 2010
[7] E Estrada H Nazeran F Ebrahimi and M Mikaeili EEG signal features forcomputer-aided sleep stage detection in Proceedings of the IEEE EMBSConference on Neural Engineering 669ndash672 2009
[8] ZM Hana1047297ah KFM Yunos ZH Murat MN Taib S Lias EEG brainwavepattern for smoking behaviour after horizontal rotation treatment in Pro-ceedings of the IEEE Student Conference on Research and Development 2009
[9] ZH Murat R Shilawani SA Kadir RM Isa and MN Taib The effects of
mobile phone usage on human braingwave using EEG in Proceedings of the13th International Conference on Modelling and Simulation 2011
[10] K Lehnertz RG Andrzejak J Arnhold T Kreuz F Mormann C Rieke
G Widman CE Elger Nonliner EEG analysis in Epilepsy its possible use forinterictal focus localization seizure ancipation and prevention J ClinNeurophsysiol 18 (3) (2001) 209ndash222
[11] R Panda PS Khobragade PD Jambhule SN Jengthe PR Pal and TK GandhiClassi1047297cation of EEG signal using wavelet transform and support vectormachine for epileptic seizure diction in Proceedings of the InternationalConference on Systems in Medicine and Biology (ICSMB) 405ndash408 2010
[12] SG Dastidar H Adeli N Dadmehr Mixed-band wavelet chaos-neural net-
work methodology for epilepsy and epileptic seizure detection IEEE TransBiomeical Eng 54 (9) (2007) 1545ndash1551
[13] H Ocak Optimal classi1047297cation of epileptic seizures in EEG using waveletanalysis and genetic algorithm Signal Process 88 (2008) 1858ndash1867
[14] K Polat S Gunes Classi1047297cation of epileptiform EEG using a hybrid systembased on decision tree classi1047297er and fast Fourier transform Appl Math
Comput 187 (2) (2007) 1017ndash1026[15] SF Liang HC Wang WL Chang Combination of EEG complexity and spectral
analysis for epilepsy diagnosis and seizure detection EURASIP J Adv SignalProcess 853434 (2010) 1ndash15 (Article ID)
[16] RB Pachori Discrimination between ictal and seizure-free EEG signals usingempirical mode decomposition Res Lett Signal Process 293056 (2008) 1ndash5(Article ID)
[17] D Birvinskas V Jusas I Martisius and R Damasevicius EEG dataset reductionand feature extraction using discrete cosine transform in Proceedings of theEuropean symposium on computer modeling and simulation (EMS) 199-2042012
[18] V Bajaj RB Pachori Classi1047297cation of seizure and non-seizure EEG signalsusing empirical mode decomposition IEEE Trans Inf Technol Biomed 16 (6)(2012) 1135ndash1142
[19] Small dataset Epilepsy dataset (text format) langhttpepileptologie-bonndecmsfront_contentphpidcat=193amplang=3ampchangelang=3rang (accessed 280412)
[20] J Harrison Fast and accurate bessel function computation in Proceedings of the19th IEEE international symposium on computer arithmetic 104ndash113
2009[21] X Zhang W Diao Z Cheng Wavelet transform and singular value decom-
position of EEG signal for pattern recognition of complicated hand activitiesLect Notes Comput Sci Digit Hum Model 4561 (2007) 294ndash303
[22] Large dataset EEG Data set from Epilepsy Center of the University Hospital of
Freiburg langhttpepilepsyuni-freiburgdefreiburg-seizure-prediction-projecteeg-databaserang (accessed 100612)
[23] KD Brabanter P Karsmakers F Ojeda C Alzate and JD Brabanter KPelckmans B De Moor J Vandewalle JAK Suykens LS-SVMlab ToolboxUsers Guide version 18 Katholieke Universiteit Leuven 2011
[24] S Sanei JA Chambers EEG Signal Processing John Wiley amp Sons Ltd UK
2007 (Centre of Digital Signal Processing Cardiff University)[25] S Abe Support Vector Machine for Pattern Classi1047297cation Second Edition
Springer London Dordrecht Heidelberg New York 2010
[26] H Wang and D Hu Comparison of SVM and LS-SVM for regression inProceedings of the international conference on neural networks and brain 1279ndash283 2005
[27] H Xing M Ha B Hu D Tian Linear feature-weighted support vectormachine Fuzzy Inf Eng 1 (3) (2009) 289ndash305
[28] A Subasi EEG signal classi1047297cation using wavelet feature extraction and amixture of expert model Expert Syst Appl 32 (4) (2007) 1084ndash1093(Elsevier)
[29] JR Williamson DW Bliss DW Browne JT Narayanan Seizure predictionusing EEG spatiotemporal correlation structure Epilepsy Behav 25 (2) (2012)230ndash238
[30] TP Jung S Makeig C Humphries et al Removing electroencephalographicsartifacts by blind source separation Psychophysiology 37 (2000) 163ndash178
[31] W De Clercq A Vergult B Vanrumste WV Paesschen SV Huffel Canonicalcorrelation analysis applied to remove muscle artifacts from the electroence-phalogram IEEE Trans Biomeical Eng 53 (12) (2006) 2583ndash2587
[32] J Ma P Tao S Bayram V Svetnik Muscle artifacts in multichannel EEGcharacteristics and reduction ClinNeurophysiol 123 (8) (2012) 1676ndash1686
[33] GA Worrell L Parish SD Cranstoun R Jonas G Baltuch B Litt High-frequency oscillations and seizure generation in neocortical epilepsy Brain127 (2004) 1496ndash1506
[34] M Paul W Lin CT Lau BndashS Lee A long term reference frame for hierarchicalb-picture based video coding IEEE Trans Circuits Syst Video Technol (2014)httpdxdoiorg101109TCSVT20142302555
[35] Z Gu W Lin CT Lau BndashS Lee and M Paul Two dimensional singular valuedecomposition (2D-SVD) based video coding in Proceedings of the IEEEconference on image processing (IEEE ICIP-10) 181ndash184 2010
Mohammad Zavid Parvez received BScEng (hons)Degree in Computer Science and Engineering fromAsian University of Bangladesh in 2003 and MScDegree in Electrical Engineering with emphasis onSignal Processing from Blekinge Institute of TechnologySweden in 2010 Parvezs research interests includebiomedical signal processing and machine learning
He has more than three year industry experience asa software developer Moreover he was involved sev-eral software projects in Denmark Currently he is aPhD student and teaching staff in Charles Sturt Uni-versity
Parvez is a Graduate student member of IEEE He hasreceived full funded scholarship from the Faculty He has published several journaland conference papers in the area of EEG signal analysis and cognitive radio
Manoranjan Paul received BScEng (hons) Degree inComputer Science and Engineering from BangladeshUniversity of Engineering and Technology (BUET) Ban-gladesh in 1997 and PhD Degree from Monash Uni-versity Australia in 2005 He was an Assistant Professorin Ahsanullah University of Science and Technology Hewas a Post-Doctoral Research Fellow in the Universityof New South Wales in 2005ndash2006 Monash Universityin 2006ndash2009 and Nanyang Technological Universityin 2009ndash2011 He has joined in the School of Comput-ing and Mathematics Charles Sturt University (CSU) at2011 Currently he is a Senior Lecturer and AssociateDirector of the Center for Research in Complex Systems
(CRiCS) in CSUDr Paul is a senior member of the IEEE and ACS He is in editorial board of three
international journals including EURASIP Journal on Advances in Signal Processing(JASP) Dr Paul has served as a guest editor of Journal of Multimedia and Journal of Computers for 1047297ve special issues He has been in Technical Program Committees of more than 25 international conferences Dr Paul obtained more than A$1Mcompetitive grant money including Australian Research Council (ARC) DiscoveryProject He has published more than 65 refereed publications including journalsand conferences He organized a special session on ldquoNew video coding technolo-giesrdquo in IEEE ISCAS 2010 He was a keynote speaker on ldquoVision friendly videocodingrdquo in IEEE ICCIT 2010 and tutorial speaker on ldquoMultiview Video Codingrdquo inDICTA 2013 His major research interests are in the 1047297elds of video coding computervision and EEG Signals analysis Dr Paul received Vice Chancellor ResearchExcellence Award in Faculty level at 2013
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200200
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 711
In the experiment we extract two features such as STD of the
1047297rst IMF and STD of normalized PSD of the 1047297rst IMF From the 1047297rst
two rows in Fig 5(b) it can be observed that the 1047297rst IMF of ictal
signal carries the higher amplitude compared to the interictal
signal
The third row of the 1047297gure represents the STD of the 1047297rst IMF
and STD of PSD of the same IMF from using ictal and interictal
signals The STDs for both cases of interictal signal are signi1047297cantly
larger than that of ictal signal Thus the two new features
extracted by EMD method can be good features for classi1047297cation
Table 1 presents the summary of all corresponding features use
for ictal and interictal classi1047297cation by each method
36 Classi 1047297er
We extract a number of features from different transforma-
tionsdecompositions To classify ictal and interictal signals we
need to use a classi1047297er The goal of a classi1047297er is to 1047297nd patients
states such as ictal (class 1) and interictal (class 2) using machine
learning approaches with cross-validation The challenge is to 1047297nd
the mapping that generalizes from training sets to unseen test
sets For the cross-validation data are partitioned into training set
and test set In our experiment we use 4-fold cross validation
where each dataset was randomly divided into 4 splits where
three of them were used for training and the fourth one used for
testing
To classify the ictal and interictal signals we extract various
features from DCT DWT SVD and IMF of EMD For classi1047297cation we
use SVM-based classi1047297er as the SVM is one of the best classi1047297ers in
the EEG signal analysis SVM is a potential methodology for solving
problem in linear and nonlinear classi1047297cation function estimation
and kernel based learning methods [25] It can minimize the
operational error and maximize the margin hyperplane as a result
it will maximize the classi1047297cation performance [25] LS-SVM is the
extended version of SVM and it is closely related to regularization
networks and Gaussian process and it also has primal-dual inter-
pretations [23] The major drawback of SVM is in its higher
computational burden of the constrained optimization program-
ming However LS-SVM can solve this problem [26]
The classi1047297er is a LS-SVM which learns nonlinear mapping
from the training set features xifrac141hellipnT where nT is the number
of training features into the patients state ictal ( thorn1) and interictal( 1) For unbiased classi1047297cation results [27] we divide the whole
1000 trials into four subsets in each subset we randomly select
75 samples for the validation training set and the remaining 25
for the validation testing set Let f yigi frac14 C 12 designate the LS-SVM
validation test outputs mapping to class 1 or class 2 The equation
of LS-SVM can be de1047297ned in [18] as
f eth xTHORN frac14 sign sumN
i 1
α i yiK eth x xiTHORN thornc
eth6THORN
where K ( x xi) is a kernel function α i are the Lagrange multipliers c
is the bias term xi is the training input and yi is the training
output pairs Linear RBF and Morlet kernel are used in our
experiments and RBF and Morlet functions can be de1047297ned in [18]
Fig 6 Extracted IMFs by EMD from Temporal lobe (Patient 12 channel one 15th datablock for ictal and 35th datablock for interictal)
Table 1Extracted features of proposed and existing methods
