Research ArticleClassification of Task-State fMRI Data Based on Circle-EMD andMachine Learning

Renzhou Gui Tongjie Chen and Han Nie

e Department of Information and Communication Engineering Tongji University Shanghai 201804 China

Correspondence should be addressed to Renzhou Gui rzguitongjieducn

Received 16 July 2019 Revised 31 May 2020 Accepted 7 July 2020 Published 1 August 2020

Academic Editor Rasit Koker

Copyright copy 2020 Renzhou Gui et al -is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

In the research work of the brain-computer interface and the function of human brain work the state classification of multitaskstate fMRI data is a problem -e fMRI signal of the human brain is a nonstationary signal with many noise effects and in-terference Based on the commonly used nonstationary signal analysis method HilbertndashHuang transform (HHT) we propose animproved circle-EMD algorithm to suppress the end effect -e algorithm can extract different intrinsic mode functions (IMFs)decompose the fMRI data to filter out low frequency and other redundant noise signals and more accurately reflect the truecharacteristics of the original signal For the filtered fMRI signal we use three existing different machine learningmethods logisticregression (LR) support vector machine (SVM) and deep neural network (DNN) to achieve effective classification of differenttask states -e experiment compares the results of these machine learning methods and confirms that the deep neural networkhas the highest accuracy for task-state fMRI data classification and the effectiveness of the improved circle-EMD algorithm

1 Introduction

-e brain-computer interface (BCI) allows people to controlthe machine through signals generated by brain activity toimplement communications between people and the envi-ronment and expand the ability of humans to controlmachines [1] In recent years functional magnetic resonanceimaging (fMRI) has become an important means of studyingthe high-level cortical function of the human brain [2ndash5]Based on blood-oxygen-level-dependent (BOLD) imaging itis a tool that allows functional changes to be studied in vivo[6] and has been used to display the morphology and lo-cation of the cortical center -e process of data acquisitionis noninvasive and nonradiative and the image resolution ofthe data is very high and easy to combine with conventionalMR anatomical images

One major feature of the fMRI signal is that it does notsatisfy linearity and stationarity and has both temporal andspatial distribution characteristics [7 8] In 1999 Hil-bertndashHuang transform (HHT) [9] which is a combination ofempirical mode decomposition (EMD) and Hilbert trans-form (HT) is an adaptive time-frequency analysis method

and very suitable for feature extraction and analysis of thenonlinear and nonstationary signals [10] HHT has goodtemporal and spatial resolution [11ndash13] and does not requirea priori function basis [14] so that the original signal can besmoothed and decomposed to different scales of fluctuationsand trends step by step [15] -e application field is veryextensive [1 16 17] and very conducive to biomedical signalextraction [18]

However a tricky problem with applying the EMDmethod is that when the cubic spline function is used tocalculate both ends of the data sequence an end effectoccurs causing the result to be distorted In response to thisproblem Deng et al [19] used neural networks to extend thevarious intrinsic mode function (IMF) components toachieve accurate EMD decomposition Yang and Jia [20]optimized the EMD and HT processes through time seriesmodeling and prediction methods and achieved efficientfrequency domain decomposition on stationary and simplenonstationary time series Cheng et al [21] Peng et al [16]and Yang et al [22] used the SVM approach to solve the endeffect in the EMD method Wang and Gang [23] proposed anew data expansion method based on minimum similarity

HindawiComputational Intelligence and NeuroscienceVolume 2020 Article ID 7691294 10 pageshttpsdoiorg10115520207691294

distance which effectively suppresses the divergence at bothends of the signal during EMD decomposition thus im-proving the frequency resolution On the basis of cubicspline interpolation Tai and Deng [24] proposed an im-proved EMD method based on multiobjective optimizationwhich achieved a good solution From the above we can seethat there are many improved methods for nonstationarysignal decomposition but there is still a need to find asuitable fast and efficient solution for spatial signals

After smoothing the fMRI signal and extracting fea-tures machine state learning can be used to classify featurelearning and reconstruction and so on for the task-statefMRI data In recent years many classification methodshave been used and have important significance for fMRIresearch such as support vector machine (SVM) [25]neighbor algorithm (kNN) [26] Gaussian Bayesian (GNB)linear discriminant analysis (LDA) [27] logistic regression(LR) [28] and deep learning [29ndash31] Compared withtraditional machine learning methods deep learning hasstrong learning ability and can make better use of data setsfor feature extraction [32] Nowadays the use of deeplearning to achieve in-depth analysis of fMRI data has avery high practical value However at present the high-precision classification of multitask state fMRI data is still aproblem

-ere are many noise effects in the task-state data andthe accuracy is generally lower than that of the rest state dataXin et al [33] proposed a new classification algorithm fordepression called weighted discriminant dictionary learning(WDDL) of fMRI data in task state with an accuracy rate of7931 Ertugrul et al [34] proposed a new framework toencode the local connectivity patterns of the brain -eyclassify the cognitive state of the Human ConnectomeProject (HCP) task fMRI dataset by training the SVMWhencharacterizing the pairwise correlation between pairs of boldresponses in all regions the classification accuracy was7749

In view of the above analysis we use EMD to process thespatial frequency information of fMRI data at each timepoint and improve the effectiveness of machine learning forfMRI data by stripping out the spatial frequency data ofdifferent stages -erefore based on the fMRI spatial datathis paper proposed an effective circle-EMD method fordecomposing spatial nonstationary signals for the researchand classification of multitask state fMRI data and solved theend effect of the traditional EMDmethodWe deeply studiedand compared the characteristics and contribution of eachIMF of fMRI data verified the effectiveness of this methodby using a variety of machine learning algorithms and fi-nally achieved a higher precision classification of fMRIspatial data

2 Materials and Methods

21 Basic Principles of the EMD Method -e purpose ofEMD is to decompose the signal into the sum of the IMFsand obtain the components of the original signal withdifferent frequencies -e main steps of EMD can besummarized as follows

(1) First local extrema including the maximum andminimum values of the signal x(t) are obtained andthe upper envelope signals eupp(t) and the lowerenvelope signal elow(t) are respectively constructedby cubic spline interpolation according to thesevalues And then we get the average of the envelopem(t) [eupp(t) + elow(t)]2

(2) We calculate the signal h1(t) which has high fre-quency by reducing the average from the signalh1(t) x(t) minus m(t)

(3) If the signal h1(t) satisfies the two conditions of theIMF [35] (1) the function must have the samenumber of local extremum points and zero crossingsover the entire time range or at most one differenceand (2) at any point in time the envelope of the localmaximum (upper envelope) and the envelope of thelocal minimum (lower envelope) must be zero onaverage -en we set this signal to IMF and name itC1(t) and the residual signal r1(t) x(t) minus C1(t)If the condition is not satisfied the signals h1(t) arerepeated as steps 1 and 2 as the original signal x(t)and it is worthwhile to obtain an IMF that satisfiesthe condition

(4) Steps 1sim3 are repeated until the residual signal rn(t)

is a monotonic function or constant

Finally we decompose the signal to generate n IMFs anda residual signal namely x(t) 1113936

ni1 Ci(t) + rn(t)

22 Improvement of EMD Algorithm Circle-EMD Sincethe EMD algorithm uses a cubic spline difference to fit theenvelope of the maxima and minima it produces end effectsat both ends By analyzing the spatial signal sequence offMRI (at one time) we find that the signals of each ROI offMRI have no temporal correlations but have spatial cor-relation such as a net and should not be regarded as timeseries but network -erefore in this paper the data queuesof each ROI are connected to form a circle (as shown in theleft part of Figure 1 the gray dotted lines are the back-endpart and front-end part of the data supplementary to thefront-end and back-end of this original data) so that there isno end effect when calculating the envelope and it can betterhelp extract the space feature

Let the signal x(t) be n+ 1 values from 0 to n and S(j)denotes the fitting function between the jth and j+ 1thvalues and the second derivative of each endpoint isSjPrime Mj (j 0 1 2 n) -en the traditional cubic splineinterpolation equations are shown in the following equation

μ1M0 + 2M1 + λ1M2 6f t0 t1 t21113858 1113859

μ2M1 + 2M2 + λ2M3 6f t1 t2 t31113858 1113859

μnminus 1Mnminus 2 + 2Mnminus 1 + λnminus 1Mn 6f tnminus 2 tnminus 1 tn1113858 1113859

⎧⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎩

(1)

