mring

G) sactsyss ofify ss rnal

classify four different arm movement signals. Prior to classication, proper feature vectors are derived

1. Introduction

takenpotentlt of diignal isetecteactivi

ing of feature extraction and classication, have been applied (Par-ker, Englehart, & Hudgins 2004). This concept has been used for thedevelopment of myoelectric prosthesis control systems obtainedby classication of EMG signals (Choi & Kim, 2007; Englehart, Hud-gins, & Parker, 2001; Hu & Nenov, 2004; Kumar et al., 2001; Lucas,Gaufriau, Pascual, Doncarli, & Farina, 2008; Parker & Scott, 1986;

The success of a myoelectric control scheme depends largely onthe classication accuracy. Englehart et al. (2001), proposed a no-vel approach that demonstrates greater accuracy than their previ-ous work. They used wavelet-based feature set, reduced indimension by principal components analysis. Further, it is exposedthat four channels of myoelectric data increases the classicationaccuracy, as compared to one or two channels. They claimed thata robust online classier is constructed, which produces class deci-sions on a continuous stream of data.

Choi and Kim (2007) investigated to design an assistive real-time system for the upper limb disabled to access a computer via

Corresponding author.E-mail addresses: aalkan@ksu.edu.tr (A. Alkan), mucahidgunay@gmail.com

Expert Systems with Applications 39 (2012) 4447

Contents lists availab

Expert Systems w

.e(M. Gnay).during contraction or relaxation.Each movement of muscles corresponds to a specic pattern of

activation of several muscle bres; therefore multi-channel EMGrecordings can be used to identify the movement. Due to the com-plex nature of the signal, detailed analysis and classication is of-ten difcult, especially if the EMG relates to movement (Kumar,Ma, & Burton, 2001).

For this purpose, different pattern recognition schemes, consist-

imum classication error, as estimated from the learning signal set.Then themethodwas applied to the classication of six handmove-ments with recording of the surface EMG from eight locations overthe forearm. For all subjects using the eight channels they reportedamisclassication rate as (mean S.D.) 4.7 3.7%with theproposedapproach while it was 11.1 10.0% without proposed technique.They stated that DWT and SVM can be implemented with fast algo-rithms and their method is suitable for real-time applications.Bioelectrical signals are usuallyproduced by the sum of electricalspecialized tissue or organ as a resuevents happening in the body. EMG sbioelectrical signals which can be dand are generated by the electrical0957-4174/$ - see front matter 2011 Elsevier Ltd. Adoi:10.1016/j.eswa.2011.06.043from the signal. The feature vectors are generated by using mean absolute value (MAV). These featurevectors are provided as inputs to the identication/classication system. Discriminant analysis using vedifferent approaches, classication accuracy rates achieved from very good (98%) to poor (96%) by using10-fold cross validation. SVM classier gives a very good average accuracy rate (99%) for four movementswith the classication error rate 1%. Correct classication rates of the applied techniques are very highwhich can be used to classify EMG signals for prosperous arm prosthesis control studies.

2011 Elsevier Ltd. All rights reserved.

to be electric currentsial differences across afferent electrochemicalone of the best-known

d over the skin surfacety of the muscle bres

Parker et al., 2004; Tscharner, 2000; Wojtczak, Amaral, Dias, Wol-czowski, & Kurzynski 2009.

As in Lucas et al. (2008), the discrete wavelet transforms (DWT)based representation space is used for supervised classication ofmulti-channel surface electromyography signals with the aim ofcontrolling myoelectric prostheses. They applied a support vectormachine (SVM) approach to classify a multichannel representationspace. They optimized themotherwaveletwith the criterion ofmin-terns properly. Discriminant analysis and support vector machine (SVM) classier have been used toIdentication of EMG signals using discri

