INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 55

Muscular-Contraction Classification :Comparison Study Between IndependentComponent Analysis and Artificial Neural

Network

D. Sueaseenak1∗, T. Chanwimalueang2,

W. Iampa1, Guest members, and M. Sangworasil1, Member

ABSTRACT

We developed a multi-channel electromyogram ac-quisition system using PSOC microcontroller to ac-quire multi-channel EMG signals. An array of 4 x 4surface electrodes was used to record the EMG sig-nal. The obtained signals were classified by a back-propagation-type artificial neural network. B-splineinterpolation technique has been utilized to map theEMG signal on the muscle surface. The topologi-cal mapping of the EMG is then analyzed to classifythe pattern of muscle contraction using independentcomponent analysis. The proposed system was suc-cessfully demonstrated to record EMG data and itssurface mapping. The comparison study of muscu-lar contraction classification using independent com-ponent analysis and artificial neural network demon-strates shows that performance of ANN classificationis as comparable as that of the ICA. The computa-tional time of ANN is also less than that of the ICA.

Keywords: Electromyography, Principal Compo-nent Analysis, Independent Component Analysis, Ar-tificial Neural Network

1. INTRODUCTION

Electromyography (EMG) is the study of muscleelectrical signals. EMG is sometimes referred to asmyoelectric activity. Many muscular abnormalitiessuch as muscular dystrophy, inflammation of mus-cle, peripheral nerve damages could results in anabnormal electromyogram [1-6]. Figure 1 shows aschematic representation of muscle units and its com-ponents. EMG can be recorded by two types of elec-trodes; invasive electrode the so-called wire or nee-dle electrodes and non-invasive electrode the so-called

* Corresponding author.Manuscript received on December 25, 2009.,

1D. Sueaseenak, W. Iampa and M. Sangworasil, Depart-ment of Electronics, Faculty of Engineering, King Mongkut’sInstitute of Technology Ladkrabang, Thailand

2T. Chanwimalueang, Biomedical Medical Engineering pro-gramme, Faculty of Engineering, Srinakharinwirot University,Thailand

E-mail addresses: emg7849@gmail.com (D. Sueaseenak)

Fig.1: Muscle unit [7]

surface electrode. Wire or needle electrodes recordsindividual muscle fiber action potentials which is anideal choice to evaluate the muscle activity [10]. How-ever, fine wire intramuscular electrodes require a nee-dle for insertion into the muscle and may cause a sig-nificant pain. The choice of surface electrode is thenpreferable. However, when EMG is acquired fromsurface electrodes mounted directly on the skin, thesignal is a composite of all the muscle fiber action po-tentials occurring in the muscles underlying the skin.Estimating this force in general is a hard problemdue to difficulties in activating a single muscle in iso-lation, isolating the signal generated by a muscle fromthat of its neighbors, and other associated problems[8-9]. The clinical application of EMG can be classi-fied into two mains categories. (i) Standard EMG [9]is recorded from discrete sites on a muscle and thusprovides only a limited picture of the actual muscularelectrical activity in the vicinity of the recording elec-trode. (ii) Array EMG recorded by an array electrodewhich facilitates the clinical interpretation of electri-cal activities through the mapping of these signals onthe muscle surface [10]. In this paper, a PSOC-basedmulti-channel surface electrode array data acquisi-tion system is developed to acquire EMG data. TheEMG signals are then mapped using B-spline inter-polation technique. The EMG topological Mappingis then used for classification of muscular contraction.There exist a number of 2D-pattern classifications

56 D. Sueaseenak et al: Muscular-Contraction Classification : Comparison Study ... (55-62)

Fig.2: Muscular contraction classification system

[10-13]. In this research, we compared the EMG-contraction classification technique between indepen-dent component analysis (ICA) and artifial neuralnetwork (ANN) [15]. The results show that the ANNclassification yields as good performance as the ICAin the faster computational time.