-ere is a total of n minus 1 equations and n+ 1 unknownvariables where hj tj+1 minus tj 1 μj (hjminus 1(hjminus 1 +

hj)) 05 λJ (hj(hjminus 1 + hj)) 05 and f[tjminus 1 tj tj+1]

2 Computational Intelligence and Neuroscience

(1(hjminus 1 + hj))[((x(tj) minus x(tjminus 1))hjminus 1) minus ((x(tj+1) minus

x(tj)) hj)] 2x(tj) minus x(tjminus 1) minus x(tj+1) After connectingthe beginning and end of the data queue we set the fitfunction between x(n) and x(0) to S(n)

-rough the formulas SPrime(n)0 SPrime(0)n andSPrime(n)n SPrime(n minus 1)n there are complementary equations

μ0Mn + 2M0 + λ0M1 6f tn t0 t11113858 1113859

μnMnminus 1 + 2Mn + λnM0 6f tnminus 1 tn t01113858 11138591113896 (2)

At this time a total of n+ 1 equations and n+ 1 unknownvariables can calculate all the unknown coefficients -eindividual IMF components of the EMD decomposition canbe further determined by the new annular envelope asshown in the right part of Figure 1 All the IMF signalsdecomposed by circle-EMD can also be regarded as a cir-cular data as shown in Figure 2 in which the black signal isthe original signal and the other signals with other colors areIMF signals

23 Machine Learning Algorithms andModels We use threeexisting machine learning methods LR SVM and DNN toachieve effective classification of different task states

231 Nonlinear SVM SVM is a binary classifier based onsupervised learning and has been widely used in variousscientific fields [36ndash38] -e boundary of classification is themaximum margin hyperplane for solving learning samplesWhen encountering some linearly inseparable problems(feature surfaces have hypersurfaces) nonlinear functionscan be used to transform these into linear separable prob-lems by mapping data from the original feature space tohigher-dimensional Hilbert spaces Common kernel func-tions in SVM are polynomial kernel RBF kernel Laplaciankernel and Sigmoid kernel -is paper uses the RBF kernelfunction also known as the Gaussian kernel whose cor-responding mapping function can map the sample space toan infinite dimensional space -e analytical expression isshown in the following equation

K X1 X2( 1113857 exp minus X1 minus X221113872 11138732σ2)1113872 11138731113872 (3)

232 Logistic Regression for Multiclassification LR is ageneralized linear regression analysis model that can handleboth classification and regression problems -e improvedLR can even handle multiclassification problems [39] -erelationship between the intermediate value y and the input

x represents the linear part of the model which isy 1113936

mi1(wi lowast xi) + b where wi is the weight matrix and b is

the bias -rough the obtained y the final result of theregression function can be obtained after passing theSoftmax function -e formula for Softmax isS(y) (1(1 + eminus y)) whose the value range is distributed in[01]

233 Deep Neural Network -e neural network was firstdeveloped into an important part [40] of machine learningby being inspired by neurons in the brain-e basis of neuralnetworks is the perceptron model [41] A deep neuralnetwork is a neural network that contains multiple layers ofhidden layers DNN can be divided into three categoriesaccording to the location of different layers input layerhidden layer and output layer

-e deep learning model designed in this paper is shownin Figure 3 -ere are three hidden layers -e number ofneurons in each layer is 300 200 and 100 respectively -eactivation function is ReLU function Each layer usesDropout and BatchNorm2d regularization to help preventoverfitting of the training -e loss function is the cross-entropy loss function and the batch_size size is 32

C minus1n

1113944x

[y ln a +(1 minus y)ln(1 minus z)] (4)

-e cross-entropy represents the distance between theactual output (probability) and the expected output(probability) that is the smaller the value of the cross-entropy the closer the two probability distributions are It isvery helpful for the training of machine learning [42]

elow (t)

Residualsignal

IMF-1IMF-2IMF-3IMF-4IMF-5

IMF-n

helliphellip

helliphellip

m (t) = [eupp(t) + elow(t)]2

m (t)x (t)

eupp (t)

x (t)

Figure 1 flowchart of the proposed circle-EMD

11 9 7 5 3 1 1 3 5 7 9 11

IMF numberfM

RI v

alue

IMF number

1510

50

ndash5ndash10ndash15

20

ndash2011 9 7 5 3 1 1 3 5 7 9 11

Figure 2 IMF signals with different frequency domains decom-posed by circle-EMD

Computational Intelligence and Neuroscience 3

Assuming that the probability distribution y is the desiredoutput the probability distribution z is the actual output-e formula for the cross-entropy loss function is shown inthe following equation

We used 10-fold cross-validation to evaluate the modelresults -e data are divided into 10 groups and 9 experi-ments are used as the training set and 1 group is used as thetest set for each experiment A total of 10 experiments wereperformed so each group of data could be calculated as a testset Finally the accuracy of all test sets was averaged toobtain the final experimental results

3 Experimental Process and Results

-e experimental process is shown in Figure 4 fMRI data arefirst preprocessed by the Brain Decoder Toolbox and thenIMF signals are obtained by circle-EMD decomposition-ree different machine learning tools are used as experi-ments to verify the effective role of circle-EMD

31 fMRI Data and Preprocessing -e data and fMRIdecoding tools used in this article were obtained from thewebsite (httpwwwcnsatrjpdniendownloadsbrain-decoder-toolbox) -e data in this paper were obtainedon a 15 T MRI scanner (Shimadzu-Marconi) with a TRTEflip angle 5 s50ms90 degree field of view (FOV)

192mm acquisition matrix 64times 64 for functional imagesand 50 slices During the fMRI data acquisition process thesubjects performed three gestures of stone scissors andpaper in the MRI according to the instructions -ere aretotally 10 runs and each run has a 20-second break at thebeginning and end In each run there are 32 states including8 rest states and 24 task states so the task-state data have 240samples We divided into 10-fold cross-validation 216samples as the training set and 24 samples as the test set perexperiment as shown in Table 1

We use the Brain DecoderToolbox tool (the process flowand parameters are shown in Table 2) to smooth decontourand regularize the fMRI data and filter out the data of the sixROIs most relevant to the pose which are ldquoM1_RHandrdquo

ldquoSMA_RHandrdquo ldquoCB_RHandrdquo ldquoM1_LHandrdquo ldquoSMA_L-Handrdquo and ldquoCB_LHandrdquo thus obtaining 5333 voxels foreach sample -e locations of each ROI data are shown inTable 3 Based on the statistics of each voxelchannel (thedata value is distributed between minus 544 and 2412919) wefilter out the top 200 data with the highest statistical value(the data value is distributed between 599434 and 2412919)to initially filter unnecessary redundancy (-rough theactual classification and comparison in the experiment if theoriginal 5333 data are directly processed and classifiedwithout the filter the accuracy of the classification is rela-tively low whether the SVM LR or DNN method is used)

By counting the position and number of six ROIs in thefMRI data sequence we obtain the results shown in Table 3

32 Decomposing Data by Using Circle-EMD

321 Decomposed IMF Signal As shown in Figure 5 it is apartial result of the circle-EMD decomposition of the pre-processed fMRI data sequence where the ldquoCurrent IMFrdquo isthe sequence number of the decomposed IMF signal and theldquoSift Iterrdquo is the number of iterations for extracting the IMFsignal For our data set the number of IMFs decomposed isroughly between 4 and 7

Using the original EMD and the circle-EMD algorithmproposed in this paper we decompose the preprocessedfMRI data for each IMF signal as shown in Figure 6 -euppermost black signal is an original data signal with a labelof stone and each IMF signal is decomposed by two EMDmethods (the IMF signals are sequentially decomposed fromtop to bottom) By comparing the distribution results inTable 1 we find that the IMF signals obtained by the twodecompositions are very similar in the high-frequency partFor example in ldquoCB_RHandrdquo these data have strongcharacteristic signals at both high and low frequenciesHowever in the low-frequency part the signal character-istics are more obvious such as the data in ldquoM1_RHandrdquo

-e frequency ranges for each IMF in Figure 6 are shownin Figure 7 For the convenience of display fast Fouriertransform (FFT) is used to transform these spatial IMF