Ahmet Alkan , Mcahid GnayKahramanmaras Sutcu Imam University, Department of Electrical & Electronics Enginee

a r t i c l e i n f o

Keywords:EMGDiscriminant analysisCross validationSVM

a b s t r a c t

The electromyography (EMtary or involuntary musclecontrolled by the nervouscal/physiological propertienique is proposed to classEMG signals. This work useferent movements. Each sig

journal homepage: wwwll rights reserved.inant analysis and SVM classier

, Kahramanmaras, Turkey

ignal is a bioelectrical signal variation, generated in muscles during volun-ivities. The muscle activities such as contraction or relaxation are alwaystem. The EMG signal is a complicated biomedical signal due to anatomi-the muscles and its noisy environment. In this paper, a classication tech-ignals required for a prosperous arm prosthesis control by using surfaceecorded EMG signals generated by biceps and triceps muscles for four dif-has one single pattern and it is essential to separate and classify these pat-

le at ScienceDirect

ith Applications

lsevier .com/locate /eswa

corded and ltered using signal statistics such as mean and vari-

dures were developed to calculate intensity. It is reported that the

ual patterns are used to elaborate the features of the signals(Englehart & Hudgins, 2003). Then these patterns are windowedand average values of the each window are used as the feature val-ues. Using the variable m to represent window index, xk to repre-sent the EMG data point at time k, and L to representclassication window length the estimate of mean absolute valueis given by:

Xm 1LXLk1

jxkj 1

Window length is the single parameter affecting the output of pre-processing method (MAV). The output of MAV has inuenced indi-rectly upon the success of classication because of being applied asan input to the classiers. That is why, it is so important to denethe window length. Different window lengths are tried and regard-ing to the classication accuracy, it is decided to use 32 samples asan optimum window length for the data. Thus, 512 points are re-duced to 16 points by calculating absolute value and then takingthe average of each of the windows. So, for each pattern, insteadof 512 values 16 values can be used for further processing system.The EMG classication/analysis steps can be seen in Fig. 2. Also, asample pre-processed EMG signal from four arm movements areshown in Fig. 3.

2.2.2. Discriminant analysisDiscriminant analysis is a well-known statistical classication

s wmethod resolves events within the EMG signal.Hu and Nenov (2004), compared the performance of two feature

extraction methods for multichannel EMG based arm movementclassication. Theyuseda scalar autoregressivemodel (sAR) for eachchannel and amultivariate ARmodel (mAR)whichmodels all chan-nels as a whole. Leave-one-out cross-validation was adopted forevaluating the classication performance using a parametric statis-tical classier. Theyprocessed a total of 216EMGsegments obtainedfrom repeated 18 performances by three normal subjects. It is re-ported that mAR model based feature set achieved a better classi-cation accuracy than sAR did for each conguration.

Articial neural networks are used to classify EMG signals to con-trolmultifunction prosthesis. Fingermotions discrimination is takenas the key problem inWojtczak et al. (2009). The EMG signal classi-cation system was established using the linear neural network. It isreported that the experimental results show a promising perfor-mance in classication of motions based on bio-signal patterns.

Multi-channel surface electromyography (SEMG) providesinformation on motion detection of exion and extension of n-gers, wrist, forearm, and arm. A portable hand motion classier(HMC) is developed to identify hand motion from the SEMG signalswith an electrode conguration system (ECS) and recognition usinggrey relational analysis (GRA) based classier in Dua, Hung, Shyu,and Chen (2010). They reported that the ECS consists of seven ac-tive electrodes place around the forearm to acquire the multi-channel SEMG signals of corresponding muscle groups and theGRA-based classier could be further recommend to implementin prosthesis control, robotic manipulator or hand motion classi-cation applications.

2. Materials and methods

2.1. Material

In this study, four upper arm movements were designed. Theyincluded elbow exion (EF), elbow extension (EE), forearm prona-tion (FP) and forearm supination (FS). For each movement there are100 patterns and each of the patterns has 512 points. First 256points obtained from biceps and the last 256 points from tricepsmuscles. Sampling frequency of the acquisition system is1000 Hz (Englehart & Hudgins, 2003; Kocyigit & Korurek, 2005).A sample EMG pattern can be seen in Fig. 1.

