Comparison Study of Muscular-ContractionClassification Between Independent Component

Analysis and Artificial Neural NetworkDirek Sueaseenak*, Sunu Wibirama*, Theerasak Chanwimalueangt, Chuchart Pintavirooj * and Manus Sangworasil

Research Center for Communication and Information Technology (ReCCIT), andDepartment of Electronics, Faculty of Engineering,

King Mongkut's Institute of Technology Ladkrabang, Bangkok, ThailandtFaculty of Engineering, Srinakharinwirot University, Nakhon-Nayok, Thailand

kpchuchagkmitlac.th

Abstract- We developed a multi-channel electromyogram clinical interpretation of electrical activities through theacquisition system using PSOC microcontroller to acquire multi- mapping of these signals on the muscle surface [9]. In thischannel EMG signals. An array of 4 x 4 surface electrodes was paper, a PSOC-based multi-channel surface electrode arrayused to record the EMG signal. The obtained signals were data acquisition system is developed to acquire EMG data. Theclassified by a back-propagation-type artificial neural network. EMG signals are then mapped using B-spline interpolationB-spline interpolation technique has been utilized to map the technique. The EMG topological Mapping is then used forEMG signal on the muscle surface. The topological mapping of classification of muscular contraction. There exist a number ofthe EMG is then analyzed to classify the pattern of muscley ~~~~~2D-pattemn classifications [9-12]. In this research, wecontraction using independent component analysis. The proposed compared the EMG-contraction classification techniquesystem was successfully demonstrated to record EMG data and between independent component analysis (ICA) and artifialits surface mapping. The comparison study of muscular neural network (ANN) [14]. The results show that the ANNcontraction classification using independent component analysis neuration (ANN)

as

]. perre s that theand artificial neural network demonstrates shows that classification yields as good performance as the ICA in theperformance of ANN classification is as comparable as that of faster computational time.the ICA. The computational time of ANN is also less than that of The paper is organized as follows: Section II is devoted tothe ICA. the design concept multi-channel electromyogram system.

Keywords-EMG, PCA ICA, ANN Section III described feature extraction and topologicalmapping process. Section IV briefly introduced Independent

I. INTRODUCTION component analysis. Section V briefly introduced artificialneural network.The experiment and result is shown in section

Electromyography (EMG) is the study of muscle electrical VIDicsonadCclinispvdeinetonIIsignals. EMG is sometimes referred to as myoelectric activity.Many muscular abnormalities such as muscular dystrophy, II. DESIGN AND CONSTRUCTION OF MULTI-inflammation of muscle, peripheral nerve damages could CHANNEL EMGresults in an abnormal electromyogram [1-6]. EMG can be EMG measurement is accomplished by the instrumentrecorded by two types of electrodes; invasive electrode the so- Ed metromyp h stem in the cnsts of

called wire or needle electrodes and non-invasive electrode the called electromyograph The system, in general, consists ofso-called surface electrode. Wire or needle electrodes records instrumentation amplifier, notch filter, offset adjustment,individual muscle fiber action potentials which is an ideal isolator, main amplification, and the CRT display. Thechoice to evaluate the muscle activity [9]. However, fine wire instrument amplifier is a front-end, high CMRR differentialintramuscular electrodes require a needle for insertion into the amplifier which functions to pick-up a low amflitude signalmuscle and may cause a significant pain. The choice of surface submersed in the high-frequency noise. The notch filter gets ridelectrode is then preferable. However, when EMG is acquired of the 50Hz noise while keeping the EMG signal intact. Thefrom surface electrodes mounted directly on the skin, the signal offset adjustment maintains the baseline level especially duringis a composite of all the muscle fiber action potentials the subjects motion. The function of isolator is to separate theoccurring in the muscles underlying the skin. Estimating this front-end section from the rear-end section to protect theforce in general is a hard problem due to difficulties in possible electrical shock tobhepatient. The main amplificationactivating a single muscle in isolation, isolating the signal conditions the EMG prior to be display with CRT. Thegenerated by a muscle from that of its neighbors, and other complexity of the electronic circuit becomes realized with theassociated problems [7-8]. The clinical application of EMG necessity to monitor the multi-channel of EMG. Suchcan be classified into two mains categories. (i) Standard EMG complicate designs, however, are made possible by the[8] is recorded from discrete sites on a muscle and thus creation of entirely reconfiguration and programmableprovides only a limited picture of the actual muscular electrical components the so-called Programmable System On Chipactivity in the vicinity of the recording electrode. (ii) Array Microcontroller (PSOC Microcontroller).EMG recorded by an array electrode which facilitates the