Stone

Scissors

Paper

Inputs = 200

hellip

hellip

hellip

Hidden layerfc1 = 300 Hidden layer

fc2 = 200 Hidden layerfc3 = 100

Weights 1200 times 300

Weights 2300 times 200

Weights 3200 times 100

Weights 4100 times 3

Somax layeroutputs = 3

Figure 3 Deep learning model in this paper

4 Computational Intelligence and Neuroscience

(1(hjminus 1 + hj))[((x(tj) minus x(tjminus 1))hjminus 1) minus ((x(tj+1) minus

x(tj)) hj)] 2x(tj) minus x(tjminus 1) minus x(tj+1) After connectingthe beginning and end of the data queue we set the fitfunction between x(n) and x(0) to S(n)

-rough the formulas SPrime(n)0 SPrime(0)n andSPrime(n)n SPrime(n minus 1)n there are complementary equations

μ0Mn + 2M0 + λ0M1 6f tn t0 t11113858 1113859

μnMnminus 1 + 2Mn + λnM0 6f tnminus 1 tn t01113858 11138591113896 (2)

At this time a total of n+ 1 equations and n+ 1 unknownvariables can calculate all the unknown coefficients -eindividual IMF components of the EMD decomposition canbe further determined by the new annular envelope asshown in the right part of Figure 1 All the IMF signalsdecomposed by circle-EMD can also be regarded as a cir-cular data as shown in Figure 2 in which the black signal isthe original signal and the other signals with other colors areIMF signals

23 Machine Learning Algorithms andModels We use threeexisting machine learning methods LR SVM and DNN toachieve effective classification of different task states

231 Nonlinear SVM SVM is a binary classifier based onsupervised learning and has been widely used in variousscientific fields [36ndash38] -e boundary of classification is themaximum margin hyperplane for solving learning samplesWhen encountering some linearly inseparable problems(feature surfaces have hypersurfaces) nonlinear functionscan be used to transform these into linear separable prob-lems by mapping data from the original feature space tohigher-dimensional Hilbert spaces Common kernel func-tions in SVM are polynomial kernel RBF kernel Laplaciankernel and Sigmoid kernel -is paper uses the RBF kernelfunction also known as the Gaussian kernel whose cor-responding mapping function can map the sample space toan infinite dimensional space -e analytical expression isshown in the following equation

K X1 X2( 1113857 exp minus X1 minus X221113872 11138732σ2)1113872 11138731113872 (3)

232 Logistic Regression for Multiclassification LR is ageneralized linear regression analysis model that can handleboth classification and regression problems -e improvedLR can even handle multiclassification problems [39] -erelationship between the intermediate value y and the input

x represents the linear part of the model which isy 1113936

mi1(wi lowast xi) + b where wi is the weight matrix and b is

the bias -rough the obtained y the final result of theregression function can be obtained after passing theSoftmax function -e formula for Softmax isS(y) (1(1 + eminus y)) whose the value range is distributed in[01]

233 Deep Neural Network -e neural network was firstdeveloped into an important part [40] of machine learningby being inspired by neurons in the brain-e basis of neuralnetworks is the perceptron model [41] A deep neuralnetwork is a neural network that contains multiple layers ofhidden layers DNN can be divided into three categoriesaccording to the location of different layers input layerhidden layer and output layer

-e deep learning model designed in this paper is shownin Figure 3 -ere are three hidden layers -e number ofneurons in each layer is 300 200 and 100 respectively -eactivation function is ReLU function Each layer usesDropout and BatchNorm2d regularization to help preventoverfitting of the training -e loss function is the cross-entropy loss function and the batch_size size is 32

C minus1n

1113944x

[y ln a +(1 minus y)ln(1 minus z)] (4)

-e cross-entropy represents the distance between theactual output (probability) and the expected output(probability) that is the smaller the value of the cross-entropy the closer the two probability distributions are It isvery helpful for the training of machine learning [42]

elow (t)

Residualsignal

IMF-1IMF-2IMF-3IMF-4IMF-5

IMF-n

helliphellip

helliphellip

m (t) = [eupp(t) + elow(t)]2

m (t)x (t)

eupp (t)

x (t)

Figure 1 flowchart of the proposed circle-EMD

11 9 7 5 3 1 1 3 5 7 9 11

IMF numberfM

RI v

alue

IMF number

1510

50

ndash5ndash10ndash15

20

ndash2011 9 7 5 3 1 1 3 5 7 9 11

Figure 2 IMF signals with different frequency domains decom-posed by circle-EMD

Computational Intelligence and Neuroscience 3

Assuming that the probability distribution y is the desiredoutput the probability distribution z is the actual output-e formula for the cross-entropy loss function is shown inthe following equation

We used 10-fold cross-validation to evaluate the modelresults -e data are divided into 10 groups and 9 experi-ments are used as the training set and 1 group is used as thetest set for each experiment A total of 10 experiments wereperformed so each group of data could be calculated as a testset Finally the accuracy of all test sets was averaged toobtain the final experimental results

3 Experimental Process and Results

-e experimental process is shown in Figure 4 fMRI data arefirst preprocessed by the Brain Decoder Toolbox and thenIMF signals are obtained by circle-EMD decomposition-ree different machine learning tools are used as experi-ments to verify the effective role of circle-EMD

31 fMRI Data and Preprocessing -e data and fMRIdecoding tools used in this article were obtained from thewebsite (httpwwwcnsatrjpdniendownloadsbrain-decoder-toolbox) -e data in this paper were obtainedon a 15 T MRI scanner (Shimadzu-Marconi) with a TRTEflip angle 5 s50ms90 degree field of view (FOV)

192mm acquisition matrix 64times 64 for functional imagesand 50 slices During the fMRI data acquisition process thesubjects performed three gestures of stone scissors andpaper in the MRI according to the instructions -ere aretotally 10 runs and each run has a 20-second break at thebeginning and end In each run there are 32 states including8 rest states and 24 task states so the task-state data have 240samples We divided into 10-fold cross-validation 216samples as the training set and 24 samples as the test set perexperiment as shown in Table 1

We use the Brain DecoderToolbox tool (the process flowand parameters are shown in Table 2) to smooth decontourand regularize the fMRI data and filter out the data of the sixROIs most relevant to the pose which are ldquoM1_RHandrdquo

ldquoSMA_RHandrdquo ldquoCB_RHandrdquo ldquoM1_LHandrdquo ldquoSMA_L-Handrdquo and ldquoCB_LHandrdquo thus obtaining 5333 voxels foreach sample -e locations of each ROI data are shown inTable 3 Based on the statistics of each voxelchannel (thedata value is distributed between minus 544 and 2412919) wefilter out the top 200 data with the highest statistical value(the data value is distributed between 599434 and 2412919)to initially filter unnecessary redundancy (-rough theactual classification and comparison in the experiment if theoriginal 5333 data are directly processed and classifiedwithout the filter the accuracy of the classification is rela-tively low whether the SVM LR or DNN method is used)

By counting the position and number of six ROIs in thefMRI data sequence we obtain the results shown in Table 3

32 Decomposing Data by Using Circle-EMD

321 Decomposed IMF Signal As shown in Figure 5 it is apartial result of the circle-EMD decomposition of the pre-processed fMRI data sequence where the ldquoCurrent IMFrdquo isthe sequence number of the decomposed IMF signal and theldquoSift Iterrdquo is the number of iterations for extracting the IMFsignal For our data set the number of IMFs decomposed isroughly between 4 and 7

Using the original EMD and the circle-EMD algorithmproposed in this paper we decompose the preprocessedfMRI data for each IMF signal as shown in Figure 6 -euppermost black signal is an original data signal with a labelof stone and each IMF signal is decomposed by two EMDmethods (the IMF signals are sequentially decomposed fromtop to bottom) By comparing the distribution results inTable 1 we find that the IMF signals obtained by the twodecompositions are very similar in the high-frequency partFor example in ldquoCB_RHandrdquo these data have strongcharacteristic signals at both high and low frequenciesHowever in the low-frequency part the signal character-istics are more obvious such as the data in ldquoM1_RHandrdquo

-e frequency ranges for each IMF in Figure 6 are shownin Figure 7 For the convenience of display fast Fouriertransform (FFT) is used to transform these spatial IMF

Stone

Scissors

Paper

Inputs = 200

hellip

hellip

hellip

Hidden layerfc1 = 300 Hidden layer

fc2 = 200 Hidden layerfc3 = 100

Weights 1200 times 300

Weights 2300 times 200

Weights 3200 times 100

Weights 4100 times 3

Somax layeroutputs = 3

Figure 3 Deep learning model in this paper

4 Computational Intelligence and Neuroscience

Assuming that the probability distribution y is the desiredoutput the probability distribution z is the actual output-e formula for the cross-entropy loss function is shown inthe following equation