The paper is organized as follows: Section 2 is de-voted to the designed concept of multi-channel elec-tromyogram system. Section 3 briefly reviews fourier-based features, Section 4 describes feature extractionand topological mapping process. Section 5 brieflyintroduces Independent Component Analysis. Sec-tion 6 briefly introduces Artificial Neural Network.The experiment and results are shown in Section 7.Discussion and Conclusion are provided in Section 8.

2. DESIGN AND CONSTRUCTION OFMULTI-CHANNEL EMG

EMG measurement is accomplished by the instru-ment called electromyograph. The system, in general,consists of instrumentation amplifier, notch filter, off-set adjustment, isolator, main amplification, and theCRT display. The instrument amplifier is a front-end,high CMRR differential amplifier which functions topick-up a low amplitude signal submersed in the high-frequency noise. The notch filter gets rid of the 50Hznoise while keeping the EMG signal intact. The off-set adjustment maintains the baseline level especiallyduring the subject’s motion. The function of isolatoris to separate the front-end section from the rear-endsection to protect the possible electrical shock to thepatient. The main amplification conditions the EMGprior to be display with CRT. The complexity of theelectronic circuit becomes realized with the necessityto monitor the multi-channel of EMG. Such compli-cate designs, however, are made possible by the cre-ation of entirely reconfiguration and programmablecomponents the so-called Programmable System onChip Microcontroller (PSOCMicrocontroller).

The designed EMG system is capable of monitor-ing 16 channels of EMG simultaneously. Each chan-nel consists of 2 main parts; (i) EMG signal process-

Fig.3: Multi-channel electromyogram acquisitionsystem

Fig.4: Raw EMG signal on laptop computer

ing unit and PSOC microcontroller. Figure 2 showsthe muscular contraction classification system Fig-ure 3 shows Multi-channel electromyogram acquisi-tion system.

The EMG signal processing units consists of 3 sub-units

(i) Instrumentation Amplifier. This subunit usesthe INA 2128 BUR-BROWN Integrated Circuit. TheIC can achieve a CMRR up to 120 dB and gain up to1000.

(ii) Noise filter. The function of the filter is to

INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 57

get rid of the 10-20 Hz noise which is classified as amotion artifact.

(iii) Amplifier and Offset Adjustment. The objec-tive of this sub-unit is to Amplifier EMG signal andmaintains the appropriate offset voltage prior to in-terface with the PSOC.

The PSOC microcontroller consists of 4 subunits(i) PGA (Programmable Gain Amplification) This

subunit acts as the buffer and the main amplificationof EMG.

(ii) Low pass filter. The function of the filter isto remove of the high frequency noise. The cut-offfrequency is at 500 Hz.

(iii) DELTA-SIGMA. This subunit functions as a8-bit analog to digital converter.

(iv) UART. This subunit functions to perform RS-232 interfacing unit with PC

3. CLASSICAL FOURIER-BASED FEATURES

Fourier transform is a useful technique that rep-resents a stationary signal in terms of a function offrequency by determining the frequency component.The equation is represented as:

X (f) =N−1∑n=0

x (n) e(−j2πfn/N) (1)

Where x(n) is a sequence of EMG input and X (f)is complex vector provided by the sinusoidal coeffi-cients. The classical Fourier transform was appliedto the EMG signal. Then, the power spectral densityof the energy at various frequencies was computedfrom square magnitude of the Fourier transform asrepresented in Equation 2. Figure 5 shows the energycontent of EMG signal in form of the power spectrum.

PS (f) = |X (f)|2 (2)

Fig.5: EMG Power spectrum

4. EMG FEATURE EXTRACTION ANDMAPPING

Fig.6: Feature extraction and mapping

Figure 6 shows the feature extraction and map-ping process. Each of the 16 EMG channels will beconverted to frequency domain by taking the fouriertransform. The energy content of the EMG signalis then evaluated by computing area under the mag-nitude squared of the fourier transform. The energycontent on the 4x4 grid corresponding to the 4x4 elec-trode shown in figure 7 is used for artificial neuralnetwork classification. The 4x4 grid data was inter-polated to derive the 49x49 topological maps whichare later applied to ICA for muscular contraction clas-sification. Figure 8 shows the topographical mappingof various muscular contractions.