Proposed methods Feature-1 Feature-2
Q4WHF_SVD (Quartile four weighted high frequency component using SVD ) STD on 25 weighted high frequency
singular values
Energy on 25 weighted high frequency
singular values
Q4WHF_DCT (Quartile four weighted high frequency component using DCT ) Energy on 25 weighted high
frequency DCT coef 1047297cients
Entropy on 25 weighted high frequency DCT
coef 1047297cients
FIMF_EMD (First IMF using EMD ) STD on the 1047297rst IMF (ie high
frequency components)
STD on normalized PSD
Q4WHF_DCT ndash HF_DWT (Quartile four weighted high frequency component
using DCT and high frequency from DWT)
STD on 25 weighted high frequency
DCT coef 1047297cients
STD on detail signals (ie high frequency
components) after apply DWT
Combined Combined all above features
Existing methods Feature-1 Feature-2
EMD[18] Amplitude modulation bandwidth Frequency modulation bandwidth
DCT [17] Features from 6 of quantized DCT coef 1047297cients at lower frequency
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200196
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 811
as
keth x xiTHORN frac14 expeth x xi2=2s
2THORN eth7THORN
where s controls the width of the RBF function
keth x xiTHORN frac14 prodd
k frac14 1
cos ω0etheth xk xki THORN=aTHORN
h iexpeth xk xk
i =2a2THORN eth8THORN
where d is the dimension a is the 1047298exible coef 1047297cient and xki is the
k-th component of i-th training data Matlab codes for LS-SVMclassi1047297er are available at httpwwwesatkuleuvenacbesista
lssvmlab
37 Computational time
Feature extractions using IMF through EMD take longer time
due to two-fold iterations to 1047297nd all IMFs In the 1047297rst iteration
EMD decomposes a signal and checks whether the signal makes an
IMF and in the second iteration it subtracts the current IMF from
the signal and continues for the next IMF Even if we stop the
iteration for the 1047297rst IMF the EMD-based technique takes a full
1047297rst iteration for calculating the IMF The experimental data
reveals that the proposed technique using DCT DWT and SVD
takes less computational time compared to the EMD-based tech-nique as DCT DWT and SVD transformation is faster compared to
the EMD decomposition Thus the proposed methods (except
FIMF_EMD method ie First IMF using EMD) take insigni1047297cant
time compared to the existing method [18]
4 Experimental results
Sometimes EEG signals have line noise and other kinds of
artifacts due to muscle and body movements Notch 1047297lter and
independent component analysis (ICA)-based method are recom-
mended to remove the line noise and artifacts respectively for
suf 1047297cient amount of EEG signals [30ndash32] However in this experi-
ment we do not use any explicit noiseartifacts removal technique
to see the tolerance of the proposed method against noise
artifacts
We propose four methods and the experimental results of the
proposed methods are compared with existing methods [1718]
and corresponding features are presented in Table 1Our ultimate
target is to classify ictal and interictal EEG signals using LS-SVM
with different kernels such as RBF and Morlet where the values of regularization parameter and kernel parameter are 10 and 6
respectively for both kernels We also use linear kernel In this
paper four methods are proposed based on the new features of
ictal and interictal EEG signals captured from Frontal and Temporal
lobes using DCT DWT SVD and IMF of EMD to see the strength of
individual transformationdecomposition techniques for seizure
detection Then we also test the classi1047297cation accuracy by combin-
ing all features In our experiment we use 200 signals from ictal
and 800 signals from interictal EEG signals After training and
testing of all features we calculate average value from the four
subsets to compute sensitivity speci1047297city and accuracy [18] The
sensitivity speci1047297city and accuracy are de1047297ned as
Sensitivity frac14 ethTP =ethTP thornFN THORNTHORN 100 eth9THORN
Specificity frac14 ethTN =ethTN thornFP THORNTHORN 100 eth10THORN
Accuracy frac14 ethethTP thornTN THORN=ethTP thornTN thornFP thorn FN THORNTHORN 100 eth11THORN
where TP and TN represent the total number of detected true
positive events and true negative events respectively The FP and
FN represent false positive and false negative respectively
The technique in [18] claims that the second IMF provides
better classi1047297cation results while they use frequency and ampli-
tude modulation bandwidth features for the dataset [19] We
conduct experiments using 1047297rst four IMFs however we provide
results using only the 1047297rst IMF presented in Table 2 as it carries the
Table 2
Sensitivity speci1047297city and accuracy for different features of ictal and interictal EEG signals from Frontal and Temporal lobes
Classi1047297er Brain location Criteria Existing method Proposed method
Kernel Lobe EMD[18] DCT[17] Q4WHF_SVD Q4WHF_DCT FIMF_EMD Q4WHF_DCT ndashHF_DWT Combined
RBF Frontal SEN 2941 870 6666 100 9583 7678 4914
SPE 8051 7973 8220 8130 8186 8337 8382
ACC 7790 7810 8150 8160 8220 8300 7980
Temporal SEN 6667 400 100 9824 9710 9747 9647
SPE 8028 9863 8351 8473 8571 8665 8710
ACC 8020 7970 8420 8550 8650 8750 8790
Average SEN 4804 635 8333 9912 9647 8713 7281
SPE 8040 8918 8286 8302 8379 8501 8546
ACC 7905 7890 8285 8355 8435 8525 8385
Morlet Frontal SEN 3382 000 6383 8889 9333 6000 2442
SPE 8101 7978 8216 8126 8112 8396 8042
ACC 7780 789 8130 8140 8130 8180 7560
Temporal SEN 8462 200 100 9841 9012 9368 9468
SPE 8354 9900 8430 8527 8618 8632 8775
ACC 8360 7960 8510 8610 8650 8690 8840
Average SEN 5922 100 8192 9365 9173 7684 5955
SPE 8228 8939 8323 8327 8365 8514 8409
ACC 8070 7925 8320 8375 8390 8435 8200
Linear Frontal SEN 2895 2728 8095 100 100 100 8750
SPE 8035 8061 8131 8089 8155 8081 8223
ACC 7840 766 8130 8110 8190 8100 8240
Temporal SEN 8421 600 9800 9722 9811 9800 9836
SPE 8528 9613 8411 8288 8437 8412 8509
ACC 8520 7810 8480 8340 8510 8480 8590
Average SEN 5658 1664 8948 9861 9906 9900 9293
SPE 8282 8837 8271 8189 8296 8247 8366
ACC 8180 7735 8305 8225 8350 8290 8415
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 197
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 911
best result compared to other IMFs For the RBF kernel function
the proposed methods outperform the state-of-the-art method
consistently and the average sensitivity using the proposed meth-
ods such as Quartile four weighted high frequency component using
SVD (Q4WHF_SVD) Quartile four weighted high frequency compo-
nent using DCT (Q4WHF_DCT) First IMF using EMD (FIMF_EMD)
and Quartile four weighted high frequency component using DCT and
high frequency from DWT (Q4WHF_DCTndashHF_DWT) are 8333
9912 9647 and 8713 for RBF kernel respectively whereas
the average sensitivity using the state-of-art-methods [17] and
[18] are 635 and 4804 respectively Using Morlet kernel the
average sensitivity of the proposed methods is higher compared to
the existing methods [17] and [18] Moreover we justify our result
applying linear kernel Table 2 also presents that in most of the
cases the proposed methods with linear kernel provide better
results compared to the proposed methods with Morlet and RBF
kernels in terms of sensitivity Table 2 also presents that the
proposed Q4WHF_DCT method is the best among all proposed
methods in terms of sensitivity It is also interesting to note that
the accuracy of the proposed Q4WHF_DCTndashHF_DWT method is
reasonably consistence in terms of three criteria for RBF and
Morlet kernel According to Table 2 all proposed methods always
outperform the existing methods [1718] in terms of sensitivity (i
e seizure and ictal detection rate) however in some cases the
existing methods [1718] outperform one or more proposed
methods in terms of speci1047297city (interictal detection rate) For
seizure detection sensitivity is more important compared to the
speci1047297city detection as the cost of failure detecting ictal is heavier
compared to interictal In overall classi1047297cation the proposed
methods are also better compared to the existing methods We
combine all features to classify ictal and interictal EEG signalsHowever combining is not good because LS-SVM may not prop-
erly combine all available features to make a good prediction
model The average sensitivity of the proposed methods is 9136
(average result of all kernel and lobe) whereas the sensitivities of
the state-of-the-arts methods are 5461 [18] and 799 [17]
respectively
The performance of the LS-SVM is evaluated by (ROC) plot
shown in Fig 7(a) ROC illustrates the performance of a binary
classi1047297er system where it is created by plotting the fraction of true
positives from the positives ie true positive rate (TPR) vs the
fraction of false positives from negatives ie false positive rate
(FPR) with various threshold settings TPR is known as sensitivity
and FPR is one minus the speci1047297city or true negative rate Fig 7(a)
shows that the proposed four methods are carrying good
classi1047297cation results than that of the state-of-the-art method
[18] using dataset from Frontal lobe of the third subset of training
and testing EEG signals
Fig 7(b) shows the distribution of normalized feature values
(ie energy amp entropy in the Q4WHF_DCT method STD in the
Q4WHF_DCTndashHF_DWT method STD amp energy in the Q4WHF_SVD
method and STD in the FIMF EMD method) of ictal and interictal
from the Frontal lobe dataset [22] The 1047297gure shows that the range
of feature values for ictal is different from that of the interictal
signals This phenomenon is the other evidence why we use those
features to classify ictal and interictal EEG signals in the proposed
methods
First column of Fig 8 shows classi1047297cation comparison using the
proposed methods (ie FIMF EMD and Q4WHF_DCTndashHF_DWT)
and the state-of-the-art method [18] Fig 8(b) and (c) shows the
LS-SVM classi1047297cation results using the proposed FIMF EMD and
Q4WHF_DCTndashHF_DWT against the state-of-the-art method (ie
EMD Fig 8(a)) We generate 1047297rst column of Fig 8 from Frontal lobe
for the second subset of training and testing and obtain the
classi1047297cation accuracy 9120 for the Q4WHF_DCTndashHF_DWT
method 8720 for the FIMF_EMD method and 8080 for the
EMD method It can be concluded from these 1047297gures that the
proposed method based on the DCTndashDWT and the IMF outperform
the state-of-the-art method (ie EMD) The same classi1047297cation
comparisons are also provided in second column of Fig 8 using the
state-of-the-art method based on the EMD and the proposed
method based on the SVD and DCT respectively We consider
Temporal lobe for the fourth subset of training and testing obtain
classi1047297cation accuracy 9120 for Q4WHF_SVD 8960 for
Q4WHF_DCT and 8520 for EMD [18]Thus it is concluded that
the proposed methods based on DCT and SVD also outperform theEMD method [18]
5 Conclusion
Unlike seizure and non-seizure EEG signals classi1047297cation of
ictal and interictal EEG signals is a challenging task due to non
abrupt phenomena (ie visually non-separable) between ictal and
interictal EEG signals Moreover the features of ictal and interictal
signals are not consistence in different brain locations patient