978-1-4244-2336-1/08/$25.OO © 2008 IEEE

468

notB1ion D:EMG signal processing EMG spectrum analysis EMG mapping Patern classificaion

Hnd close Hand open FFT Spliine hidepenideitJ| interpolationi component

Wrist Wrist aayextension flexion 44Eetoe)(IC A)E 4x4 Electrode 0 a

Radial Ulnarflexion flexion data= Aitficial

4x4affay neu~~~~~~~~~~~~~~~~~~~~~~~~~~~9x9rialWrtist Vltist nntixietwoik-

pronation supination c MCU ta

Fig. 1 Muscular contraction classification system

The designed EMG system is capable of monitoring 16channels of EMG simultaneously. Each channel consists of 2main parts; (i) EMG signal processing unit and PSOCmicrocontroller. Figure I shows the muscular contraction _classification system Figure 2 shows Multi-channel -_E____electromyogram acquisition system. I_____X_

The EMG signal processing units consists of 3 sub-units

(i) Instrumentation Amplifier. This subunit uses the INA l_______2128 BUR-BROWN Integrated Circuit. The IC can achieve aCMRR up to 120 dB and gain up to 1000.

(ii) Noise filter. The function of the filter is to get rid of the10-20 Hz noise which is classified as a motion artifact. Fig. 3 Raw EMG signal on laptop computer

(iii) Amplifier and Offset Adjustment. The objective of thissub-unit is to Amplifier EMG signal and maintains the III. EMG FEATURE EXTRACTION AND MAPPINGappropriate offset voltage prior to interface with the PSOC.

The PSOC microcontroller consists of 4 subunits Multi-channel EMG signal Fast fourier(i) PGA (Programmable Gain Amplification) This subunit EMG data condition transform

acts as the buffer and the main amplification ofEMG.

(ii) Low pass filter. The function of the filter is to remove Neural network | 4x4 dataof the high frequency noise. The cut-off frequency is at 500 Hz. (ANN)

(iii) DELTA-SIGMA. This subunit functions as a 8-bit Classification 49x49 data Splineanalog to digital converter. (ICA) interpolation

(iv) UART. This subunit functions to perform RS-232 Fig. 4 Feature extraction and mappinginterfacing unit with PC.

Figure 4 shows the feature extraction and mapping process.Each of the 16 EMG channels will be converted to frequencydomain by taking the fourier transform. The energy content ofthe EMG signal is then evaluated by computing under the

magnitude squared of the fourier transform. The energy contenton the 4x4 grid corresponding to the 4x4 electrode shown infigure 5 is used for artificial neural network classification. The4x4 grid data was interpolated to derive the 49x49 topologicalmaps which are later appliedtoICWA for muscular contraction

X.-111.__ ~~~~~~~~classification. Figure 6 shows the topographical mapping Of5ti lll ~~~~~~~~~~variousmuscular contractions.

Fig. 2 Multi-channel electromyogram acquisition system

2008 International Symposium on Communications and Information Technologies (ISCIT 2008) 469

solution to the face recognition dilemma. It uses much moreS l .,= lill1llll-l1lll 1 1 1 1ll linformation by classifying faces based on general facial

patterns. Here we focus on the application of PCA formuscular-contraction classification

The procedure for using PCA is divided into 2 steps. (i)

Fig. 5 16 Channel electrode placements Training step and (ii) Classification step.The Training step is as follow:

(i) Convert each cropped topological mapping matrixinto a vector T1 of length N (N= map width*map height). For

from the mean vector $2 [$l':Dx2,..-:DM] which iSWs1; exension Wtst flexion defined as

1

the datse ,wel Tjth trinn set rpentedby

(iii) Compute the covariance matrix C which is defined as

Wst su prton Wnist pronaion 1 )4 (3)

(iv) Compute theE igenvalue and Eigenvector of Cwhst eAoionhstflexion drepresented as

CV i=p i (4)

Xwhere ,u; isthe correspondingEig envalue of EigenHad closed Hand open vector v

I(v) Project each training set on the Eigenspace using the_ ~~~~~~~~~~~~~~operation OQ _< 52 = V * W) (5)

Where V is the Eigen matrix where each row is the= onEigenvector V . can be written as