Fig.7: 16 Channel electrode placements

5. FROM PRINCIPAL COMPONENT ANAL-YSIS TO INDEPENDENT COMPONENTANALYSIS

Principal Component Analysis (PCA) is a statis-tical technique which used to describe a large dimen-sional space with a relative small set of vectors. Itis a popular technique for finding patterns in data ofhigh dimension, and is used commonly in both facerecognition and image compression. [13] Applicationof PCA to face recognition is known as Eigen face.The Eigen face technique is a powerful yet simple so-lution to the face recognition dilemma. It uses muchmore information by classifying faces based on gen-eral facial patterns. Here we focus on the applicationof PCA for muscular-contraction classification

The procedure for using PCA is divided into 2steps. (i) Training step and (ii) Classification step.

58 D. Sueaseenak et al: Muscular-Contraction Classification : Comparison Study ... (55-62)

Fig.8: Topological mapping

The Training step is as follow:(i) Convert each cropped topological mapping ma-

trix into a vector Ti of length N (N= map width*mapheight). For the M data set, we let the training setrepresented by, T1, T2, T3, ..., TM where M is thevector of N2

(ii) Compute the mean vector Ψ and the set ofdeviation from the mean vector Φ = [Φi,Φ2, ...ΦM ]which is defined as

Ψ =1M

M∑i=1

Ti (3)

Φi = Ti −Ψ (4)

(iii) Compute the covariance matrix C which isdefined as

C =1M

ΦΦT (5)

(iv) Compute the Eigenvalue and Eigenvector of Crepresented as

Cνi = µiνi (6)

where µi is the corresponding Eigenvalue of Eigenvector νi

(v) Project each training set on the Eigenspaceusing the operation

Ω = V · Φ (7)

Where V is the Eigen matrix where each row isthe Eigenvector νi can be written as

Ω = [ω1, ω2, . . . , ωM ] (8)

Where ωi is the coefficient of the training map ith

The Classification step is as follow:Project vector form of the tested topological map-

ping matrix Tp to the Eigenspace using equation (5)to derive ωs as

ωs = V · [Tn − ψ]

The tested topological mapping matrix is classifiedto class k which minimized

ε2k = ‖ωs − ωk‖2 (9)

with 1 ≤ k ≤MThe goal of independent component analysis (ICA)

is to minimize the statistical dependence between thebasic vectors. Mathematically, we can write

WX = U (10)

ICA searches for a linear transformation W thatminimizes the statistical dependence between eachrow of U. There exists a number of iterative algorithmto solve for W [16,17]. Most of them are optimizedfor the dependence criteria including Kurtosis, Ne-gentropy, etc.[18]. In this paper, we applied the wellknown ICA algorithm the so-called InfoMax purposedby Bell and Sejnowski [19]. The idea of InfoMax hasbeen applied to Eigenvector of PCA by Barlett et. al.[20] by minimize the statistical dependence betweeneach row of U in

WV = U (11)

where V as an Eigen Basis matrix where each rowis the Eigen vector νi defined in (4). The new basisW−1U is then used in place of V. The Projectionof each training set on the new basis -space is hencedefined as

ωs =(W−1U

)· [Tn − ψ] (12)

INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 59

6. ARTIFICIAL NEURAL NETWORK

Two-layered artificial neural network using back-propagation training protocol was used as a classifier.The 4x4 EMG grid data served as the input of theneural network. The eight outputs of neural networkthat correspond to the eight classes of the muscularcontraction. Structure of artificial neural network isshown in figure 9.