types of seizure and hospital setting during epileptic period of a
patient Thus the classi1047297cation performance for ictal and interictal
EEG signals using the existing features extraction and classi1047297cation
techniques is not at expected level Our initial observation
Fig 7 (a) The receiver operating characteristics (ROC) curves of the third subset of training and testing EEG signals by three proposed methods ( Q4WHF_SVD FIMF_EMD
and Q4WHF_DCTndashHF_DWT) against the EMD method using LS-SVM with RBF kernel from Frontal lobe and (b) Normalized distribution of feature values extracted from ictal
and interictal EEG signals using DCT DWT SVD and IMF based method where 200 samples are taken from each of ictal and interictal categories in Frontal lobe
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200198
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1011
indicates that high frequency components of EEG signals have
distinguished characteristics To improve the accuracy in this
paper we propose novel feature extractions strategies from
the high frequency components based on a number of estab-
lished transformation and decompositions The average sensi-
tivity (which is the measure for seizure) of the proposed
methods is 9136 (average result of all kernel and lobe)
whereas the sensitivity of the state-of-the-arts methods are
5461 [18] and 799 [17] respectively The experimental
results show that the proposed methods outperform the state-
of-the-art methods in terms of accuracy sensitivity and speci-
1047297city for the ictal and interictal EEG signals with greater
consistence from the large benchmark dataset in two different
brain locations
Fig 8 First column represents the classi1047297cation of ictal and interictal EEG signals from Frontal lobe for the second subset of training and testing using (a) the state-of-the-art
method against (b) the proposed IMF and (c) the DCT-DWT method Second column represents the classi1047297cation of ictal and interictal EEG signals from Temporal lobe for the
fourth subset of training and testing using (d) the state-of-the-art method against (e) the proposed SVD and (f) the DCT methods
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 199
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1111
References
[1] Neurology Applied How Science is Bringing Music Instruction Back toExpressive Development URL langhttpwwwthomasjwestmusiccomneurolo
gyappliedhtmrang (accessed 270812)[2] EEG Signal The science behind Bo-Tau URL langhttpwwwbo-taucoukindex
phpaction=the_science_behind_bo_tau rang (accessed 270812)[3] Epilepsy-SeizuresURL langhttpwwwhealthcommunitiescomepilepsy-sei
zuresabout-seizuresshtml rang (accessed 270812)[4] M Soleymani T Pun M Pantic Multi-modal emotion recognition in response
to videos IEEE Trans Affect Comput 3 (2) (2012) 211ndash223[5] S Scholler S Bosse MS Treder B Blankertz G Curio K-R Muumlller
T Wiegand Toward a direct measure of video quality perception using EEGIEEE Trans Image Process 21 (5) (2012) 2619ndash2629
[6] W Di C Zhihua F Ruifang L Guangyu and L Tian Study on human brainafter consuming alcohol based on EEG signal in Proceedings of the Interna-tional Conference on Computer Science and Information Technology (ICCSIT)5 406ndash409 2010
[7] E Estrada H Nazeran F Ebrahimi and M Mikaeili EEG signal features forcomputer-aided sleep stage detection in Proceedings of the IEEE EMBSConference on Neural Engineering 669ndash672 2009
[8] ZM Hana1047297ah KFM Yunos ZH Murat MN Taib S Lias EEG brainwavepattern for smoking behaviour after horizontal rotation treatment in Pro-ceedings of the IEEE Student Conference on Research and Development 2009
[9] ZH Murat R Shilawani SA Kadir RM Isa and MN Taib The effects of
mobile phone usage on human braingwave using EEG in Proceedings of the13th International Conference on Modelling and Simulation 2011
[10] K Lehnertz RG Andrzejak J Arnhold T Kreuz F Mormann C Rieke
G Widman CE Elger Nonliner EEG analysis in Epilepsy its possible use forinterictal focus localization seizure ancipation and prevention J ClinNeurophsysiol 18 (3) (2001) 209ndash222
[11] R Panda PS Khobragade PD Jambhule SN Jengthe PR Pal and TK GandhiClassi1047297cation of EEG signal using wavelet transform and support vectormachine for epileptic seizure diction in Proceedings of the InternationalConference on Systems in Medicine and Biology (ICSMB) 405ndash408 2010
[12] SG Dastidar H Adeli N Dadmehr Mixed-band wavelet chaos-neural net-
work methodology for epilepsy and epileptic seizure detection IEEE TransBiomeical Eng 54 (9) (2007) 1545ndash1551
[13] H Ocak Optimal classi1047297cation of epileptic seizures in EEG using waveletanalysis and genetic algorithm Signal Process 88 (2008) 1858ndash1867
[14] K Polat S Gunes Classi1047297cation of epileptiform EEG using a hybrid systembased on decision tree classi1047297er and fast Fourier transform Appl Math
Comput 187 (2) (2007) 1017ndash1026[15] SF Liang HC Wang WL Chang Combination of EEG complexity and spectral
analysis for epilepsy diagnosis and seizure detection EURASIP J Adv SignalProcess 853434 (2010) 1ndash15 (Article ID)
[16] RB Pachori Discrimination between ictal and seizure-free EEG signals usingempirical mode decomposition Res Lett Signal Process 293056 (2008) 1ndash5(Article ID)
[17] D Birvinskas V Jusas I Martisius and R Damasevicius EEG dataset reductionand feature extraction using discrete cosine transform in Proceedings of theEuropean symposium on computer modeling and simulation (EMS) 199-2042012
[18] V Bajaj RB Pachori Classi1047297cation of seizure and non-seizure EEG signalsusing empirical mode decomposition IEEE Trans Inf Technol Biomed 16 (6)(2012) 1135ndash1142
[19] Small dataset Epilepsy dataset (text format) langhttpepileptologie-bonndecmsfront_contentphpidcat=193amplang=3ampchangelang=3rang (accessed 280412)
[20] J Harrison Fast and accurate bessel function computation in Proceedings of the19th IEEE international symposium on computer arithmetic 104ndash113
2009[21] X Zhang W Diao Z Cheng Wavelet transform and singular value decom-
position of EEG signal for pattern recognition of complicated hand activitiesLect Notes Comput Sci Digit Hum Model 4561 (2007) 294ndash303
[22] Large dataset EEG Data set from Epilepsy Center of the University Hospital of
Freiburg langhttpepilepsyuni-freiburgdefreiburg-seizure-prediction-projecteeg-databaserang (accessed 100612)
[23] KD Brabanter P Karsmakers F Ojeda C Alzate and JD Brabanter KPelckmans B De Moor J Vandewalle JAK Suykens LS-SVMlab ToolboxUsers Guide version 18 Katholieke Universiteit Leuven 2011
[24] S Sanei JA Chambers EEG Signal Processing John Wiley amp Sons Ltd UK
2007 (Centre of Digital Signal Processing Cardiff University)[25] S Abe Support Vector Machine for Pattern Classi1047297cation Second Edition
Springer London Dordrecht Heidelberg New York 2010
[26] H Wang and D Hu Comparison of SVM and LS-SVM for regression inProceedings of the international conference on neural networks and brain 1279ndash283 2005
[27] H Xing M Ha B Hu D Tian Linear feature-weighted support vectormachine Fuzzy Inf Eng 1 (3) (2009) 289ndash305
[28] A Subasi EEG signal classi1047297cation using wavelet feature extraction and amixture of expert model Expert Syst Appl 32 (4) (2007) 1084ndash1093(Elsevier)
[29] JR Williamson DW Bliss DW Browne JT Narayanan Seizure predictionusing EEG spatiotemporal correlation structure Epilepsy Behav 25 (2) (2012)230ndash238
[30] TP Jung S Makeig C Humphries et al Removing electroencephalographicsartifacts by blind source separation Psychophysiology 37 (2000) 163ndash178
[31] W De Clercq A Vergult B Vanrumste WV Paesschen SV Huffel Canonicalcorrelation analysis applied to remove muscle artifacts from the electroence-phalogram IEEE Trans Biomeical Eng 53 (12) (2006) 2583ndash2587
[32] J Ma P Tao S Bayram V Svetnik Muscle artifacts in multichannel EEGcharacteristics and reduction ClinNeurophysiol 123 (8) (2012) 1676ndash1686
[33] GA Worrell L Parish SD Cranstoun R Jonas G Baltuch B Litt High-frequency oscillations and seizure generation in neocortical epilepsy Brain127 (2004) 1496ndash1506
[34] M Paul W Lin CT Lau BndashS Lee A long term reference frame for hierarchicalb-picture based video coding IEEE Trans Circuits Syst Video Technol (2014)httpdxdoiorg101109TCSVT20142302555
[35] Z Gu W Lin CT Lau BndashS Lee and M Paul Two dimensional singular valuedecomposition (2D-SVD) based video coding in Proceedings of the IEEEconference on image processing (IEEE ICIP-10) 181ndash184 2010
Mohammad Zavid Parvez received BScEng (hons)Degree in Computer Science and Engineering fromAsian University of Bangladesh in 2003 and MScDegree in Electrical Engineering with emphasis onSignal Processing from Blekinge Institute of TechnologySweden in 2010 Parvezs research interests includebiomedical signal processing and machine learning
He has more than three year industry experience asa software developer Moreover he was involved sev-eral software projects in Denmark Currently he is aPhD student and teaching staff in Charles Sturt Uni-versity
Parvez is a Graduate student member of IEEE He hasreceived full funded scholarship from the Faculty He has published several journaland conference papers in the area of EEG signal analysis and cognitive radio
Manoranjan Paul received BScEng (hons) Degree inComputer Science and Engineering from BangladeshUniversity of Engineering and Technology (BUET) Ban-gladesh in 1997 and PhD Degree from Monash Uni-versity Australia in 2005 He was an Assistant Professorin Ahsanullah University of Science and Technology Hewas a Post-Doctoral Research Fellow in the Universityof New South Wales in 2005ndash2006 Monash Universityin 2006ndash2009 and Nanyang Technological Universityin 2009ndash2011 He has joined in the School of Comput-ing and Mathematics Charles Sturt University (CSU) at2011 Currently he is a Senior Lecturer and AssociateDirector of the Center for Research in Complex Systems
(CRiCS) in CSUDr Paul is a senior member of the IEEE and ACS He is in editorial board of three
international journals including EURASIP Journal on Advances in Signal Processing(JASP) Dr Paul has served as a guest editor of Journal of Multimedia and Journal of Computers for 1047297ve special issues He has been in Technical Program Committees of more than 25 international conferences Dr Paul obtained more than A$1Mcompetitive grant money including Australian Research Council (ARC) DiscoveryProject He has published more than 65 refereed publications including journalsand conferences He organized a special session on ldquoNew video coding technolo-giesrdquo in IEEE ISCAS 2010 He was a keynote speaker on ldquoVision friendly videocodingrdquo in IEEE ICCIT 2010 and tutorial speaker on ldquoMultiview Video Codingrdquo inDICTA 2013 His major research interests are in the 1047297elds of video coding computervision and EEG Signals analysis Dr Paul received Vice Chancellor ResearchExcellence Award in Faculty level at 2013
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200200
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 811
as
keth x xiTHORN frac14 expeth x xi2=2s
2THORN eth7THORN
where s controls the width of the RBF function
keth x xiTHORN frac14 prodd
k frac14 1