2.2. Methodsance. In order to control movement and clicking of a cursor fromthe obtained signals, they classied six patterns, applying a super-vised multi-layer neural network trained by a back propagationalgorithm. Also, they developed an on-screen keyboard, making itpossible to enter Roman and Korean letters on the computer. It isreported that using this computer interface, the user can browsethe Internet and read/send e-mail.

Tscharner (2000), developed a time-frequency analysis of theintensities in time series for the analysis of surface myoelectric sig-nals. The author proposed an intensity analysis which uses a lterbank of non-linearly scaled wavelets with specied time-resolu-tion to extract time-frequency representations of the signal. Toapproximate the power of the signal in time domain, certain proce-residual muscle activities without standard computer interfaces.For this idea, EMG signals from muscles in the lower arm were re-

A. Alkan, M. Gnay / Expert System2.2.1. Pre-processingPre-processing or feature extraction is very important issue in

many signal processing applications because of the very complexnatures of the biomedical signals. As in the case of many bioelectri-cal signals surface myoelectric signals are routinely pre-processedusing different techniques to reduce noise and highlight for theanalysis. For this aim, some popular techniques such as time do-main features, spectral analysis, zero crossing and turns counting,root mean square, integral of RMS and wavelet analysis are used(Ma, Kumar, & Pah, 2001).

In this study a time domain feature extraction method, namely,mean absolute value (MAV) is used. Absolute values of the individ-

0 100 200 300 400 500-5

0

5(a) Elbow flextion

0 100 200 300 400 500-5

0

5(b) Elbow extension

0 100 200 300 400 500-5

0

5(c) Forearm pronation

0 100 200 300 400 500-5

0

5(d) Forearm supination

Fig. 1. A sample EMG data from four different arm movements.

ith Applications 39 (2012) 4447 45technique which uses training data to estimate the parameters ofdiscriminant functions of the predictor variables. Discriminantfunctions determine boundaries in predictor space between

s wPre-processing

Classification

EMG

MeanAverageValue (MAV)

(EF, EE, FP, FS)

Discriminant Analysis

SVM

Decisions

Fig. 2. EMG classication steps.

46 A. Alkan, M. Gnay / Expert Systemvarious classes. The resulting classier discriminates among theclasses (the categorical levels of the response) based on the predic-tor data. The details of the discriminant analysis classier can befound in Cao and Sanders (1996).

The feature vectors are provided to the discriminant analysisclassier which is so simple to implement and much faster to train.The performances of the following ve different types of Discrim-inant analysis classiers have been investigated in the identica-tion of surface myoelectric records to identify the movement.

Linear discriminant function ts a multivariate normal densityto each group, with a pooled estimate of the covariance matrix.Diagonal linear discriminant function is similar to linear discrimi-nant function except the estimate of covariance matrix being diag-onal, not pooled. This diagonal covariance matrix is estimated bytaking only the diagonal of the estimated sample (pooled) covari-ance matrix, and ignoring the rest. Quadratic discriminant functionts multivariate normal densities with covariance estimates strat-ied by group. Diagonal quadratic discriminant function is similarto quadratic discriminant function except the estimate of covari-ance matrix being diagonal. Mahalanobis discriminant functionuses Mahalanobis distances with stratied covariance estimates.

0 5 10 150

0.5

1

1.5(a)

Ampl

itude

0 5 10 150

1

2

3

Ampl

itude

(b)

0 5 10 150

0.5

1

1.5

Ampl

itude

(c)

0 5 10 150

1

2

3

Ampl

itude

(d)

Fig. 3. A sample pre-processed EMG signal from four arm movements: (a) Elbowexion. (b) Elbow extension. (c) Forearm pronation. (d) Forearm supination.For all of these classiers, the underlying analysis is based onthe evaluation of discriminant functions; therefore a general name,discriminant analysis, is used for this type of classier design.