Radial flexion Ulrv enfSexion (6)

Fig. 6 Topological mappingWhere W). is the coefficient of the training map i1h

IV. FROM PRINCIPAL COMPONENT ANALYSIS TO The Classification step is as follow:INDEPENDENT COMPONENT ANALYSIS Project vector form of the tested topological mapping

Principal Component Analysis (PCA) is a statistical matrix T to the Eigenspace using equation (5) to derive (Otechnique which used to describe a large dimensional spacewith a relative small set of vectors. It is a popular technique for asfinding patterns in data of high dimension, and is used e=V*[T-ucommonly in both face recognition and image compression. V [Tn[13] Application ofPCA to face recognition is known as Eigen Tetse oooia apn arxi lsiidtface. Thne Eigen face technique iS a powerfUll yet simple clskwhhmimze

470 2008 International Symposium on Communications and Information Technologies (ISCIT 2008)

2 2 VI. EXPERIMENT AND RESULTSck -Cos cok (7) The 15 patterns of topographical mapping of eight muscular

contraction of forearm (120 maps) were used in the trainingwith 1 < k <M process of ICA and used 4x4 grid (120 data) for training

The goal of independent component analysis (ICA) is to process of ANN. The topographical mapping of the 15minimize the sascunknown contractions was then used as the ICA tested set. The

Minimizeathemsticatiscal dritepnecbtenh aivtos 15 unknown 4x4 grid data was then used as ANN testd set.Figure 8 shows the ICA training sets, the derived ICA basis and

WX = U (8) the result of ICA classification. Table 1 shows the accuracy ofICA and ANN classification.

ICA searches for a linear transformation W that minimizesthe statistical dependence between each row of U. There exists TABLE I. ACCURACY OF ICA AND ANN CLASSIFICATIONa number of iterative algorithm to solve for W [15,16]. Most of Muscular ICA ANNthem are optimized for the dependence criteria includingKurtosis, Negentropy, etc.[17]. In this paper, we applied the Contraction correct 0 correct 00well known ICA algorithm the so-called InfoMax purposed by Wrist extension 15 100 15 100Bell and Sejnowski [18]. The idea of InfoMax has been applied Wrist flexion 5 33 13 86.7to Eigenvector of PCA by Barlett et. al. [19] by minimize thestatistical dependence between each row ofU in Wrist supination 14 93.3 15 100

WV = U (9) Wrist pronation 8 53.3 15 100

Radial flexion 14 93.3 12 80where V as an Eigen Basis matrix where each row is the

-I Ulnar flexion 14 93.3 13 86.7Eigen vector vi defined in (4). The new basis w u is then

Hand close 13 86.7 12 80used in place of V. The Projection of each training set on thenew basis -space is hence defined as Handopen 15 100 15 100

co (W-1u) [Tn (Il0) Total 112 93.3 115 95.8

V. ARTIFICIAL NEURAL NETWORK VII. DISCUSSION AND CONCLUSIONTwo-layered artificial neural network using back- A multi-channel electromyogram acquisition system using

propagation training protocol was used as a classifier. The 4x4 PSOC microcontroller was designed and constructed to aquiregrid data served as the input of the neural network. The eight multi-channel EMG signals.The 16 EMG channels was 4x4outputs of neural network that correspond to the eight classes grid data for classification by using artificial neural network.of the muscular contraction. Structure of artificial neural The 4x4 grid data was performed to a topological map of EMGnetwork is shown in figure 7. signal on the muscle surface. The mapping for various pattern

of muscular contraction were then recorded and later analyzedIiillt Hk1deHiddH11ti with independent component analysis to classify the pattern of

muscular contraction pattern. The comparison study ofNo g .classification result demonstrates that ANN provides the

comparable performance as the ICA. Yet the ANNcomputational time is noticeably less than that of the ICA.

Fi.rtfca nera networusdi th clssifcaio

Fig.7Art 2008aInteralnatiorkuena SympolasiumcatonCmuiain n nomto ehoois(SI 08 7

ACKNOWLEDGMENT

The authors wish to thank DEMAMEDICAL CO., LTD ford! *1 41 w I st ilt t support the ECG/EMG surface electrode and lead wire for

t 4- _ _ __Z1!<-< < measuring EMG signal in this research.

REFERENCES. . [1] G. A. Bekey, C Chang, J Perry, M.M. Hoffer,"Pattern recognition of

multiple EMG signals applied to the description of human gait,"Proceedings of IEEE, vol. 65, pp. 674-689, 1977.