cos ω0etheth xk xki THORN=aTHORN
h iexpeth xk xk
i =2a2THORN eth8THORN
where d is the dimension a is the 1047298exible coef 1047297cient and xki is the
k-th component of i-th training data Matlab codes for LS-SVMclassi1047297er are available at httpwwwesatkuleuvenacbesista
lssvmlab
37 Computational time
Feature extractions using IMF through EMD take longer time
due to two-fold iterations to 1047297nd all IMFs In the 1047297rst iteration
EMD decomposes a signal and checks whether the signal makes an
IMF and in the second iteration it subtracts the current IMF from
the signal and continues for the next IMF Even if we stop the
iteration for the 1047297rst IMF the EMD-based technique takes a full
1047297rst iteration for calculating the IMF The experimental data
reveals that the proposed technique using DCT DWT and SVD
takes less computational time compared to the EMD-based tech-nique as DCT DWT and SVD transformation is faster compared to
the EMD decomposition Thus the proposed methods (except
FIMF_EMD method ie First IMF using EMD) take insigni1047297cant
time compared to the existing method [18]
4 Experimental results
Sometimes EEG signals have line noise and other kinds of
artifacts due to muscle and body movements Notch 1047297lter and
independent component analysis (ICA)-based method are recom-
mended to remove the line noise and artifacts respectively for
suf 1047297cient amount of EEG signals [30ndash32] However in this experi-
ment we do not use any explicit noiseartifacts removal technique
to see the tolerance of the proposed method against noise
artifacts
We propose four methods and the experimental results of the
proposed methods are compared with existing methods [1718]
and corresponding features are presented in Table 1Our ultimate
target is to classify ictal and interictal EEG signals using LS-SVM
with different kernels such as RBF and Morlet where the values of regularization parameter and kernel parameter are 10 and 6
respectively for both kernels We also use linear kernel In this
paper four methods are proposed based on the new features of
ictal and interictal EEG signals captured from Frontal and Temporal
lobes using DCT DWT SVD and IMF of EMD to see the strength of
individual transformationdecomposition techniques for seizure
detection Then we also test the classi1047297cation accuracy by combin-
ing all features In our experiment we use 200 signals from ictal
and 800 signals from interictal EEG signals After training and
testing of all features we calculate average value from the four
subsets to compute sensitivity speci1047297city and accuracy [18] The
sensitivity speci1047297city and accuracy are de1047297ned as
Sensitivity frac14 ethTP =ethTP thornFN THORNTHORN 100 eth9THORN
Specificity frac14 ethTN =ethTN thornFP THORNTHORN 100 eth10THORN
Accuracy frac14 ethethTP thornTN THORN=ethTP thornTN thornFP thorn FN THORNTHORN 100 eth11THORN
where TP and TN represent the total number of detected true
positive events and true negative events respectively The FP and
FN represent false positive and false negative respectively
The technique in [18] claims that the second IMF provides
better classi1047297cation results while they use frequency and ampli-
tude modulation bandwidth features for the dataset [19] We
conduct experiments using 1047297rst four IMFs however we provide
results using only the 1047297rst IMF presented in Table 2 as it carries the
Table 2
Sensitivity speci1047297city and accuracy for different features of ictal and interictal EEG signals from Frontal and Temporal lobes
Classi1047297er Brain location Criteria Existing method Proposed method
Kernel Lobe EMD[18] DCT[17] Q4WHF_SVD Q4WHF_DCT FIMF_EMD Q4WHF_DCT ndashHF_DWT Combined
RBF Frontal SEN 2941 870 6666 100 9583 7678 4914
SPE 8051 7973 8220 8130 8186 8337 8382
ACC 7790 7810 8150 8160 8220 8300 7980
Temporal SEN 6667 400 100 9824 9710 9747 9647
SPE 8028 9863 8351 8473 8571 8665 8710
ACC 8020 7970 8420 8550 8650 8750 8790
Average SEN 4804 635 8333 9912 9647 8713 7281
SPE 8040 8918 8286 8302 8379 8501 8546
ACC 7905 7890 8285 8355 8435 8525 8385
Morlet Frontal SEN 3382 000 6383 8889 9333 6000 2442
SPE 8101 7978 8216 8126 8112 8396 8042
ACC 7780 789 8130 8140 8130 8180 7560
Temporal SEN 8462 200 100 9841 9012 9368 9468
SPE 8354 9900 8430 8527 8618 8632 8775
ACC 8360 7960 8510 8610 8650 8690 8840
Average SEN 5922 100 8192 9365 9173 7684 5955
SPE 8228 8939 8323 8327 8365 8514 8409
ACC 8070 7925 8320 8375 8390 8435 8200
Linear Frontal SEN 2895 2728 8095 100 100 100 8750
SPE 8035 8061 8131 8089 8155 8081 8223
ACC 7840 766 8130 8110 8190 8100 8240
Temporal SEN 8421 600 9800 9722 9811 9800 9836
SPE 8528 9613 8411 8288 8437 8412 8509
ACC 8520 7810 8480 8340 8510 8480 8590
Average SEN 5658 1664 8948 9861 9906 9900 9293
SPE 8282 8837 8271 8189 8296 8247 8366
ACC 8180 7735 8305 8225 8350 8290 8415
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 197
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 911
best result compared to other IMFs For the RBF kernel function
the proposed methods outperform the state-of-the-art method
consistently and the average sensitivity using the proposed meth-
ods such as Quartile four weighted high frequency component using
SVD (Q4WHF_SVD) Quartile four weighted high frequency compo-
nent using DCT (Q4WHF_DCT) First IMF using EMD (FIMF_EMD)
and Quartile four weighted high frequency component using DCT and
high frequency from DWT (Q4WHF_DCTndashHF_DWT) are 8333
9912 9647 and 8713 for RBF kernel respectively whereas
the average sensitivity using the state-of-art-methods [17] and
[18] are 635 and 4804 respectively Using Morlet kernel the
average sensitivity of the proposed methods is higher compared to
the existing methods [17] and [18] Moreover we justify our result
applying linear kernel Table 2 also presents that in most of the
cases the proposed methods with linear kernel provide better
results compared to the proposed methods with Morlet and RBF
kernels in terms of sensitivity Table 2 also presents that the
proposed Q4WHF_DCT method is the best among all proposed
methods in terms of sensitivity It is also interesting to note that
the accuracy of the proposed Q4WHF_DCTndashHF_DWT method is
reasonably consistence in terms of three criteria for RBF and
Morlet kernel According to Table 2 all proposed methods always
outperform the existing methods [1718] in terms of sensitivity (i
e seizure and ictal detection rate) however in some cases the
existing methods [1718] outperform one or more proposed
methods in terms of speci1047297city (interictal detection rate) For
seizure detection sensitivity is more important compared to the
speci1047297city detection as the cost of failure detecting ictal is heavier
compared to interictal In overall classi1047297cation the proposed
methods are also better compared to the existing methods We
combine all features to classify ictal and interictal EEG signalsHowever combining is not good because LS-SVM may not prop-
erly combine all available features to make a good prediction
model The average sensitivity of the proposed methods is 9136
(average result of all kernel and lobe) whereas the sensitivities of
the state-of-the-arts methods are 5461 [18] and 799 [17]
respectively
The performance of the LS-SVM is evaluated by (ROC) plot
shown in Fig 7(a) ROC illustrates the performance of a binary
classi1047297er system where it is created by plotting the fraction of true
positives from the positives ie true positive rate (TPR) vs the
fraction of false positives from negatives ie false positive rate
(FPR) with various threshold settings TPR is known as sensitivity
and FPR is one minus the speci1047297city or true negative rate Fig 7(a)
shows that the proposed four methods are carrying good
classi1047297cation results than that of the state-of-the-art method
[18] using dataset from Frontal lobe of the third subset of training
and testing EEG signals
Fig 7(b) shows the distribution of normalized feature values
(ie energy amp entropy in the Q4WHF_DCT method STD in the
Q4WHF_DCTndashHF_DWT method STD amp energy in the Q4WHF_SVD
method and STD in the FIMF EMD method) of ictal and interictal
from the Frontal lobe dataset [22] The 1047297gure shows that the range
of feature values for ictal is different from that of the interictal
signals This phenomenon is the other evidence why we use those
features to classify ictal and interictal EEG signals in the proposed
methods
First column of Fig 8 shows classi1047297cation comparison using the
proposed methods (ie FIMF EMD and Q4WHF_DCTndashHF_DWT)
and the state-of-the-art method [18] Fig 8(b) and (c) shows the
LS-SVM classi1047297cation results using the proposed FIMF EMD and
Q4WHF_DCTndashHF_DWT against the state-of-the-art method (ie
EMD Fig 8(a)) We generate 1047297rst column of Fig 8 from Frontal lobe
for the second subset of training and testing and obtain the
classi1047297cation accuracy 9120 for the Q4WHF_DCTndashHF_DWT
method 8720 for the FIMF_EMD method and 8080 for the
EMD method It can be concluded from these 1047297gures that the
proposed method based on the DCTndashDWT and the IMF outperform
the state-of-the-art method (ie EMD) The same classi1047297cation
comparisons are also provided in second column of Fig 8 using the
state-of-the-art method based on the EMD and the proposed
method based on the SVD and DCT respectively We consider
Temporal lobe for the fourth subset of training and testing obtain
classi1047297cation accuracy 9120 for Q4WHF_SVD 8960 for
Q4WHF_DCT and 8520 for EMD [18]Thus it is concluded that
the proposed methods based on DCT and SVD also outperform theEMD method [18]
5 Conclusion
Unlike seizure and non-seizure EEG signals classi1047297cation of
ictal and interictal EEG signals is a challenging task due to non
abrupt phenomena (ie visually non-separable) between ictal and
interictal EEG signals Moreover the features of ictal and interictal
signals are not consistence in different brain locations patient
types of seizure and hospital setting during epileptic period of a
patient Thus the classi1047297cation performance for ictal and interictal
EEG signals using the existing features extraction and classi1047297cation
techniques is not at expected level Our initial observation
Fig 7 (a) The receiver operating characteristics (ROC) curves of the third subset of training and testing EEG signals by three proposed methods ( Q4WHF_SVD FIMF_EMD
and Q4WHF_DCTndashHF_DWT) against the EMD method using LS-SVM with RBF kernel from Frontal lobe and (b) Normalized distribution of feature values extracted from ictal
and interictal EEG signals using DCT DWT SVD and IMF based method where 200 samples are taken from each of ictal and interictal categories in Frontal lobe
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200198
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1011
indicates that high frequency components of EEG signals have
distinguished characteristics To improve the accuracy in this
paper we propose novel feature extractions strategies from
the high frequency components based on a number of estab-
lished transformation and decompositions The average sensi-
tivity (which is the measure for seizure) of the proposed