2.2.3. Cross validation (CV)CV is a method that divides the data set into two groups as

training and sample for supervised learning. Kfold CV rstly ob-tains k subsets having equal number of members from the data setand members are randomly shared by the subsets. For one classi-cation process, each of the subsets is used as a training set andall of the other subsets as a sample set. So, the resultant predictionaccuracy rate is calculated by the average of k prediction accuracyrates (Bollen & Gu, 2006).

2.2.4. SVM classierVapnik has presented a new computational method called sup-

port vector machine. His theory has been advanced between years19951998 (Qian, Mao, Xiang, & Wang, 2010). SVM takes the inputdata as an n-dimensional feature space. Then an (n 1) dimen-sional hyperplane separates the space into two parts. n-dimen-sional input data xi (i = 1, 2, . . . , l) is labelled as yi = 1 for class 1and as yi = 1 for class 2 by yi matrix. A hyperplane can be denedfor linearly separable data.

f x x x b Xni1xixi b 0 2

Sgn (f(x)) is the decision function. In Eq. (2), x is an n-dimen-sional vector and b is a scalar. These determine the position ofthe hyperplane that completely separates the space has to obeythe limits:

yixi x b 1P 0()f xi xi x bP 1 yi 1f xi xi x b 6 1 yi 1

3

The hyperplane that creates maximum limit is called an optimalhyperplane. In the equation below, ni is the independent variableand C is the error penalty. The minimized solution of the hyper-plane is as followed:

/x; n 1=2x x CXli1

ni

!4

depending on

yixi x bP 1 ni; i 1;2; . . . ; l 5ni measures the distance between the limit and the sample xi on

the other side of the limit. This calculation can be simplied asfollowed:

Va Xli1ai 12

Xli;j1

aiajyiyjKerxi:xj 6

depending on

Xli1

yiai 0; C P aP 0; i 1;2; . . . ; l 7

The function Ker(xi xj) is called as kernel function returns the dotproduct of the feature space maps of the original data points. De-tails of the method can be found in the literature (Wang, Yuan,Liu, Yu, & Li, 2009).

3. Results

ith Applications 39 (2012) 4447Using the classify routine of discriminant analysis; ve differenttypes of discriminant analysis classiers are implemented. Inperformance analysis of different types of classiers one needs to

s with Applications 39 (2012) 4447 47report prediction/classication error and/or accuracy rate. (Theclassication accuracy is 1 classication error rate).

The issue of assessment of prediction error of a classier alsodeserves much attention. The experimental classication error isthe ratio of wrong decisions to the total number of cases studied.The true error rate is statistically dened as the error rate of a clas-sier on an asymptotically large number of new cases that con-verge in the limit to the actual population distribution. Duringtraining, underlying parameters of a classier are adjusted usingthe information contained in the training samples. The predictionaccuracy can initially be evaluated by testing the classier backon the training set and noting the resultant training or re-substitu-tion error. This type of assessment of classier performance, basedon training error, is instrumental during the design phase.

If the training set contains too many outliers or excessive train-ing is done, the generalizability performance of the classier willbe poor. Therefore, while evaluating prediction accuracy of classi-cation methods, it is important not to use the training error only(Asyali, Colak, Demirkaya, & Inan, 2006). In general, the training er-ror rates tend to be biased optimistically, i.e., the true error rate isalmost invariably higher than the training error rate. If there areplenty of training samples available, one can partition the overalltraining set into two sets and use one for training and the otherfor testing. If the classier is designed based on a small trainingset, the generalizability performance of the classier will be pooragain.

After the pre-processing step, features are applied to classica-tion system to classify EMG signals related to four upper armmovements. In this study for discriminant analysis 10-fold CV isused. Table 1 reports the 10-fold CV error rates for the ve differenttypes of discriminant analysis classiers.

The same features are applied to a SVM classier to compare theclassication results of all methods with the same pre-processingtechnique. In the SVM classier, the overall data set is randomly di-vided into two equal subsets, selecting the half of the data fortraining and the other half for testing.

4. Discussion

In this study we have proposed a surface EMG signal classica-

Table 1Classication results: Training (resubstitution) and 10-fold error rates for differenttypes of classiers based on the discriminant analysis and SVM.