[2] S. Boisset, F Goubel, "Integrated electromygraphy activity and muscle* w wH -_ < < _ _ X _ work," J Applied Physiol, vol 35, pp. 695-702, 1972.

[3] C.J. DeLuca, "Use of the surface EMG signal for performancet _' t a

X 1 _ _ _ _j _ _evaluation of back muscle," Muscle & Nerve, vol. 16, pp. 210-216,T ,P T t K, t t K _ =1993.~ I ~ jj * [4] R. Plonsey, "The active fiber in a volume conductor," IEEE Trans

Biomed Eng, vol. 21, pp. 371-381, 1974.[5] D.A. Winter, "Pathologic gait diagnosis with computer averaged

electromyographic profiles," Arch Phys Med Rehabil, vol. 65, pp. 393-(a) 398, 1984.

[6] K. Lyons, J Perry, J.K. Gronley, L Bbarnes, D. Antonelli, "Timing and* ~~~~~~~ ~~~~~ ~~relative intensity of hip extensor and abductor muscle action during level

- . |- E ;: =_ E = 0 *--and stair ambulation: An EMG study," Physical Therapy, vol. 63, pp.1597-1605, 1983.

[7] K.S. Turker, T.S Miles, "Cross talk from other muscles can contaminateelectromyographic signals in reflex studies of the human leg,"Neuroscience Letters, vol. 11I1, pp. 164-169, 1990.

[8] J.W. Morrenhof, H.J. Abbink, "Cross-correlation and cross talk insurface electromyography," Electromyography, vol. 25, pp. 73-79, 1985.

Is in is Is 11 1in FM in 9 is Is[9] Basmajian JV, de Luca CJ. Muscles Alive - The Functions Revealed byElectromyography. The Williams & Wilkins Company; Baltimore, 1985

*toin n 11 Is isIn 11 IN Is FM [10] J. Cartinhour, " A Bayes classifier when the class distributions comefrom a common multivariate normal distribution " ,IEEE Transactionson Reliability, vol. 41, Issue 2, pp. 124 -126, 1992.

[11] N.B. Karayiannis, M.M. Randolph-Gips, " Soft learning vectorquantization and clustering algorithms based on non-Euclidean norms:single-norm algorithms " JIEEE Transactions on Neural Networks, vol.

3-= _; _.N. 0 f, 201 tt - lt 16, Issue 2, pp. 423 -435,2005.[12] E. Alpaydin, M.I. Jordan, " Local linear perceptrons for classification"

J,EEE Transactions on Neural Networks, vol. 7, Issue 3, pp. 788 - 794,1996.

(b) [13] Zhujie, Y.L. Yu, " Face recognition with Eigen faces " ,Proceedings ofthe IEEE International Conference on Industrial Technology, pp. 434 -

Anonumobus input of Hand Close no: 2 438, 1994.

[II [14] Comon, P., Independent component analysis; A new concept? SignalProcessing, vol 36, no. 3, pp. 287-314,1994

30 [15] Cardoso, J.-F., Infomax and Maximum Likelihood for SourceSeparation. IEEE Letters on Signal Processing, vol. 4. pp. 112-114, 199740

[16] Hyvarinen, A., The Fixed-point Algorithm and Maximum Likelihood5 10 1 20 25 30 35 40 45 Estimation for Independent Component Analysis. Neural Processing

EMG idenification is Hand Close no:13 Letters, vol. 10: pp. 1-5, 1999.[17] Hyvarinen, A. and E. Oja, Independent Component Analysis:

10 Algorithms and Applications. Neural Networks, 2000. vol. 13, no. 4-5,20 pp. 411-430, 2000.30 [18] Bell, A.J. and T.J. Sejnowski, An informationmaximization Approach to400 Blind Separation and Blind Deconvolution. Neural Computation, vol. 7,

no. 6, pp. 1129-1159, 19955 10 15 21] 25 30 35 40 45 [19] Bartlett, M.S., H.M. Lades, and T.J. Sejnowski. Independent component

representations for face recognition. in SPIE Symposium on ElectronicImaging: Science and Technology; Conference on Human Vision andElectronic Imaging III, San Jose, CA, 1998

Fig. 8 (a) Training Topological Mapping Input of ICA;

(b) ICA Basis; (c) Result of classification

472 2008 International Symposium on Communications and Information Technologies (ISCIT 2008)