Fig.9: Artificial neural network used in the classifi-cation

7. EXPERIMENT AND RESULTS

The 15 patterns of topographical mapping of eightmuscular contraction of forearm (120 maps) wereused in the training process of ICA and used 4x4grid (120 data) for training process of ANN. The to-pographical mapping of the 15 unknown contractionswas then used as the ICA tested set. The 15 unknown4x4 grid data was then used as ANN testd set. Figure8 shows the ICA training sets, the derived ICA basisand the result of ICA classification. Table 1 showsthe accuracy of ICA and ANN classification.

8. DISCUSSION AND CONCLUSION

A multi-channel electromyogram acquisition sys-tem using PSOC microcontroller was designed andconstructed to aquire multi-channel EMG signals.The16 EMG channels was 4x4 grid data for classificationby using artificial neural network. The 4x4 grid datawas performed to a topological map of EMG signal

Table 1: ACCURACY OF ICA AND ANN CLAS-SIFICATION

on the muscle surface. The mapping for various pat-tern of muscular contraction were then recorded andlater analyzed with independent component analysisto classify the pattern of muscular contraction pat-tern. The comparison study of classification resultdemonstrates that ANN provides the comparable per-formance as the ICA. Yet the ANN computationaltime is noticeably less than that of the PCA.

9. ACKNOWLEDGMENT

The authors wish to thanks ASEAN UniversityNetwork/Southeast Asia Engineering Education De-velopment Network (AUN/SEED-Net) to supportscholarship and The DEMAMEDICAL CO., LTD tosupport the ECG/EMG surface electrode and leadwire to measure EMG signal in this research.

References

[1] G. A. Bekey, C Chang, J Perry, M.M. Hof-fer,“Pattern recognition of multiple EMG signalsapplied to the description of human gait,” Pro-ceedings of IEEE, vol. 65, pp. 674-689, 1977.

[2] S. Boisset, F Goubel, “Integrated electromygra-phy activity and muscle work,” J Applied Phys-iol, vol 35, pp. 695-702, 1972.

[3] C.J. DeLuca, “Use of the surface EMG signal forperformance evaluation of back muscle,” Muscle& Nerve, vol. 16, pp. 210-216, 1993.

[4] R. Plonsey, “The active fiber in a volume con-ductor,” IEEE Trans Biomed Eng, vol. 21, pp.371-381, 1974.

[5] D.A. Winter, “Pathologic gait diagnosis withcomputer averaged electromyographic profiles,”Arch Phys Med Rehabil, vol. 65, pp. 393-398,1984.

[6] K. Lyons, J Perry, J.K. Gronley, L Bbarnes, D.Antonelli, “Timing and relative intensity of hipextensor and abductor muscle action during level

60 D. Sueaseenak et al: Muscular-Contraction Classification : Comparison Study ... (55-62)

(a)

(b)

Fig.10: (a) Training Topological Mapping Input ofICA; (b) ICA Basis

and stair ambulation: An EMG study,” PhysicalTherapy, vol. 63, pp. 1597-1605, 1983.

[7] Elaine N.Marieb., Human anatomy and physiol-ogy, Sixth Edition, The Pearson Education; pp.296,2004

[8] K.S. Turker, T.S Miles, “Cross talk from othermuscles can contaminate electromyographic sig-nals in reflex studies of the human leg,” Neuro-science Letters, vol. 111, pp. 164-169, 1990.

[9] J.W. Morrenhof, H.J. Abbink, “Cross-correlation and cross talk in surface elec-tromyography,” Electromyography, vol. 25, pp.73-79, 1985.

[10] Basmajian JV, de Luca CJ. Muscles Alive - The

(a) Hand close (b) Hand open

(c) Wrist extension (d) Wrist flexion

(e) Wrist extension (f) Wrist flexion

(g) Radial flexion (h) Ulnar flexion

Fig.11: Classification results

Functions Revealed by Electromyography. TheWilliams & Wilkins Company; Baltimore, 1985

[11] J. Cartinhour, “ A Bayes classifier when the classdistributions come from a common multivariatenormal distribution ” ,IEEE Transactions on Re-liability, vol. 41, Issue 2, pp. 124 - 126, 1992.