We used 10-fold cross-validation to evaluate the modelresults -e data are divided into 10 groups and 9 experi-ments are used as the training set and 1 group is used as thetest set for each experiment A total of 10 experiments wereperformed so each group of data could be calculated as a testset Finally the accuracy of all test sets was averaged toobtain the final experimental results

3 Experimental Process and Results

-e experimental process is shown in Figure 4 fMRI data arefirst preprocessed by the Brain Decoder Toolbox and thenIMF signals are obtained by circle-EMD decomposition-ree different machine learning tools are used as experi-ments to verify the effective role of circle-EMD

31 fMRI Data and Preprocessing -e data and fMRIdecoding tools used in this article were obtained from thewebsite (httpwwwcnsatrjpdniendownloadsbrain-decoder-toolbox) -e data in this paper were obtainedon a 15 T MRI scanner (Shimadzu-Marconi) with a TRTEflip angle 5 s50ms90 degree field of view (FOV)

192mm acquisition matrix 64times 64 for functional imagesand 50 slices During the fMRI data acquisition process thesubjects performed three gestures of stone scissors andpaper in the MRI according to the instructions -ere aretotally 10 runs and each run has a 20-second break at thebeginning and end In each run there are 32 states including8 rest states and 24 task states so the task-state data have 240samples We divided into 10-fold cross-validation 216samples as the training set and 24 samples as the test set perexperiment as shown in Table 1

We use the Brain DecoderToolbox tool (the process flowand parameters are shown in Table 2) to smooth decontourand regularize the fMRI data and filter out the data of the sixROIs most relevant to the pose which are ldquoM1_RHandrdquo

ldquoSMA_RHandrdquo ldquoCB_RHandrdquo ldquoM1_LHandrdquo ldquoSMA_L-Handrdquo and ldquoCB_LHandrdquo thus obtaining 5333 voxels foreach sample -e locations of each ROI data are shown inTable 3 Based on the statistics of each voxelchannel (thedata value is distributed between minus 544 and 2412919) wefilter out the top 200 data with the highest statistical value(the data value is distributed between 599434 and 2412919)to initially filter unnecessary redundancy (-rough theactual classification and comparison in the experiment if theoriginal 5333 data are directly processed and classifiedwithout the filter the accuracy of the classification is rela-tively low whether the SVM LR or DNN method is used)

By counting the position and number of six ROIs in thefMRI data sequence we obtain the results shown in Table 3

32 Decomposing Data by Using Circle-EMD

321 Decomposed IMF Signal As shown in Figure 5 it is apartial result of the circle-EMD decomposition of the pre-processed fMRI data sequence where the ldquoCurrent IMFrdquo isthe sequence number of the decomposed IMF signal and theldquoSift Iterrdquo is the number of iterations for extracting the IMFsignal For our data set the number of IMFs decomposed isroughly between 4 and 7

Using the original EMD and the circle-EMD algorithmproposed in this paper we decompose the preprocessedfMRI data for each IMF signal as shown in Figure 6 -euppermost black signal is an original data signal with a labelof stone and each IMF signal is decomposed by two EMDmethods (the IMF signals are sequentially decomposed fromtop to bottom) By comparing the distribution results inTable 1 we find that the IMF signals obtained by the twodecompositions are very similar in the high-frequency partFor example in ldquoCB_RHandrdquo these data have strongcharacteristic signals at both high and low frequenciesHowever in the low-frequency part the signal character-istics are more obvious such as the data in ldquoM1_RHandrdquo

-e frequency ranges for each IMF in Figure 6 are shownin Figure 7 For the convenience of display fast Fouriertransform (FFT) is used to transform these spatial IMF

Stone

Scissors

Paper

Inputs = 200

hellip

hellip

hellip

Hidden layerfc1 = 300 Hidden layer

fc2 = 200 Hidden layerfc3 = 100

Weights 1200 times 300

Weights 2300 times 200

Weights 3200 times 100

Weights 4100 times 3

Somax layeroutputs = 3

Figure 3 Deep learning model in this paper

4 Computational Intelligence and Neuroscience

Preprocessing Brain Decoder Toolbox

Circle-EMD

EMD

SVM (rbf)

Logistic regression

Deep learning

fMRI data

Figure 4 -e whole process in this paper

Table 1 Number of samples per experiment

Data type Rock Scissors PaperTraining data 72 72 72Testing data 8 8 8

Table 2 Process name and parameters used when using the Brain Decoder Toolbox

Order index of process flow Process name Parameters1 shiftData shift 1 (shiftTR)2 fmri_selectRoi rois_use 1 1 1 1 1 13 selectChanByTvals num_chans 200(nVoxels) tvals_min 324 reduceOutliers std_thres 4 num_its 25 detrend_bdtb None6 normByBaseline base_conds 1 47 zNorm_bdtb app_dim 2(alongspace) smode 18 selectConds conds 2 4

Table 3 -e locations of each ROI datum in the fMRI data sequence

ROIname TRTEflip angle SMA_RHand CB_RHand M1_LHand SMA_LHand CB_LHand

Index 5 s50ms90degree

75-76 85ndash88 9395ndash109

110ndash114116ndash151 115 152ndash200 75-76 85ndash88 93

95ndash10967 71ndash74 77ndash84 89ndash92

94

Figure 5 Circle-EMD results of decomposing fMRI signals

Computational Intelligence and Neuroscience 5

signals and the frequency ranges are shown in the range of[minus π2 π2] (only positive frequency is plotted) From Fig-ure 7 it can also prove that the original signal has suc-cessfully been decomposed into IMF signals with differentfrequency bands and energy

By further comparing more results between the twodifferent methods the common EMD algorithm has heavyend effects -e difference between the first and last of thesignal is often interrupted -is means that when cubicspline interpolation is performed there is a phenomenon ofspectral leakage distortion after complementing 0 and theside lobes are large -is kind of signal has serious inter-ference to the subsequent machine learning classificationwhich affects the classification accuracy

-e data signal in this paper should not have a temporalcorrelation nor the order of the data -e IMF signaldecomposed by our algorithm has better continuity at thefirst and last of the signal as shown in Figure 8 -us thesignal does not have a starting position and an endingposition in the actual sense and can be connected in a circleshape And it is truly in the low-frequency signal portion thesignals have more significant data characteristics Since thereis no zero-compensation operation there is no spectraldistortion and the side lobes are extremely low which canhelp improve the accuracy of classification

322 e Role of Different IMFs We use the SVM methodwith RBF kernal to classify each IMF signal and signals formedby IMF combination and view the contribution of each IMFsignal to the classification SVM uses MATLABrsquos libsvm-mattool version 30-1 (httpswwwcsientuedutwsimcjlinlibsvm)the kernel function parameter is set to degree 3 gamma 0Coef0 0 nu 05 cache memory size 100 and tolerance oftermination criterion 1e minus 3 p 01 shrinking 1 and theweight matrix after training is 200lowast 3

-e average accuracy of the 10-fold cross-validationresults is statistical as shown in Table 4 Since the number ofIMFs obtained by decomposing each fMRI signal is notfixed here we only extract the first 4 IMF signals whencomparing the IMF separately and finally show the results ofall IMF accumulations In the right half of the table we haverealistic SVM predictions and real label statistics For ex-ample the third value of 40 in the first row indicates that 40samples were correctly predicted under the 80 samples withthe real label as stones

By comparison we find that the IMF signal decomposedfirst has higher frequency characteristics and also has ahigher contribution to the classification result -e more theIMF signals are superimposed the higher the accuracy of thedata Comparing with Table 5 we found that although theclassification accuracy of a single IMF signal there is no high

20

0

ndash20fMRI

val

ue5

0

ndash5fMRI

val

ue

5

0

ndash5fMRI

val

ue

10

ndash1ndash2

fMRI

val

uefM

RI v

alue

1

0

ndash1fMRI

val

ue

0 50 100 150fMRI number

200 0 50 100 150fMRI number

200

0 50 100 150fMRI number

200 0 50 100 150fMRI number

200

0 50 100 150fMRI number

200 0 50 100 150fMRI number

200

0 50 100 150fMRI number

200 0 50 100 150fMRI number

200

0 50 100 150fMRI number

200 0 50 100 150fMRI number

200

0 50 100 150fMRI number

(a) (b)