methods is 9136 (average result of all kernel and lobe)
whereas the sensitivity of the state-of-the-arts methods are
5461 [18] and 799 [17] respectively The experimental
results show that the proposed methods outperform the state-
of-the-art methods in terms of accuracy sensitivity and speci-
1047297city for the ictal and interictal EEG signals with greater
consistence from the large benchmark dataset in two different
brain locations
Fig 8 First column represents the classi1047297cation of ictal and interictal EEG signals from Frontal lobe for the second subset of training and testing using (a) the state-of-the-art
method against (b) the proposed IMF and (c) the DCT-DWT method Second column represents the classi1047297cation of ictal and interictal EEG signals from Temporal lobe for the
fourth subset of training and testing using (d) the state-of-the-art method against (e) the proposed SVD and (f) the DCT methods
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 199
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1111
References
[1] Neurology Applied How Science is Bringing Music Instruction Back toExpressive Development URL langhttpwwwthomasjwestmusiccomneurolo
gyappliedhtmrang (accessed 270812)[2] EEG Signal The science behind Bo-Tau URL langhttpwwwbo-taucoukindex
phpaction=the_science_behind_bo_tau rang (accessed 270812)[3] Epilepsy-SeizuresURL langhttpwwwhealthcommunitiescomepilepsy-sei
zuresabout-seizuresshtml rang (accessed 270812)[4] M Soleymani T Pun M Pantic Multi-modal emotion recognition in response
to videos IEEE Trans Affect Comput 3 (2) (2012) 211ndash223[5] S Scholler S Bosse MS Treder B Blankertz G Curio K-R Muumlller
T Wiegand Toward a direct measure of video quality perception using EEGIEEE Trans Image Process 21 (5) (2012) 2619ndash2629
[6] W Di C Zhihua F Ruifang L Guangyu and L Tian Study on human brainafter consuming alcohol based on EEG signal in Proceedings of the Interna-tional Conference on Computer Science and Information Technology (ICCSIT)5 406ndash409 2010
[7] E Estrada H Nazeran F Ebrahimi and M Mikaeili EEG signal features forcomputer-aided sleep stage detection in Proceedings of the IEEE EMBSConference on Neural Engineering 669ndash672 2009
[8] ZM Hana1047297ah KFM Yunos ZH Murat MN Taib S Lias EEG brainwavepattern for smoking behaviour after horizontal rotation treatment in Pro-ceedings of the IEEE Student Conference on Research and Development 2009
[9] ZH Murat R Shilawani SA Kadir RM Isa and MN Taib The effects of
mobile phone usage on human braingwave using EEG in Proceedings of the13th International Conference on Modelling and Simulation 2011
[10] K Lehnertz RG Andrzejak J Arnhold T Kreuz F Mormann C Rieke
G Widman CE Elger Nonliner EEG analysis in Epilepsy its possible use forinterictal focus localization seizure ancipation and prevention J ClinNeurophsysiol 18 (3) (2001) 209ndash222
[11] R Panda PS Khobragade PD Jambhule SN Jengthe PR Pal and TK GandhiClassi1047297cation of EEG signal using wavelet transform and support vectormachine for epileptic seizure diction in Proceedings of the InternationalConference on Systems in Medicine and Biology (ICSMB) 405ndash408 2010
[12] SG Dastidar H Adeli N Dadmehr Mixed-band wavelet chaos-neural net-
work methodology for epilepsy and epileptic seizure detection IEEE TransBiomeical Eng 54 (9) (2007) 1545ndash1551
[13] H Ocak Optimal classi1047297cation of epileptic seizures in EEG using waveletanalysis and genetic algorithm Signal Process 88 (2008) 1858ndash1867
[14] K Polat S Gunes Classi1047297cation of epileptiform EEG using a hybrid systembased on decision tree classi1047297er and fast Fourier transform Appl Math
Comput 187 (2) (2007) 1017ndash1026[15] SF Liang HC Wang WL Chang Combination of EEG complexity and spectral
analysis for epilepsy diagnosis and seizure detection EURASIP J Adv SignalProcess 853434 (2010) 1ndash15 (Article ID)
[16] RB Pachori Discrimination between ictal and seizure-free EEG signals usingempirical mode decomposition Res Lett Signal Process 293056 (2008) 1ndash5(Article ID)
[17] D Birvinskas V Jusas I Martisius and R Damasevicius EEG dataset reductionand feature extraction using discrete cosine transform in Proceedings of theEuropean symposium on computer modeling and simulation (EMS) 199-2042012
[18] V Bajaj RB Pachori Classi1047297cation of seizure and non-seizure EEG signalsusing empirical mode decomposition IEEE Trans Inf Technol Biomed 16 (6)(2012) 1135ndash1142
[19] Small dataset Epilepsy dataset (text format) langhttpepileptologie-bonndecmsfront_contentphpidcat=193amplang=3ampchangelang=3rang (accessed 280412)
[20] J Harrison Fast and accurate bessel function computation in Proceedings of the19th IEEE international symposium on computer arithmetic 104ndash113
2009[21] X Zhang W Diao Z Cheng Wavelet transform and singular value decom-
position of EEG signal for pattern recognition of complicated hand activitiesLect Notes Comput Sci Digit Hum Model 4561 (2007) 294ndash303
[22] Large dataset EEG Data set from Epilepsy Center of the University Hospital of
Freiburg langhttpepilepsyuni-freiburgdefreiburg-seizure-prediction-projecteeg-databaserang (accessed 100612)
[23] KD Brabanter P Karsmakers F Ojeda C Alzate and JD Brabanter KPelckmans B De Moor J Vandewalle JAK Suykens LS-SVMlab ToolboxUsers Guide version 18 Katholieke Universiteit Leuven 2011
[24] S Sanei JA Chambers EEG Signal Processing John Wiley amp Sons Ltd UK
2007 (Centre of Digital Signal Processing Cardiff University)[25] S Abe Support Vector Machine for Pattern Classi1047297cation Second Edition
Springer London Dordrecht Heidelberg New York 2010
[26] H Wang and D Hu Comparison of SVM and LS-SVM for regression inProceedings of the international conference on neural networks and brain 1279ndash283 2005
[27] H Xing M Ha B Hu D Tian Linear feature-weighted support vectormachine Fuzzy Inf Eng 1 (3) (2009) 289ndash305
[28] A Subasi EEG signal classi1047297cation using wavelet feature extraction and amixture of expert model Expert Syst Appl 32 (4) (2007) 1084ndash1093(Elsevier)
[29] JR Williamson DW Bliss DW Browne JT Narayanan Seizure predictionusing EEG spatiotemporal correlation structure Epilepsy Behav 25 (2) (2012)230ndash238
[30] TP Jung S Makeig C Humphries et al Removing electroencephalographicsartifacts by blind source separation Psychophysiology 37 (2000) 163ndash178
[31] W De Clercq A Vergult B Vanrumste WV Paesschen SV Huffel Canonicalcorrelation analysis applied to remove muscle artifacts from the electroence-phalogram IEEE Trans Biomeical Eng 53 (12) (2006) 2583ndash2587
[32] J Ma P Tao S Bayram V Svetnik Muscle artifacts in multichannel EEGcharacteristics and reduction ClinNeurophysiol 123 (8) (2012) 1676ndash1686
[33] GA Worrell L Parish SD Cranstoun R Jonas G Baltuch B Litt High-frequency oscillations and seizure generation in neocortical epilepsy Brain127 (2004) 1496ndash1506
[34] M Paul W Lin CT Lau BndashS Lee A long term reference frame for hierarchicalb-picture based video coding IEEE Trans Circuits Syst Video Technol (2014)httpdxdoiorg101109TCSVT20142302555
[35] Z Gu W Lin CT Lau BndashS Lee and M Paul Two dimensional singular valuedecomposition (2D-SVD) based video coding in Proceedings of the IEEEconference on image processing (IEEE ICIP-10) 181ndash184 2010
Mohammad Zavid Parvez received BScEng (hons)Degree in Computer Science and Engineering fromAsian University of Bangladesh in 2003 and MScDegree in Electrical Engineering with emphasis onSignal Processing from Blekinge Institute of TechnologySweden in 2010 Parvezs research interests includebiomedical signal processing and machine learning
He has more than three year industry experience asa software developer Moreover he was involved sev-eral software projects in Denmark Currently he is aPhD student and teaching staff in Charles Sturt Uni-versity
Parvez is a Graduate student member of IEEE He hasreceived full funded scholarship from the Faculty He has published several journaland conference papers in the area of EEG signal analysis and cognitive radio
Manoranjan Paul received BScEng (hons) Degree inComputer Science and Engineering from BangladeshUniversity of Engineering and Technology (BUET) Ban-gladesh in 1997 and PhD Degree from Monash Uni-versity Australia in 2005 He was an Assistant Professorin Ahsanullah University of Science and Technology Hewas a Post-Doctoral Research Fellow in the Universityof New South Wales in 2005ndash2006 Monash Universityin 2006ndash2009 and Nanyang Technological Universityin 2009ndash2011 He has joined in the School of Comput-ing and Mathematics Charles Sturt University (CSU) at2011 Currently he is a Senior Lecturer and AssociateDirector of the Center for Research in Complex Systems
(CRiCS) in CSUDr Paul is a senior member of the IEEE and ACS He is in editorial board of three
international journals including EURASIP Journal on Advances in Signal Processing(JASP) Dr Paul has served as a guest editor of Journal of Multimedia and Journal of Computers for 1047297ve special issues He has been in Technical Program Committees of more than 25 international conferences Dr Paul obtained more than A$1Mcompetitive grant money including Australian Research Council (ARC) DiscoveryProject He has published more than 65 refereed publications including journalsand conferences He organized a special session on ldquoNew video coding technolo-giesrdquo in IEEE ISCAS 2010 He was a keynote speaker on ldquoVision friendly videocodingrdquo in IEEE ICCIT 2010 and tutorial speaker on ldquoMultiview Video Codingrdquo inDICTA 2013 His major research interests are in the 1047297elds of video coding computervision and EEG Signals analysis Dr Paul received Vice Chancellor ResearchExcellence Award in Faculty level at 2013
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200200
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 911
best result compared to other IMFs For the RBF kernel function
the proposed methods outperform the state-of-the-art method
consistently and the average sensitivity using the proposed meth-
ods such as Quartile four weighted high frequency component using
SVD (Q4WHF_SVD) Quartile four weighted high frequency compo-
nent using DCT (Q4WHF_DCT) First IMF using EMD (FIMF_EMD)
and Quartile four weighted high frequency component using DCT and
high frequency from DWT (Q4WHF_DCTndashHF_DWT) are 8333
9912 9647 and 8713 for RBF kernel respectively whereas
the average sensitivity using the state-of-art-methods [17] and
[18] are 635 and 4804 respectively Using Morlet kernel the
average sensitivity of the proposed methods is higher compared to
the existing methods [17] and [18] Moreover we justify our result
applying linear kernel Table 2 also presents that in most of the
cases the proposed methods with linear kernel provide better
results compared to the proposed methods with Morlet and RBF
kernels in terms of sensitivity Table 2 also presents that the
proposed Q4WHF_DCT method is the best among all proposed
methods in terms of sensitivity It is also interesting to note that
the accuracy of the proposed Q4WHF_DCTndashHF_DWT method is
reasonably consistence in terms of three criteria for RBF and
Morlet kernel According to Table 2 all proposed methods always