[12] N.B. Karayiannis, M.M. Randolph-Gips, “ Softlearning vector quantization and clustering algo-rithms based on non-Euclidean norms: single-norm algorithms ” ,IEEE Transactions on Neu-ral Networks, vol. 16, Issue 2, pp. 423 - 435, 2005.

[13] E. Alpaydin, M.I. Jordan, “ Local linear percep-trons for classification ” ,IEEE Transactions onNeural Networks, vol. 7, Issue 3, pp. 788 - 794,1996.

[14] Zhujie, Y.L. Yu, “ Face recognition with Eigenfaces ” ,Proceedings of the IEEE International

INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 61

Conference on Industrial Technology, pp. 434 -438, 1994.

[15] Comon, P., Independent component analysis; Anew concept? Signal Processing, vol 36, no. 3,pp. 287-314,1994

[16] Cardoso, J.-F., Infomax and Maximum Likeli-hood for Source Separation. IEEE Letters onSignal Processing, vol. 4. pp. 112-114, 1997

[17] Hyvarinen, A., The Fixed-point Algorithm andMaximum Likelihood Estimation for Indepen-dent Component Analysis. Neural ProcessingLetters, vol. 10: pp. 1-5, 1999.

[18] Hyvarinen, A. and E. Oja, Independent Com-ponent Analysis: Algorithms and Applications.Neural Networks, 2000. vol. 13, no. 4-5, pp. 411-430, 2000.

[19] Bell, A.J. and T.J. Sejnowski, An information-maximization Approach to Blind Separation andBlind Deconvolution. Neural Computation, vol.7, no. 6, pp. 1129-1159, 1995

[20] Bartlett, M.S., H.M. Lades, and T.J. Sejnowski.Independent component representations for facerecognition. in SPIE Symposium on ElectronicImaging: Science and Technology; Conferenceon Human Vision and Electronic Imaging III,San Jose, CA, 1998

D. Sueaseenak received the B.Eng.(Electrical Engineering) from Srinakhar-inwirot University, Thailand, in 2005,and the M.Eng. (Biomedical Electron-ics) from King Mongkut’s Institute ofTechnology Ladkrabang, Thailand, in2007. He is currently a Ph.D. studentin Department of Electronics, Faculty ofEngineering, KMITL.

T. Chanwimalueang received theB.Eng. (Electrical Engineering) fromKhon Kaen University, Thailand, in2000, and the M.Eng. (BiomedicalElectronics) from King Mongkut’s Insti-tute of Technology Ladkrabang, Thai-land, in 2007. He is currently a Lec-turer in Biomedical Engineering pro-gramme, Faculty of Engineering, Sri-nakharinwirot University.

W. Iampa received the B.Sc. (Radi-ological Technology) and M.Sc. (Ra-diological Technology) from MahidolUniversity, Thailand in 2005 and2008, respectively. She is currentlya Ph.D. student in Department ofElectronics, Faculty of Engineering,KMITL.

M. Sangworasil was born in Bangkok,Thailand in 1951. He received the Bach-elor of engineering and Master o Engi-neering from King Mongkut’s Instituteof Technology at Ladkrabang, Bangkok,Thailand in 1973 and 1977 respectively,and the D. Eng (Electronics) from TokaiUniversity, Japan, in 1990. Followinghis graduate studies, he worked almost28 years at Electronic De- partment,Faculty of Engineering, King Mongkut’s

Institute of Technology at Ladkrabang, Bangkok where he iscurrently an associate professor. His research interest are in thearea of image process with emphasis on Imag Reconstruction,3D modeling, Image Classication and Image Filtering.