200 0 50 100 150fMRI number

200

10

ndash1ndash2

20

0

ndash20fMRI

val

ue

5

0

ndash5fMRI

val

ue

5

0

ndash5fMRI

val

ue

10

ndash1ndash2

fMRI

val

uefM

RI v

alue

1

0

ndash1fMRI

val

ue

10

ndash1ndash2

Originalsignal

IMF 1

IMF 2

IMF 3

IMF 4

IMF 5

Figure 6 Each IMF signal (a) Original EMD (b) circle-EMD

6 Computational Intelligence and Neuroscience

signals and the frequency ranges are shown in the range of[minus π2 π2] (only positive frequency is plotted) From Fig-ure 7 it can also prove that the original signal has suc-cessfully been decomposed into IMF signals with differentfrequency bands and energy

By further comparing more results between the twodifferent methods the common EMD algorithm has heavyend effects -e difference between the first and last of thesignal is often interrupted -is means that when cubicspline interpolation is performed there is a phenomenon ofspectral leakage distortion after complementing 0 and theside lobes are large -is kind of signal has serious inter-ference to the subsequent machine learning classificationwhich affects the classification accuracy

-e data signal in this paper should not have a temporalcorrelation nor the order of the data -e IMF signaldecomposed by our algorithm has better continuity at thefirst and last of the signal as shown in Figure 8 -us thesignal does not have a starting position and an endingposition in the actual sense and can be connected in a circleshape And it is truly in the low-frequency signal portion thesignals have more significant data characteristics Since thereis no zero-compensation operation there is no spectraldistortion and the side lobes are extremely low which canhelp improve the accuracy of classification

322 e Role of Different IMFs We use the SVM methodwith RBF kernal to classify each IMF signal and signals formedby IMF combination and view the contribution of each IMFsignal to the classification SVM uses MATLABrsquos libsvm-mattool version 30-1 (httpswwwcsientuedutwsimcjlinlibsvm)the kernel function parameter is set to degree 3 gamma 0Coef0 0 nu 05 cache memory size 100 and tolerance oftermination criterion 1e minus 3 p 01 shrinking 1 and theweight matrix after training is 200lowast 3

-e average accuracy of the 10-fold cross-validationresults is statistical as shown in Table 4 Since the number ofIMFs obtained by decomposing each fMRI signal is notfixed here we only extract the first 4 IMF signals whencomparing the IMF separately and finally show the results ofall IMF accumulations In the right half of the table we haverealistic SVM predictions and real label statistics For ex-ample the third value of 40 in the first row indicates that 40samples were correctly predicted under the 80 samples withthe real label as stones

By comparison we find that the IMF signal decomposedfirst has higher frequency characteristics and also has ahigher contribution to the classification result -e more theIMF signals are superimposed the higher the accuracy of thedata Comparing with Table 5 we found that although theclassification accuracy of a single IMF signal there is no high

20

0

ndash20fMRI

val

ue5

0

ndash5fMRI

val

ue

5

0

ndash5fMRI

val

ue

10

ndash1ndash2

fMRI

val

uefM

RI v

alue

1

0

ndash1fMRI

val

ue

0 50 100 150fMRI number

200 0 50 100 150fMRI number

200

0 50 100 150fMRI number

200 0 50 100 150fMRI number

200

0 50 100 150fMRI number

200 0 50 100 150fMRI number

200

0 50 100 150fMRI number

200 0 50 100 150fMRI number

200

0 50 100 150fMRI number

200 0 50 100 150fMRI number

200

0 50 100 150fMRI number

(a) (b)

200 0 50 100 150fMRI number

200

10

ndash1ndash2

20

0

ndash20fMRI

val

ue

5

0

ndash5fMRI

val

ue

5

0

ndash5fMRI

val

ue

10

ndash1ndash2

fMRI

val

uefM

RI v

alue

1

0

ndash1fMRI

val

ue

10

ndash1ndash2

Originalsignal

IMF 1

IMF 2

IMF 3

IMF 4

IMF 5

Figure 6 Each IMF signal (a) Original EMD (b) circle-EMD

6 Computational Intelligence and Neuroscience

accuracy of the original signal However after super-imposing each IMF signal that is deleting the residualmonotonic signal the classification accuracy is the highest

33 ClassificationResults ofMachineLearning -ere are 240data sets in this paper In each experiment 216 are trainingdata and 24 are test sets For each training data the input is200-dimensional fMRI data and the output label is a 3-dimensional classified data Using LR SVM and DNN weclassify raw data and data after preprocessed by circle-EMDIn the training process the accuracy curve of the logisticregression and deep learning of the experiment is shown in

the Figure 9 (red lines indicate the training set and greenlines indicate the test set) and the accuracy rate classificationresults are shown in Table 6 To further compare the originalsignal with our circle-EMD method we still do the machinelearning classification statistics for the original signal

From Table 6 we can see that for the fMRI data of thispaper no matter which machine learning method ourcircle-EMD algorithm has achieved higher accuracy whichproves that our circle-EMD method is effective for sepa-rating spatial signals Sex can help to refer to the classifi-cation accuracy of machine learning

To further compare the sensitivity and specificity of theclassification this paper further calculated the weightedaverage of precision recall and f1-score of the classificationresults in Figure 6 (we only show the average value of 10-foldcross-validation results And since the samples are balancedthe results of macroaverage and weighted average are thesame for multiclassification mean calculation so this paperonly shows the results of weighted average)

Comparing the results of each machine learning classifi-cation deep learning performed best in each experiment andthe average accuracy of the final three categories of the test setreached 812 LR and DNN have achieved close to 100classification accuracy on the training set of this data but theperformance of LR on the test set is not enough and it does nothave good generalization -is result not only confirms thevalidity of our circle-EMDmethod but also the importance ofdeep learning for fMRI data classification and learning

100

50

0Am

plitu

de100

50

0Am

plitu

de

50

0Am

plitu

de

50

0Am

plitu

de

50

0Am

plitu

de

100

50

0Am

plitu

de

100

50

0Am

plitu

de

50

0Am

plitu

de

50

0Am

plitu

de50

0Am

plitu

de

0 05 1 15f (Hz)

0 05 1 15f (Hz)

0 05 1 15f (Hz)

0 05 1 15f (Hz)

0 05 1 15f (Hz)

0 05 1 15f (Hz)

0 05 1 15f (Hz)

0 05 1 15f (Hz)

0 05 1 15f (Hz)

0 05 1 15f (Hz)

(a) (b)

IMF 1

IMF 2

IMF 3

IMF 4

IMF 5

Figure 7 FFT frequency of each IMF signal (a) Original EMD (b) circle-EMD

5

0

ndash5

ndash106

42

02

46 6

42

02

46

IMF number

IMF number

fMRI

val

ue

Figure 8 Original signal and each decomposed circle-IMF signal

Computational Intelligence and Neuroscience 7

Table 4 Classification results of the task states of each circle-IMF signal and combined signals by using SVM

IMF Label Rock (80) Scissors (80) Paper (80)Prediction Rock Scissors Paper Rock Scissors Paper Rock Scissors Paper

1 05625 40 20 20 18 43 19 10 18 522 04833 46 25 9 31 26 23 14 22 443 04292 47 12 21 33 20 27 23 21 364 04458 51 19 10 32 30 18 25 29 261 + 2 05875 44 27 9 24 40 16 7 16 571 + 3 060 46 17 17 18 44 18 10 16 541 + 4 0575 41 23 16 17 46 17 11 18 511 + 2 + 3 06333 50 20 10 22 45 13 7 16 571 + 2 + 4 05958 48 22 10 18 43 19 8 20 521 + 2 + 3 + 4 06417 53 18 9 20 45 15 6 18 56All IMFs 06542 53 18 9 19 47 14 7 16 57

Table 5 Classification results of original signals that have not been processed by EMD

Label Rock (80) Scissors (80) Paper (80)Prediction Rock Scissors Paper Rock Scissors Paper Rock Scissors Paper

Original signal 06375 50 21 9 21 45 14 6 16 58

Accu

racy

rate

Epoch number

Accuracy of train and test

TrainTest

1

09

08

07

06

05

04

0 25 50 75 100 125 150

(a)

Accuracy of train and test

Accu

racy

rate

1

09

08

07

06

05

04

Epoch number

TrainTest

0 25 50 75 100 125 150

(b)