outperform the existing methods [1718] in terms of sensitivity (i
e seizure and ictal detection rate) however in some cases the
existing methods [1718] outperform one or more proposed
methods in terms of speci1047297city (interictal detection rate) For
seizure detection sensitivity is more important compared to the
speci1047297city detection as the cost of failure detecting ictal is heavier
compared to interictal In overall classi1047297cation the proposed
methods are also better compared to the existing methods We
combine all features to classify ictal and interictal EEG signalsHowever combining is not good because LS-SVM may not prop-
erly combine all available features to make a good prediction
model The average sensitivity of the proposed methods is 9136
(average result of all kernel and lobe) whereas the sensitivities of
the state-of-the-arts methods are 5461 [18] and 799 [17]
respectively
The performance of the LS-SVM is evaluated by (ROC) plot
shown in Fig 7(a) ROC illustrates the performance of a binary
classi1047297er system where it is created by plotting the fraction of true
positives from the positives ie true positive rate (TPR) vs the
fraction of false positives from negatives ie false positive rate
(FPR) with various threshold settings TPR is known as sensitivity
and FPR is one minus the speci1047297city or true negative rate Fig 7(a)
shows that the proposed four methods are carrying good
classi1047297cation results than that of the state-of-the-art method
[18] using dataset from Frontal lobe of the third subset of training
and testing EEG signals
Fig 7(b) shows the distribution of normalized feature values
(ie energy amp entropy in the Q4WHF_DCT method STD in the
Q4WHF_DCTndashHF_DWT method STD amp energy in the Q4WHF_SVD
method and STD in the FIMF EMD method) of ictal and interictal
from the Frontal lobe dataset [22] The 1047297gure shows that the range
of feature values for ictal is different from that of the interictal
signals This phenomenon is the other evidence why we use those
features to classify ictal and interictal EEG signals in the proposed
methods
First column of Fig 8 shows classi1047297cation comparison using the
proposed methods (ie FIMF EMD and Q4WHF_DCTndashHF_DWT)
and the state-of-the-art method [18] Fig 8(b) and (c) shows the
LS-SVM classi1047297cation results using the proposed FIMF EMD and
Q4WHF_DCTndashHF_DWT against the state-of-the-art method (ie
EMD Fig 8(a)) We generate 1047297rst column of Fig 8 from Frontal lobe
for the second subset of training and testing and obtain the
classi1047297cation accuracy 9120 for the Q4WHF_DCTndashHF_DWT
method 8720 for the FIMF_EMD method and 8080 for the
EMD method It can be concluded from these 1047297gures that the
proposed method based on the DCTndashDWT and the IMF outperform
the state-of-the-art method (ie EMD) The same classi1047297cation
comparisons are also provided in second column of Fig 8 using the
state-of-the-art method based on the EMD and the proposed
method based on the SVD and DCT respectively We consider
Temporal lobe for the fourth subset of training and testing obtain
classi1047297cation accuracy 9120 for Q4WHF_SVD 8960 for
Q4WHF_DCT and 8520 for EMD [18]Thus it is concluded that
the proposed methods based on DCT and SVD also outperform theEMD method [18]
5 Conclusion
Unlike seizure and non-seizure EEG signals classi1047297cation of
ictal and interictal EEG signals is a challenging task due to non
abrupt phenomena (ie visually non-separable) between ictal and
interictal EEG signals Moreover the features of ictal and interictal
signals are not consistence in different brain locations patient
types of seizure and hospital setting during epileptic period of a
patient Thus the classi1047297cation performance for ictal and interictal
EEG signals using the existing features extraction and classi1047297cation
techniques is not at expected level Our initial observation
Fig 7 (a) The receiver operating characteristics (ROC) curves of the third subset of training and testing EEG signals by three proposed methods ( Q4WHF_SVD FIMF_EMD
and Q4WHF_DCTndashHF_DWT) against the EMD method using LS-SVM with RBF kernel from Frontal lobe and (b) Normalized distribution of feature values extracted from ictal
and interictal EEG signals using DCT DWT SVD and IMF based method where 200 samples are taken from each of ictal and interictal categories in Frontal lobe
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200198
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1011
indicates that high frequency components of EEG signals have
distinguished characteristics To improve the accuracy in this
paper we propose novel feature extractions strategies from
the high frequency components based on a number of estab-
lished transformation and decompositions The average sensi-
tivity (which is the measure for seizure) of the proposed
methods is 9136 (average result of all kernel and lobe)
whereas the sensitivity of the state-of-the-arts methods are
5461 [18] and 799 [17] respectively The experimental
results show that the proposed methods outperform the state-
of-the-art methods in terms of accuracy sensitivity and speci-
1047297city for the ictal and interictal EEG signals with greater
consistence from the large benchmark dataset in two different
brain locations
Fig 8 First column represents the classi1047297cation of ictal and interictal EEG signals from Frontal lobe for the second subset of training and testing using (a) the state-of-the-art
method against (b) the proposed IMF and (c) the DCT-DWT method Second column represents the classi1047297cation of ictal and interictal EEG signals from Temporal lobe for the
fourth subset of training and testing using (d) the state-of-the-art method against (e) the proposed SVD and (f) the DCT methods
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 199
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1111
References
[1] Neurology Applied How Science is Bringing Music Instruction Back toExpressive Development URL langhttpwwwthomasjwestmusiccomneurolo
gyappliedhtmrang (accessed 270812)[2] EEG Signal The science behind Bo-Tau URL langhttpwwwbo-taucoukindex
phpaction=the_science_behind_bo_tau rang (accessed 270812)[3] Epilepsy-SeizuresURL langhttpwwwhealthcommunitiescomepilepsy-sei
zuresabout-seizuresshtml rang (accessed 270812)[4] M Soleymani T Pun M Pantic Multi-modal emotion recognition in response
to videos IEEE Trans Affect Comput 3 (2) (2012) 211ndash223[5] S Scholler S Bosse MS Treder B Blankertz G Curio K-R Muumlller
T Wiegand Toward a direct measure of video quality perception using EEGIEEE Trans Image Process 21 (5) (2012) 2619ndash2629
[6] W Di C Zhihua F Ruifang L Guangyu and L Tian Study on human brainafter consuming alcohol based on EEG signal in Proceedings of the Interna-tional Conference on Computer Science and Information Technology (ICCSIT)5 406ndash409 2010
[7] E Estrada H Nazeran F Ebrahimi and M Mikaeili EEG signal features forcomputer-aided sleep stage detection in Proceedings of the IEEE EMBSConference on Neural Engineering 669ndash672 2009
[8] ZM Hana1047297ah KFM Yunos ZH Murat MN Taib S Lias EEG brainwavepattern for smoking behaviour after horizontal rotation treatment in Pro-ceedings of the IEEE Student Conference on Research and Development 2009
[9] ZH Murat R Shilawani SA Kadir RM Isa and MN Taib The effects of
mobile phone usage on human braingwave using EEG in Proceedings of the13th International Conference on Modelling and Simulation 2011
[10] K Lehnertz RG Andrzejak J Arnhold T Kreuz F Mormann C Rieke
G Widman CE Elger Nonliner EEG analysis in Epilepsy its possible use forinterictal focus localization seizure ancipation and prevention J ClinNeurophsysiol 18 (3) (2001) 209ndash222
[11] R Panda PS Khobragade PD Jambhule SN Jengthe PR Pal and TK GandhiClassi1047297cation of EEG signal using wavelet transform and support vectormachine for epileptic seizure diction in Proceedings of the InternationalConference on Systems in Medicine and Biology (ICSMB) 405ndash408 2010
[12] SG Dastidar H Adeli N Dadmehr Mixed-band wavelet chaos-neural net-
work methodology for epilepsy and epileptic seizure detection IEEE TransBiomeical Eng 54 (9) (2007) 1545ndash1551
[13] H Ocak Optimal classi1047297cation of epileptic seizures in EEG using waveletanalysis and genetic algorithm Signal Process 88 (2008) 1858ndash1867
[14] K Polat S Gunes Classi1047297cation of epileptiform EEG using a hybrid systembased on decision tree classi1047297er and fast Fourier transform Appl Math
Comput 187 (2) (2007) 1017ndash1026[15] SF Liang HC Wang WL Chang Combination of EEG complexity and spectral
analysis for epilepsy diagnosis and seizure detection EURASIP J Adv SignalProcess 853434 (2010) 1ndash15 (Article ID)
[16] RB Pachori Discrimination between ictal and seizure-free EEG signals usingempirical mode decomposition Res Lett Signal Process 293056 (2008) 1ndash5(Article ID)
[17] D Birvinskas V Jusas I Martisius and R Damasevicius EEG dataset reductionand feature extraction using discrete cosine transform in Proceedings of theEuropean symposium on computer modeling and simulation (EMS) 199-2042012
[18] V Bajaj RB Pachori Classi1047297cation of seizure and non-seizure EEG signalsusing empirical mode decomposition IEEE Trans Inf Technol Biomed 16 (6)(2012) 1135ndash1142
[19] Small dataset Epilepsy dataset (text format) langhttpepileptologie-bonndecmsfront_contentphpidcat=193amplang=3ampchangelang=3rang (accessed 280412)
[20] J Harrison Fast and accurate bessel function computation in Proceedings of the19th IEEE international symposium on computer arithmetic 104ndash113
2009[21] X Zhang W Diao Z Cheng Wavelet transform and singular value decom-
position of EEG signal for pattern recognition of complicated hand activitiesLect Notes Comput Sci Digit Hum Model 4561 (2007) 294ndash303
[22] Large dataset EEG Data set from Epilepsy Center of the University Hospital of
Freiburg langhttpepilepsyuni-freiburgdefreiburg-seizure-prediction-projecteeg-databaserang (accessed 100612)
[23] KD Brabanter P Karsmakers F Ojeda C Alzate and JD Brabanter KPelckmans B De Moor J Vandewalle JAK Suykens LS-SVMlab ToolboxUsers Guide version 18 Katholieke Universiteit Leuven 2011
[24] S Sanei JA Chambers EEG Signal Processing John Wiley amp Sons Ltd UK
2007 (Centre of Digital Signal Processing Cardiff University)[25] S Abe Support Vector Machine for Pattern Classi1047297cation Second Edition
Springer London Dordrecht Heidelberg New York 2010
[26] H Wang and D Hu Comparison of SVM and LS-SVM for regression inProceedings of the international conference on neural networks and brain 1279ndash283 2005
[27] H Xing M Ha B Hu D Tian Linear feature-weighted support vectormachine Fuzzy Inf Eng 1 (3) (2009) 289ndash305
[28] A Subasi EEG signal classi1047297cation using wavelet feature extraction and amixture of expert model Expert Syst Appl 32 (4) (2007) 1084ndash1093(Elsevier)
[29] JR Williamson DW Bliss DW Browne JT Narayanan Seizure predictionusing EEG spatiotemporal correlation structure Epilepsy Behav 25 (2) (2012)230ndash238
[30] TP Jung S Makeig C Humphries et al Removing electroencephalographicsartifacts by blind source separation Psychophysiology 37 (2000) 163ndash178
[31] W De Clercq A Vergult B Vanrumste WV Paesschen SV Huffel Canonicalcorrelation analysis applied to remove muscle artifacts from the electroence-phalogram IEEE Trans Biomeical Eng 53 (12) (2006) 2583ndash2587
[32] J Ma P Tao S Bayram V Svetnik Muscle artifacts in multichannel EEGcharacteristics and reduction ClinNeurophysiol 123 (8) (2012) 1676ndash1686
[33] GA Worrell L Parish SD Cranstoun R Jonas G Baltuch B Litt High-frequency oscillations and seizure generation in neocortical epilepsy Brain127 (2004) 1496ndash1506
[34] M Paul W Lin CT Lau BndashS Lee A long term reference frame for hierarchicalb-picture based video coding IEEE Trans Circuits Syst Video Technol (2014)httpdxdoiorg101109TCSVT20142302555
[35] Z Gu W Lin CT Lau BndashS Lee and M Paul Two dimensional singular valuedecomposition (2D-SVD) based video coding in Proceedings of the IEEEconference on image processing (IEEE ICIP-10) 181ndash184 2010
Mohammad Zavid Parvez received BScEng (hons)Degree in Computer Science and Engineering fromAsian University of Bangladesh in 2003 and MScDegree in Electrical Engineering with emphasis onSignal Processing from Blekinge Institute of TechnologySweden in 2010 Parvezs research interests includebiomedical signal processing and machine learning
He has more than three year industry experience asa software developer Moreover he was involved sev-eral software projects in Denmark Currently he is aPhD student and teaching staff in Charles Sturt Uni-versity
Parvez is a Graduate student member of IEEE He hasreceived full funded scholarship from the Faculty He has published several journaland conference papers in the area of EEG signal analysis and cognitive radio
Manoranjan Paul received BScEng (hons) Degree inComputer Science and Engineering from BangladeshUniversity of Engineering and Technology (BUET) Ban-gladesh in 1997 and PhD Degree from Monash Uni-versity Australia in 2005 He was an Assistant Professorin Ahsanullah University of Science and Technology Hewas a Post-Doctoral Research Fellow in the Universityof New South Wales in 2005ndash2006 Monash Universityin 2006ndash2009 and Nanyang Technological Universityin 2009ndash2011 He has joined in the School of Comput-ing and Mathematics Charles Sturt University (CSU) at2011 Currently he is a Senior Lecturer and AssociateDirector of the Center for Research in Complex Systems
(CRiCS) in CSUDr Paul is a senior member of the IEEE and ACS He is in editorial board of three
international journals including EURASIP Journal on Advances in Signal Processing(JASP) Dr Paul has served as a guest editor of Journal of Multimedia and Journal of Computers for 1047297ve special issues He has been in Technical Program Committees of more than 25 international conferences Dr Paul obtained more than A$1Mcompetitive grant money including Australian Research Council (ARC) DiscoveryProject He has published more than 65 refereed publications including journalsand conferences He organized a special session on ldquoNew video coding technolo-giesrdquo in IEEE ISCAS 2010 He was a keynote speaker on ldquoVision friendly videocodingrdquo in IEEE ICCIT 2010 and tutorial speaker on ldquoMultiview Video Codingrdquo inDICTA 2013 His major research interests are in the 1047297elds of video coding computervision and EEG Signals analysis Dr Paul received Vice Chancellor ResearchExcellence Award in Faculty level at 2013
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200200
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1011
indicates that high frequency components of EEG signals have
distinguished characteristics To improve the accuracy in this
paper we propose novel feature extractions strategies from
the high frequency components based on a number of estab-
lished transformation and decompositions The average sensi-
tivity (which is the measure for seizure) of the proposed
methods is 9136 (average result of all kernel and lobe)
whereas the sensitivity of the state-of-the-arts methods are
5461 [18] and 799 [17] respectively The experimental
results show that the proposed methods outperform the state-
of-the-art methods in terms of accuracy sensitivity and speci-
1047297city for the ictal and interictal EEG signals with greater
consistence from the large benchmark dataset in two different
brain locations
Fig 8 First column represents the classi1047297cation of ictal and interictal EEG signals from Frontal lobe for the second subset of training and testing using (a) the state-of-the-art
method against (b) the proposed IMF and (c) the DCT-DWT method Second column represents the classi1047297cation of ictal and interictal EEG signals from Temporal lobe for the
fourth subset of training and testing using (d) the state-of-the-art method against (e) the proposed SVD and (f) the DCT methods
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200 199
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1111
References
[1] Neurology Applied How Science is Bringing Music Instruction Back toExpressive Development URL langhttpwwwthomasjwestmusiccomneurolo
gyappliedhtmrang (accessed 270812)[2] EEG Signal The science behind Bo-Tau URL langhttpwwwbo-taucoukindex
phpaction=the_science_behind_bo_tau rang (accessed 270812)[3] Epilepsy-SeizuresURL langhttpwwwhealthcommunitiescomepilepsy-sei
zuresabout-seizuresshtml rang (accessed 270812)[4] M Soleymani T Pun M Pantic Multi-modal emotion recognition in response
to videos IEEE Trans Affect Comput 3 (2) (2012) 211ndash223[5] S Scholler S Bosse MS Treder B Blankertz G Curio K-R Muumlller
T Wiegand Toward a direct measure of video quality perception using EEGIEEE Trans Image Process 21 (5) (2012) 2619ndash2629
[6] W Di C Zhihua F Ruifang L Guangyu and L Tian Study on human brainafter consuming alcohol based on EEG signal in Proceedings of the Interna-tional Conference on Computer Science and Information Technology (ICCSIT)5 406ndash409 2010
[7] E Estrada H Nazeran F Ebrahimi and M Mikaeili EEG signal features forcomputer-aided sleep stage detection in Proceedings of the IEEE EMBSConference on Neural Engineering 669ndash672 2009
[8] ZM Hana1047297ah KFM Yunos ZH Murat MN Taib S Lias EEG brainwavepattern for smoking behaviour after horizontal rotation treatment in Pro-ceedings of the IEEE Student Conference on Research and Development 2009
[9] ZH Murat R Shilawani SA Kadir RM Isa and MN Taib The effects of
mobile phone usage on human braingwave using EEG in Proceedings of the13th International Conference on Modelling and Simulation 2011
[10] K Lehnertz RG Andrzejak J Arnhold T Kreuz F Mormann C Rieke
G Widman CE Elger Nonliner EEG analysis in Epilepsy its possible use forinterictal focus localization seizure ancipation and prevention J ClinNeurophsysiol 18 (3) (2001) 209ndash222
[11] R Panda PS Khobragade PD Jambhule SN Jengthe PR Pal and TK GandhiClassi1047297cation of EEG signal using wavelet transform and support vectormachine for epileptic seizure diction in Proceedings of the InternationalConference on Systems in Medicine and Biology (ICSMB) 405ndash408 2010
[12] SG Dastidar H Adeli N Dadmehr Mixed-band wavelet chaos-neural net-
work methodology for epilepsy and epileptic seizure detection IEEE TransBiomeical Eng 54 (9) (2007) 1545ndash1551
[13] H Ocak Optimal classi1047297cation of epileptic seizures in EEG using waveletanalysis and genetic algorithm Signal Process 88 (2008) 1858ndash1867
[14] K Polat S Gunes Classi1047297cation of epileptiform EEG using a hybrid systembased on decision tree classi1047297er and fast Fourier transform Appl Math
Comput 187 (2) (2007) 1017ndash1026[15] SF Liang HC Wang WL Chang Combination of EEG complexity and spectral
analysis for epilepsy diagnosis and seizure detection EURASIP J Adv SignalProcess 853434 (2010) 1ndash15 (Article ID)
[16] RB Pachori Discrimination between ictal and seizure-free EEG signals usingempirical mode decomposition Res Lett Signal Process 293056 (2008) 1ndash5(Article ID)
[17] D Birvinskas V Jusas I Martisius and R Damasevicius EEG dataset reductionand feature extraction using discrete cosine transform in Proceedings of theEuropean symposium on computer modeling and simulation (EMS) 199-2042012
[18] V Bajaj RB Pachori Classi1047297cation of seizure and non-seizure EEG signalsusing empirical mode decomposition IEEE Trans Inf Technol Biomed 16 (6)(2012) 1135ndash1142
[19] Small dataset Epilepsy dataset (text format) langhttpepileptologie-bonndecmsfront_contentphpidcat=193amplang=3ampchangelang=3rang (accessed 280412)
[20] J Harrison Fast and accurate bessel function computation in Proceedings of the19th IEEE international symposium on computer arithmetic 104ndash113
2009[21] X Zhang W Diao Z Cheng Wavelet transform and singular value decom-
position of EEG signal for pattern recognition of complicated hand activitiesLect Notes Comput Sci Digit Hum Model 4561 (2007) 294ndash303
[22] Large dataset EEG Data set from Epilepsy Center of the University Hospital of
Freiburg langhttpepilepsyuni-freiburgdefreiburg-seizure-prediction-projecteeg-databaserang (accessed 100612)
[23] KD Brabanter P Karsmakers F Ojeda C Alzate and JD Brabanter KPelckmans B De Moor J Vandewalle JAK Suykens LS-SVMlab ToolboxUsers Guide version 18 Katholieke Universiteit Leuven 2011
[24] S Sanei JA Chambers EEG Signal Processing John Wiley amp Sons Ltd UK
2007 (Centre of Digital Signal Processing Cardiff University)[25] S Abe Support Vector Machine for Pattern Classi1047297cation Second Edition
Springer London Dordrecht Heidelberg New York 2010
[26] H Wang and D Hu Comparison of SVM and LS-SVM for regression inProceedings of the international conference on neural networks and brain 1279ndash283 2005
[27] H Xing M Ha B Hu D Tian Linear feature-weighted support vectormachine Fuzzy Inf Eng 1 (3) (2009) 289ndash305
[28] A Subasi EEG signal classi1047297cation using wavelet feature extraction and amixture of expert model Expert Syst Appl 32 (4) (2007) 1084ndash1093(Elsevier)
[29] JR Williamson DW Bliss DW Browne JT Narayanan Seizure predictionusing EEG spatiotemporal correlation structure Epilepsy Behav 25 (2) (2012)230ndash238
[30] TP Jung S Makeig C Humphries et al Removing electroencephalographicsartifacts by blind source separation Psychophysiology 37 (2000) 163ndash178
[31] W De Clercq A Vergult B Vanrumste WV Paesschen SV Huffel Canonicalcorrelation analysis applied to remove muscle artifacts from the electroence-phalogram IEEE Trans Biomeical Eng 53 (12) (2006) 2583ndash2587
[32] J Ma P Tao S Bayram V Svetnik Muscle artifacts in multichannel EEGcharacteristics and reduction ClinNeurophysiol 123 (8) (2012) 1676ndash1686
[33] GA Worrell L Parish SD Cranstoun R Jonas G Baltuch B Litt High-frequency oscillations and seizure generation in neocortical epilepsy Brain127 (2004) 1496ndash1506
[34] M Paul W Lin CT Lau BndashS Lee A long term reference frame for hierarchicalb-picture based video coding IEEE Trans Circuits Syst Video Technol (2014)httpdxdoiorg101109TCSVT20142302555
[35] Z Gu W Lin CT Lau BndashS Lee and M Paul Two dimensional singular valuedecomposition (2D-SVD) based video coding in Proceedings of the IEEEconference on image processing (IEEE ICIP-10) 181ndash184 2010