Figure 9 Accuracy rate curve (a) Logistic regression (b) Deep neural networks

Table 6 Accuracy results of extracted fMRI signals by using different machine learning algorithms

Experiment numberLogistic regression SVM Deep learning

Original All IMFs Original All IMFs Original All IMFs1 0667 0750 0708 0708 0833 08752 0667 0708 0667 0708 0792 08753 0625 0667 0708 0708 0708 08334 0667 0708 0583 0625 0667 07925 0667 0625 0583 0542 0750 07506 0708 0708 0625 0667 0750 07927 0708 0667 0583 0583 0792 08338 0667 0667 0542 0583 0542 07089 0708 0750 0625 0667 0750 083310 0750 0750 0750 0750 0750 0833Average 0683 0700 0638 0654 0733 0812

8 Computational Intelligence and Neuroscience

4 Conclusions

In this paper we propose a circle-EMD algorithm to de-compose and preprocess the ldquostone-scissors-paperrdquo multi-task state fMRI data and compare it with the original EMDmethod On this basis the three commonly used machinelearning methods logistic regression SVM and DNN areused to process the data obtained by different EMD algo-rithms We compare the effects of different EMD algorithmson the processing results through experiments and finallyobtain the classification accuracy of different data by dif-ferent machine learning methods

-e results show that the proposed algorithm caneliminate the end effect of spatial signals and extract signalswith the different spatial frequency features of multitaskstate data more effectively It also reduces the effects of lowfrequencies and noise more accurately reflects the datacharacteristics of the original signal (especially the char-acteristics of the relatively low-frequency part) shows thecorrelation between different ROIs and finally achieveshigher classification accuracy By comparing the variousIMF signals extracted by this method it is found that thecontribution from low frequency to high frequency increasesin turn and the accumulation of multiple IMF signals helpsto improve the accuracy of classification -rough com-parison with different machine learning algorithms we alsofound that the most effective algorithm is the deep learningalgorithm -e highest accuracy in this experiment is up to875 with an average of 812

In general themethod of this paper provides a new circle-EMD-based data analysis method for analyzing task-statefMRI data and combines the deep learningmethod to providea more effective solution for classification of multitask statefMRI data helping to better realize the brain-computer in-teraction -e limitation of the combination of the algorithmand machine learning is that when processing spatial fMRIsignals with small amount of data the effect is more obvious(for example 200 voxels are selected in this paper) and theuse of decomposed IMF signals is only verified in the motiondata type of this paper In future research work we willexperiment with more types of fMRI data and try to weighteach IMF signal to get a better performance and pay moreattention to the residual signal when using circle-EMD

Data Availability

For original fMRI data address please visit httpswwwnitrcorgprojectsbdtb -e processed data used to supportthe findings of this study are available from the corre-sponding author upon request

Conflicts of Interest

-e authors declare that they have no conflicts of interest

Acknowledgments

-is work was supported by the National Natural ScienceFoundation of China (61271351 and 41827807)

References

[1] M Huang P Wu Y Liu and H Chen ldquoApplication andcontrast in brain-computer interface between Hilbert-Huangtransform and wavelet transformrdquo in Proceedings of the 9thInternational Conference for Young Computer ScientistsHunan China November 2008

[2] L Xingfeng Mechanisms of Acupuncture amp MoxibustionAnalgesia and Brain Activation Regions Localization Algo-rithm Studies Institute of Automation Chinese Academy ofSciences Beijing China 2004

[3] W Jiang H Liu J Liao et al ldquoA functional MRI study ofdeception among offenders with antisocial personality dis-ordersrdquo Neuroscience vol 244 no 3 pp 90ndash98 2013

[4] S S Kannurpatti B Rypma and B B Biswal ldquoPrediction oftask-related BOLD fMRI with amplitude signatures of resting-state fMRIrdquo Frontiers in Systems Neuroscience vol 6 no 6p 7 2012

[5] J Chang M Zhang G Hitchman J Qiu and Y Liu ldquoWhenyou smile you become happy evidence from resting statetask-based fMRIrdquo Biological Psychology vol 103 pp 100ndash1062014

[6] D G Laura T Silvia P Nikolaos et al ldquo-e role of fMRI inthe assessment of neuroplasticity in MS a systematic reviewrdquoNeural Plasticity vol 2018 Article ID 3419871 27 pages 2018

[7] M Dong W Qin J Sun et al ldquoTempo-spatial analysis ofvision-related acupoint specificity in the occipital lobe usingfMRI an ICA studyrdquo Brain Research vol 1436 no 2pp 34ndash42 2012

[8] X-N Zuo A D Martino C Kelly et al ldquo-e oscillatingbrain complex and reliablerdquo Neuroimage vol 49 no 2pp 1432ndash1445 2010

[9] N E Huang Z Shen and S R Long ldquoA new view ofnonlinear water waves the Hilbert spectrumrdquo Annual Reviewof Fluid Mechanics vol 31 no 1 pp 417ndash457 1999

[10] C F Lin S W Yeh S H Chang et al An HHT-Based Time-Frequency Scheme for Analyzing the EEG Signals of ClinicalAlcoholics Nova Science Publishers New York NY USA2010

[11] P LeeW H Tsui Y Teng and T-C Hsiao ldquoAffective patternanalysis of image in frequency domain using the Hilbert-Huang transformrdquo in Proceedings of the 2012 IEEE AsiaPacific Conference on Circuits and Systems KaohsiungTaiwan December 2012

Table 7 Sensitivity and specificity of the classification

Weighted averageLogistic regression SVM Deep learning

Original All IMFs Original All IMFs Original All IMFsPrecision 0690 0711 0468 0486 0745 0825Recall 0683 0700 0638 0654 0733 0812f1-score 0681 0696 0538 0557 0729 0806

Computational Intelligence and Neuroscience 9

Table 4 Classification results of the task states of each circle-IMF signal and combined signals by using SVM

IMF Label Rock (80) Scissors (80) Paper (80)Prediction Rock Scissors Paper Rock Scissors Paper Rock Scissors Paper

1 05625 40 20 20 18 43 19 10 18 522 04833 46 25 9 31 26 23 14 22 443 04292 47 12 21 33 20 27 23 21 364 04458 51 19 10 32 30 18 25 29 261 + 2 05875 44 27 9 24 40 16 7 16 571 + 3 060 46 17 17 18 44 18 10 16 541 + 4 0575 41 23 16 17 46 17 11 18 511 + 2 + 3 06333 50 20 10 22 45 13 7 16 571 + 2 + 4 05958 48 22 10 18 43 19 8 20 521 + 2 + 3 + 4 06417 53 18 9 20 45 15 6 18 56All IMFs 06542 53 18 9 19 47 14 7 16 57

Table 5 Classification results of original signals that have not been processed by EMD

Label Rock (80) Scissors (80) Paper (80)Prediction Rock Scissors Paper Rock Scissors Paper Rock Scissors Paper

Original signal 06375 50 21 9 21 45 14 6 16 58

Accu

racy

rate

Epoch number

Accuracy of train and test

TrainTest

1

09

08

07

06

05

04

0 25 50 75 100 125 150

(a)

Accuracy of train and test

Accu

racy

rate

1

09

08

07

06

05

04

Epoch number

TrainTest

0 25 50 75 100 125 150

(b)

Figure 9 Accuracy rate curve (a) Logistic regression (b) Deep neural networks

Table 6 Accuracy results of extracted fMRI signals by using different machine learning algorithms

Experiment numberLogistic regression SVM Deep learning

Original All IMFs Original All IMFs Original All IMFs1 0667 0750 0708 0708 0833 08752 0667 0708 0667 0708 0792 08753 0625 0667 0708 0708 0708 08334 0667 0708 0583 0625 0667 07925 0667 0625 0583 0542 0750 07506 0708 0708 0625 0667 0750 07927 0708 0667 0583 0583 0792 08338 0667 0667 0542 0583 0542 07089 0708 0750 0625 0667 0750 083310 0750 0750 0750 0750 0750 0833Average 0683 0700 0638 0654 0733 0812

8 Computational Intelligence and Neuroscience

4 Conclusions

In this paper we propose a circle-EMD algorithm to de-compose and preprocess the ldquostone-scissors-paperrdquo multi-task state fMRI data and compare it with the original EMDmethod On this basis the three commonly used machinelearning methods logistic regression SVM and DNN areused to process the data obtained by different EMD algo-rithms We compare the effects of different EMD algorithmson the processing results through experiments and finallyobtain the classification accuracy of different data by dif-ferent machine learning methods