Mohammad Zavid Parvez received BScEng (hons)Degree in Computer Science and Engineering fromAsian University of Bangladesh in 2003 and MScDegree in Electrical Engineering with emphasis onSignal Processing from Blekinge Institute of TechnologySweden in 2010 Parvezs research interests includebiomedical signal processing and machine learning
He has more than three year industry experience asa software developer Moreover he was involved sev-eral software projects in Denmark Currently he is aPhD student and teaching staff in Charles Sturt Uni-versity
Parvez is a Graduate student member of IEEE He hasreceived full funded scholarship from the Faculty He has published several journaland conference papers in the area of EEG signal analysis and cognitive radio
Manoranjan Paul received BScEng (hons) Degree inComputer Science and Engineering from BangladeshUniversity of Engineering and Technology (BUET) Ban-gladesh in 1997 and PhD Degree from Monash Uni-versity Australia in 2005 He was an Assistant Professorin Ahsanullah University of Science and Technology Hewas a Post-Doctoral Research Fellow in the Universityof New South Wales in 2005ndash2006 Monash Universityin 2006ndash2009 and Nanyang Technological Universityin 2009ndash2011 He has joined in the School of Comput-ing and Mathematics Charles Sturt University (CSU) at2011 Currently he is a Senior Lecturer and AssociateDirector of the Center for Research in Complex Systems
(CRiCS) in CSUDr Paul is a senior member of the IEEE and ACS He is in editorial board of three
international journals including EURASIP Journal on Advances in Signal Processing(JASP) Dr Paul has served as a guest editor of Journal of Multimedia and Journal of Computers for 1047297ve special issues He has been in Technical Program Committees of more than 25 international conferences Dr Paul obtained more than A$1Mcompetitive grant money including Australian Research Council (ARC) DiscoveryProject He has published more than 65 refereed publications including journalsand conferences He organized a special session on ldquoNew video coding technolo-giesrdquo in IEEE ISCAS 2010 He was a keynote speaker on ldquoVision friendly videocodingrdquo in IEEE ICCIT 2010 and tutorial speaker on ldquoMultiview Video Codingrdquo inDICTA 2013 His major research interests are in the 1047297elds of video coding computervision and EEG Signals analysis Dr Paul received Vice Chancellor ResearchExcellence Award in Faculty level at 2013
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200200
7182019 Analyzing EEG
httpslidepdfcomreaderfullanalyzing-eeg 1111
References
[1] Neurology Applied How Science is Bringing Music Instruction Back toExpressive Development URL langhttpwwwthomasjwestmusiccomneurolo
gyappliedhtmrang (accessed 270812)[2] EEG Signal The science behind Bo-Tau URL langhttpwwwbo-taucoukindex
phpaction=the_science_behind_bo_tau rang (accessed 270812)[3] Epilepsy-SeizuresURL langhttpwwwhealthcommunitiescomepilepsy-sei
zuresabout-seizuresshtml rang (accessed 270812)[4] M Soleymani T Pun M Pantic Multi-modal emotion recognition in response
to videos IEEE Trans Affect Comput 3 (2) (2012) 211ndash223[5] S Scholler S Bosse MS Treder B Blankertz G Curio K-R Muumlller
T Wiegand Toward a direct measure of video quality perception using EEGIEEE Trans Image Process 21 (5) (2012) 2619ndash2629
[6] W Di C Zhihua F Ruifang L Guangyu and L Tian Study on human brainafter consuming alcohol based on EEG signal in Proceedings of the Interna-tional Conference on Computer Science and Information Technology (ICCSIT)5 406ndash409 2010
[7] E Estrada H Nazeran F Ebrahimi and M Mikaeili EEG signal features forcomputer-aided sleep stage detection in Proceedings of the IEEE EMBSConference on Neural Engineering 669ndash672 2009
[8] ZM Hana1047297ah KFM Yunos ZH Murat MN Taib S Lias EEG brainwavepattern for smoking behaviour after horizontal rotation treatment in Pro-ceedings of the IEEE Student Conference on Research and Development 2009
[9] ZH Murat R Shilawani SA Kadir RM Isa and MN Taib The effects of
mobile phone usage on human braingwave using EEG in Proceedings of the13th International Conference on Modelling and Simulation 2011
[10] K Lehnertz RG Andrzejak J Arnhold T Kreuz F Mormann C Rieke
G Widman CE Elger Nonliner EEG analysis in Epilepsy its possible use forinterictal focus localization seizure ancipation and prevention J ClinNeurophsysiol 18 (3) (2001) 209ndash222
[11] R Panda PS Khobragade PD Jambhule SN Jengthe PR Pal and TK GandhiClassi1047297cation of EEG signal using wavelet transform and support vectormachine for epileptic seizure diction in Proceedings of the InternationalConference on Systems in Medicine and Biology (ICSMB) 405ndash408 2010
[12] SG Dastidar H Adeli N Dadmehr Mixed-band wavelet chaos-neural net-
work methodology for epilepsy and epileptic seizure detection IEEE TransBiomeical Eng 54 (9) (2007) 1545ndash1551
[13] H Ocak Optimal classi1047297cation of epileptic seizures in EEG using waveletanalysis and genetic algorithm Signal Process 88 (2008) 1858ndash1867
[14] K Polat S Gunes Classi1047297cation of epileptiform EEG using a hybrid systembased on decision tree classi1047297er and fast Fourier transform Appl Math
Comput 187 (2) (2007) 1017ndash1026[15] SF Liang HC Wang WL Chang Combination of EEG complexity and spectral
analysis for epilepsy diagnosis and seizure detection EURASIP J Adv SignalProcess 853434 (2010) 1ndash15 (Article ID)
[16] RB Pachori Discrimination between ictal and seizure-free EEG signals usingempirical mode decomposition Res Lett Signal Process 293056 (2008) 1ndash5(Article ID)
[17] D Birvinskas V Jusas I Martisius and R Damasevicius EEG dataset reductionand feature extraction using discrete cosine transform in Proceedings of theEuropean symposium on computer modeling and simulation (EMS) 199-2042012
[18] V Bajaj RB Pachori Classi1047297cation of seizure and non-seizure EEG signalsusing empirical mode decomposition IEEE Trans Inf Technol Biomed 16 (6)(2012) 1135ndash1142
[19] Small dataset Epilepsy dataset (text format) langhttpepileptologie-bonndecmsfront_contentphpidcat=193amplang=3ampchangelang=3rang (accessed 280412)
[20] J Harrison Fast and accurate bessel function computation in Proceedings of the19th IEEE international symposium on computer arithmetic 104ndash113
2009[21] X Zhang W Diao Z Cheng Wavelet transform and singular value decom-
position of EEG signal for pattern recognition of complicated hand activitiesLect Notes Comput Sci Digit Hum Model 4561 (2007) 294ndash303
[22] Large dataset EEG Data set from Epilepsy Center of the University Hospital of
Freiburg langhttpepilepsyuni-freiburgdefreiburg-seizure-prediction-projecteeg-databaserang (accessed 100612)
[23] KD Brabanter P Karsmakers F Ojeda C Alzate and JD Brabanter KPelckmans B De Moor J Vandewalle JAK Suykens LS-SVMlab ToolboxUsers Guide version 18 Katholieke Universiteit Leuven 2011
[24] S Sanei JA Chambers EEG Signal Processing John Wiley amp Sons Ltd UK
2007 (Centre of Digital Signal Processing Cardiff University)[25] S Abe Support Vector Machine for Pattern Classi1047297cation Second Edition
Springer London Dordrecht Heidelberg New York 2010
[26] H Wang and D Hu Comparison of SVM and LS-SVM for regression inProceedings of the international conference on neural networks and brain 1279ndash283 2005
[27] H Xing M Ha B Hu D Tian Linear feature-weighted support vectormachine Fuzzy Inf Eng 1 (3) (2009) 289ndash305
[28] A Subasi EEG signal classi1047297cation using wavelet feature extraction and amixture of expert model Expert Syst Appl 32 (4) (2007) 1084ndash1093(Elsevier)
[29] JR Williamson DW Bliss DW Browne JT Narayanan Seizure predictionusing EEG spatiotemporal correlation structure Epilepsy Behav 25 (2) (2012)230ndash238
[30] TP Jung S Makeig C Humphries et al Removing electroencephalographicsartifacts by blind source separation Psychophysiology 37 (2000) 163ndash178
[31] W De Clercq A Vergult B Vanrumste WV Paesschen SV Huffel Canonicalcorrelation analysis applied to remove muscle artifacts from the electroence-phalogram IEEE Trans Biomeical Eng 53 (12) (2006) 2583ndash2587
[32] J Ma P Tao S Bayram V Svetnik Muscle artifacts in multichannel EEGcharacteristics and reduction ClinNeurophysiol 123 (8) (2012) 1676ndash1686
[33] GA Worrell L Parish SD Cranstoun R Jonas G Baltuch B Litt High-frequency oscillations and seizure generation in neocortical epilepsy Brain127 (2004) 1496ndash1506
[34] M Paul W Lin CT Lau BndashS Lee A long term reference frame for hierarchicalb-picture based video coding IEEE Trans Circuits Syst Video Technol (2014)httpdxdoiorg101109TCSVT20142302555
[35] Z Gu W Lin CT Lau BndashS Lee and M Paul Two dimensional singular valuedecomposition (2D-SVD) based video coding in Proceedings of the IEEEconference on image processing (IEEE ICIP-10) 181ndash184 2010
Mohammad Zavid Parvez received BScEng (hons)Degree in Computer Science and Engineering fromAsian University of Bangladesh in 2003 and MScDegree in Electrical Engineering with emphasis onSignal Processing from Blekinge Institute of TechnologySweden in 2010 Parvezs research interests includebiomedical signal processing and machine learning
He has more than three year industry experience asa software developer Moreover he was involved sev-eral software projects in Denmark Currently he is aPhD student and teaching staff in Charles Sturt Uni-versity
Parvez is a Graduate student member of IEEE He hasreceived full funded scholarship from the Faculty He has published several journaland conference papers in the area of EEG signal analysis and cognitive radio
Manoranjan Paul received BScEng (hons) Degree inComputer Science and Engineering from BangladeshUniversity of Engineering and Technology (BUET) Ban-gladesh in 1997 and PhD Degree from Monash Uni-versity Australia in 2005 He was an Assistant Professorin Ahsanullah University of Science and Technology Hewas a Post-Doctoral Research Fellow in the Universityof New South Wales in 2005ndash2006 Monash Universityin 2006ndash2009 and Nanyang Technological Universityin 2009ndash2011 He has joined in the School of Comput-ing and Mathematics Charles Sturt University (CSU) at2011 Currently he is a Senior Lecturer and AssociateDirector of the Center for Research in Complex Systems
(CRiCS) in CSUDr Paul is a senior member of the IEEE and ACS He is in editorial board of three
international journals including EURASIP Journal on Advances in Signal Processing(JASP) Dr Paul has served as a guest editor of Journal of Multimedia and Journal of Computers for 1047297ve special issues He has been in Technical Program Committees of more than 25 international conferences Dr Paul obtained more than A$1Mcompetitive grant money including Australian Research Council (ARC) DiscoveryProject He has published more than 65 refereed publications including journalsand conferences He organized a special session on ldquoNew video coding technolo-giesrdquo in IEEE ISCAS 2010 He was a keynote speaker on ldquoVision friendly videocodingrdquo in IEEE ICCIT 2010 and tutorial speaker on ldquoMultiview Video Codingrdquo inDICTA 2013 His major research interests are in the 1047297elds of video coding computervision and EEG Signals analysis Dr Paul received Vice Chancellor ResearchExcellence Award in Faculty level at 2013
MZ Parvez M Paul Neurocomputing 145 (2014) 190ndash 200200