-e results show that the proposed algorithm caneliminate the end effect of spatial signals and extract signalswith the different spatial frequency features of multitaskstate data more effectively It also reduces the effects of lowfrequencies and noise more accurately reflects the datacharacteristics of the original signal (especially the char-acteristics of the relatively low-frequency part) shows thecorrelation between different ROIs and finally achieveshigher classification accuracy By comparing the variousIMF signals extracted by this method it is found that thecontribution from low frequency to high frequency increasesin turn and the accumulation of multiple IMF signals helpsto improve the accuracy of classification -rough com-parison with different machine learning algorithms we alsofound that the most effective algorithm is the deep learningalgorithm -e highest accuracy in this experiment is up to875 with an average of 812

In general themethod of this paper provides a new circle-EMD-based data analysis method for analyzing task-statefMRI data and combines the deep learningmethod to providea more effective solution for classification of multitask statefMRI data helping to better realize the brain-computer in-teraction -e limitation of the combination of the algorithmand machine learning is that when processing spatial fMRIsignals with small amount of data the effect is more obvious(for example 200 voxels are selected in this paper) and theuse of decomposed IMF signals is only verified in the motiondata type of this paper In future research work we willexperiment with more types of fMRI data and try to weighteach IMF signal to get a better performance and pay moreattention to the residual signal when using circle-EMD

Data Availability

For original fMRI data address please visit httpswwwnitrcorgprojectsbdtb -e processed data used to supportthe findings of this study are available from the corre-sponding author upon request

Conflicts of Interest

-e authors declare that they have no conflicts of interest

Acknowledgments

-is work was supported by the National Natural ScienceFoundation of China (61271351 and 41827807)

References

[1] M Huang P Wu Y Liu and H Chen ldquoApplication andcontrast in brain-computer interface between Hilbert-Huangtransform and wavelet transformrdquo in Proceedings of the 9thInternational Conference for Young Computer ScientistsHunan China November 2008

[2] L Xingfeng Mechanisms of Acupuncture amp MoxibustionAnalgesia and Brain Activation Regions Localization Algo-rithm Studies Institute of Automation Chinese Academy ofSciences Beijing China 2004

[3] W Jiang H Liu J Liao et al ldquoA functional MRI study ofdeception among offenders with antisocial personality dis-ordersrdquo Neuroscience vol 244 no 3 pp 90ndash98 2013

[4] S S Kannurpatti B Rypma and B B Biswal ldquoPrediction oftask-related BOLD fMRI with amplitude signatures of resting-state fMRIrdquo Frontiers in Systems Neuroscience vol 6 no 6p 7 2012

[5] J Chang M Zhang G Hitchman J Qiu and Y Liu ldquoWhenyou smile you become happy evidence from resting statetask-based fMRIrdquo Biological Psychology vol 103 pp 100ndash1062014

[6] D G Laura T Silvia P Nikolaos et al ldquo-e role of fMRI inthe assessment of neuroplasticity in MS a systematic reviewrdquoNeural Plasticity vol 2018 Article ID 3419871 27 pages 2018

[7] M Dong W Qin J Sun et al ldquoTempo-spatial analysis ofvision-related acupoint specificity in the occipital lobe usingfMRI an ICA studyrdquo Brain Research vol 1436 no 2pp 34ndash42 2012

[8] X-N Zuo A D Martino C Kelly et al ldquo-e oscillatingbrain complex and reliablerdquo Neuroimage vol 49 no 2pp 1432ndash1445 2010

[9] N E Huang Z Shen and S R Long ldquoA new view ofnonlinear water waves the Hilbert spectrumrdquo Annual Reviewof Fluid Mechanics vol 31 no 1 pp 417ndash457 1999

[10] C F Lin S W Yeh S H Chang et al An HHT-Based Time-Frequency Scheme for Analyzing the EEG Signals of ClinicalAlcoholics Nova Science Publishers New York NY USA2010

[11] P LeeW H Tsui Y Teng and T-C Hsiao ldquoAffective patternanalysis of image in frequency domain using the Hilbert-Huang transformrdquo in Proceedings of the 2012 IEEE AsiaPacific Conference on Circuits and Systems KaohsiungTaiwan December 2012

Table 7 Sensitivity and specificity of the classification

Weighted averageLogistic regression SVM Deep learning

Original All IMFs Original All IMFs Original All IMFsPrecision 0690 0711 0468 0486 0745 0825Recall 0683 0700 0638 0654 0733 0812f1-score 0681 0696 0538 0557 0729 0806

Computational Intelligence and Neuroscience 9

4 Conclusions

In this paper we propose a circle-EMD algorithm to de-compose and preprocess the ldquostone-scissors-paperrdquo multi-task state fMRI data and compare it with the original EMDmethod On this basis the three commonly used machinelearning methods logistic regression SVM and DNN areused to process the data obtained by different EMD algo-rithms We compare the effects of different EMD algorithmson the processing results through experiments and finallyobtain the classification accuracy of different data by dif-ferent machine learning methods

-e results show that the proposed algorithm caneliminate the end effect of spatial signals and extract signalswith the different spatial frequency features of multitaskstate data more effectively It also reduces the effects of lowfrequencies and noise more accurately reflects the datacharacteristics of the original signal (especially the char-acteristics of the relatively low-frequency part) shows thecorrelation between different ROIs and finally achieveshigher classification accuracy By comparing the variousIMF signals extracted by this method it is found that thecontribution from low frequency to high frequency increasesin turn and the accumulation of multiple IMF signals helpsto improve the accuracy of classification -rough com-parison with different machine learning algorithms we alsofound that the most effective algorithm is the deep learningalgorithm -e highest accuracy in this experiment is up to875 with an average of 812

In general themethod of this paper provides a new circle-EMD-based data analysis method for analyzing task-statefMRI data and combines the deep learningmethod to providea more effective solution for classification of multitask statefMRI data helping to better realize the brain-computer in-teraction -e limitation of the combination of the algorithmand machine learning is that when processing spatial fMRIsignals with small amount of data the effect is more obvious(for example 200 voxels are selected in this paper) and theuse of decomposed IMF signals is only verified in the motiondata type of this paper In future research work we willexperiment with more types of fMRI data and try to weighteach IMF signal to get a better performance and pay moreattention to the residual signal when using circle-EMD

Data Availability

For original fMRI data address please visit httpswwwnitrcorgprojectsbdtb -e processed data used to supportthe findings of this study are available from the corre-sponding author upon request

Conflicts of Interest

-e authors declare that they have no conflicts of interest

Acknowledgments

-is work was supported by the National Natural ScienceFoundation of China (61271351 and 41827807)

References

[1] M Huang P Wu Y Liu and H Chen ldquoApplication andcontrast in brain-computer interface between Hilbert-Huangtransform and wavelet transformrdquo in Proceedings of the 9thInternational Conference for Young Computer ScientistsHunan China November 2008

[2] L Xingfeng Mechanisms of Acupuncture amp MoxibustionAnalgesia and Brain Activation Regions Localization Algo-rithm Studies Institute of Automation Chinese Academy ofSciences Beijing China 2004

[3] W Jiang H Liu J Liao et al ldquoA functional MRI study ofdeception among offenders with antisocial personality dis-ordersrdquo Neuroscience vol 244 no 3 pp 90ndash98 2013

[4] S S Kannurpatti B Rypma and B B Biswal ldquoPrediction oftask-related BOLD fMRI with amplitude signatures of resting-state fMRIrdquo Frontiers in Systems Neuroscience vol 6 no 6p 7 2012

[5] J Chang M Zhang G Hitchman J Qiu and Y Liu ldquoWhenyou smile you become happy evidence from resting statetask-based fMRIrdquo Biological Psychology vol 103 pp 100ndash1062014

[6] D G Laura T Silvia P Nikolaos et al ldquo-e role of fMRI inthe assessment of neuroplasticity in MS a systematic reviewrdquoNeural Plasticity vol 2018 Article ID 3419871 27 pages 2018

[7] M Dong W Qin J Sun et al ldquoTempo-spatial analysis ofvision-related acupoint specificity in the occipital lobe usingfMRI an ICA studyrdquo Brain Research vol 1436 no 2pp 34ndash42 2012

[8] X-N Zuo A D Martino C Kelly et al ldquo-e oscillatingbrain complex and reliablerdquo Neuroimage vol 49 no 2pp 1432ndash1445 2010

[9] N E Huang Z Shen and S R Long ldquoA new view ofnonlinear water waves the Hilbert spectrumrdquo Annual Reviewof Fluid Mechanics vol 31 no 1 pp 417ndash457 1999

[10] C F Lin S W Yeh S H Chang et al An HHT-Based Time-Frequency Scheme for Analyzing the EEG Signals of ClinicalAlcoholics Nova Science Publishers New York NY USA2010

[11] P LeeW H Tsui Y Teng and T-C Hsiao ldquoAffective patternanalysis of image in frequency domain using the Hilbert-Huang transformrdquo in Proceedings of the 2012 IEEE AsiaPacific Conference on Circuits and Systems KaohsiungTaiwan December 2012

Table 7 Sensitivity and specificity of the classification

Weighted averageLogistic regression SVM Deep learning

Original All IMFs Original All IMFs Original All IMFsPrecision 0690 0711 0468 0486 0745 0825Recall 0683 0700 0638 0654 0733 0812f1-score 0681 0696 0538 0557 0729 0806

Computational Intelligence and Neuroscience 9

[12] C Xia Y Wu Q Lu and B Ju ldquoSurface characteristic profileextraction based on Hilbert-Huang transformrdquoMeasurementvol 47 no 1 pp 306ndash313 2014

[13] P C Tay ldquoAM-FM image analysis using the Hilbert Huangtransformrdquo in Proceedings of the IEEE Southwest Symposiumon Image Analysis amp Interpretation Santa Fe NM USAMarch 2008

[14] N E Huang and S S P Shen Hilbert-Huang Transform andits Applications World Scientific Singapore 2005

[15] Z Lina ldquoA survey of time-frequency analysis methods fordigital signal processingrdquo Information Technology vol 6pp 26ndash28 2013

[16] L Peng L Fei Q Jiang J Chen and X Li ldquoBased on theimproved HHT and its application in the power quality de-tection of microgridrdquo in Proceedings of the InternationalConference on Electrical Machines amp Systems Beijing ChinaAugust 2011

[17] m Gong and l Xie ldquoDiscussion on the application of HHTmethod to earthquake engineeringrdquo World Information onEarthquake Engineering vol 19 no 3 pp 39ndash43 2003

[18] H Yu F Li T Wu et al ldquoFunctional brain abnormalities inmajor depressive disorder using the Hilbert-Huang trans-formrdquo Brain Imaging amp Behavior vol 12 no 6pp 1556ndash1568 2018

[19] Y DengWWang C Qian ZWang and D Dai ldquoBoundary-processing-technique in EMD method and Hilbert trans-formrdquo Chinese Science Bulletin vol 46 no 11 pp 954ndash9602001

[20] J W Yang and M P Jia ldquoStudy on processing method andanalysis of end problem of Hilbert-Huang spectrumrdquo Journalof Vibration Engineering vol 19 no 2 pp 283ndash288 2006

[21] J Cheng D Yu and Y Yang ldquoApplication of support vectorregression machines to the processing of end effects of Hil-bert-Huang transformrdquo Mechanical Systems and SignalProcessing vol 21 no 3 pp 1197ndash1211 2007

[22] J Yang P Li Y Yang and D Xu ldquoAn improved EMDmethod for modal identification and a combined static-dy-namic method for damage detectionrdquo Journal of Sound andVibration vol 420 pp 242ndash260 2018

[23] T Wang and L Gang ldquoAn improved method to solve the endeffect of EMD and its application on vibration signalrdquo inProceedings of the International Conference onMechatronics ampAutomation Changchun China August 2009

[24] G Tai and Z Deng ldquoAn improved EMDmethod based on themulti-objective optimization and its application to faultfeature extraction of rolling bearingrdquo Applied Acousticsvol 127 pp 46ndash62 2017

[25] A Khazaee A Ebrahimzadeh and A Babajani-FeremildquoApplication of advanced machine learning methods onresting-state fMRI network for identification of mild cognitiveimpairment and Alzheimerrsquos diseaserdquo Brain Imaging andBehavior vol 10 no 3 pp 799ndash817 2016

[26] A Panat A Patil and R Chauhan ldquoFeature extraction andclassification OF FMRI imagesrdquo International Journal ofElectrical Electronics amp Data Communication vol 2 no 102014

[27] B Afshin-Pour S-M Shams and S Strother ldquoA hybridLDA+gCCA model for fMRI data classification and visuali-zationrdquo IEEE Transactions on Medical Imaging vol 34 no 5pp 1031ndash1041 2015

[28] S Ryali K Supekar D A Abrams and V Menon ldquoSparselogistic regression for whole-brain classification of fMRIdatardquo Neuroimage vol 51 no 2 pp 752ndash764 2010

[29] S Koyamada Y Shikauchi K Nakae and M Koyama ldquoDeeplearning of fMRI big data a novel approach to subject-transferdecodingrdquo Computer Science httparxivorgabs150200093(2015) 2015

[30] S Sarraf and G Tofighi ldquoClassification of Alzheimerrsquos diseaseusing fMRI data and deep learning convolutional neuralnetworksrdquo 2016 httparxivorgabs1603086312016

[31] T Horikawa and Y Kamitani ldquoGeneric decoding of seen andimagined objects using hierarchical visual featuresrdquo NatureCommunications vol 8 no 1 Article ID 15037 2017

[32] X Du Y Cai S Wang and L Zhang ldquoOverview of deeplearningrdquo in Proceedings of the 2016 31st Youth AcademicAnnual Conference of Chinese Association of Automation(YAC) Wuhan China November 2017

[33] W Xin Y Ren Y Yang W Zhang and N N Xiong ldquoAweighted discriminative dictionary learning method for de-pression disorder classification using fMRI datardquo in Pro-ceedings of the IEEE International Conferences on Big Data ampCloud Computing (BDCloud) Social Computing and Net-working (SocialCom) Sustainable Computing and Commu-nications (SustainCom) (BDCloud-SocialCom-SustainCom)Atlanta GA USA October 2016

[34] I O Ertugrul M Ozay and F T Y Vural ldquoEncoding thelocal connectivity patterns of fMRI for cognitive task and stateclassificationrdquo Brain Imaging amp Behavior vol 13 no 4pp 893ndash904 2019

[35] N E Huang Z Shen S R Long and M L C Wu ldquo-eempirical mode decomposition and the Hilbert spectrum fornonlinear and non-stationary time series analysisrdquo Proceed-ings of the Royal Society of London Series A MathematicalPhysical and Engineering Sciences vol 454 no 1971pp 903ndash995 1998

[36] W-Z Lu and W-J Wang ldquoPotential assessment of theldquosupport vector machinerdquo method in forecasting ambient airpollutant trendsrdquo Chemosphere vol 59 no 5 pp 693ndash7012005

[37] M Gocic S Shamshirband Z Razak D Petkovic S Ch andS Trajkovic ldquoLong-term precipitation analysis and estimationof precipitation concentration index using three supportvector machine methodsrdquo Advances in Meteorology vol 2016Article ID 7912357 11 pages 2016

[38] S Sun ldquoA survey of multi-view machine learningrdquo NeuralComputing amp Applications vol 23 no 7-8 pp 2031ndash20382013

[39] B Zou ldquoMultiple classification using logistic regressionmodelrdquo in Lecture Notes in Computer Science C H HsuS Wang A Zhou and A Shawkat Eds Springer ChamSwitzerland 2016

[40] K Luo J Li ZWang and A Cuschieri ldquoPatient-specific deeparchitectural model for ECG classificationrdquo Journal ofHealthcare Engineering vol 2017 Article ID 410872013 pages 2017

[41] F Amato N Mazzocca F Moscato and E VivenzioldquoMultilayer perceptron an intelligent model for classificationand intrusion detectionrdquo in Proceedings of the InternationalConference on Advanced Information Networking amp Appli-cations Workshops Taipei Taiwan March 2017

[42] G Valvano G Santini N Martini et al ldquoConvolutionalneural networks for the segmentation of microcalcification inmammography imagingrdquo Journal of Healthcare Engineeringvol 2019 Article ID 9360941 9 pages 2019

10 Computational Intelligence and Neuroscience