Classiers Methods Correction rates (%) Error rates (%)

Discriminantanalysis

Linear 97.75 2.25Diaglinear 97.25 2.75Quadratic 97.75 2.25Diagquadratic 98.00 2.00Mahalanobis 96.00 4.00

SVM 99.00 1.00

A. Alkan, M. Gnay / Expert Systemtion systemwhich uses ve discriminant functions and a SVM clas-sier. The variability of accurate classication of discriminantanalysis classiers is from very good (98%) to poor (96%). If theclassication results are examined, one can see that all ve dis-criminant analysis methods give less than 5% classication errorrates. Between these ve methods, diagonal quadratic discriminantfunction gives the best classication error rate (2%). The error ratesof linear and quadratic discriminant functions are 2.25%. Diagonallinear and Mahalanobis discriminant functions have error rates2.75% and 4% respectively.

A discriminant analysis classier is a graceful method using theprobability distributions. However, the computational complexityis high when the dimension of features becomes large even for aGaussian probability density function. This often reduces the prac-tical application of discriminant analysis classiers. On the otherhand, a SVM classier minimizes the generalization error on thetest set under the structural risk minimization (SRM) principle(Bollen & Gu, 2006).

This can be seen in Table 1 that SVM classier gives a very goodaverage accuracy rate (99%) for four movements with the classi-cation error rate 1%. Since misclassied patterns are generallysame, data set plays important role for classication errors. Correctclassication rates of the applied techniques are very high whichcan be used to classify EMG signals prosperous arm prosthesis con-trol studies.

Acknowledgements

This study is supported by the Scientic Research Project Man-agement Department of Kahramanmaras Sutcu Imam Universitywith the project Identication and Classication of EMG signalsfor Arm Prosthesis, Project number: 2010/5-8 YLS.

This study was presented at the 1st International Symposiumon Computing in Science & Engineering (ISCSE 2010) June, 3-5,2010, in Kusadasi, Aydn/Turkey.

References

Asyali, M. H., Colak, D., Demirkaya, O., & Inan, M. S. (2006). Gene expression proleclassication: A review. Current Bioinformatics, 1(1), 5573.

Bollen, M. H. J., & Gu, I. Y. H. (2006). Signal Processing of power quality disturbances.Springer, pp. 684686.

Cao, J., & Sanders, D. B. (1996). Multivariate discriminant analysis of theelectromyographic interference pattern: statistical approach to discriminationamong controls, myopathies and neuropathies. Medical and BiologicalEngineering and Computing, 34(5), 369374.

Choi, C., & Kim, J. (2007). A Real-time EMG-based Assistive Computer Interface forthe Upper Limb Disabled. In Proceedings of the 2007 IEEE 10th InternationalConference on Rehabilitation Robotics, Noordwijk, The Netherlands.

Dua, Y. C., Hung, C. L., Shyu, L. Y., & Chen, T. (2010). Portable hand motion classierfor multi-channel surface electromyography recognition using grey relationalanalysis. Expert Systems with Applications, 37, 42834291.

Englehart, K., & Hudgins, B. (2003). A robust, real-time control scheme formultifunction myoelectric control. IEEE Transactions on Biomedical Engineering,50, 7.

Englehart, K., Hudgins, B., & Parker, P. A. (2001). A wavelet-based continuousclassication scheme for multifunction myoelectric control. IEEE TransactionsOn Biomedical Engineering, 48.

Hu, X., & Nenov, V. (2004). Multivariate AR modeling of electromyography for theclassication of upper arm movements. Clinical Neurophysiology, 115,12761287.

Kocyigit, Y., & Korurek, M. (2005). EMG signal classication using wavelet transformand fuzzy logic classier. ITU dergisi/d mhendislik, 4, 3.

Kumar, D. K., Ma, N., & Burton, P. (2001). Classication of dynamic multi-channelelectromyography by neural network. Electromyography and ClinicalNeurophysiology, 41(7), 401408.

Lucas, M. F., Gaufriau, A., Pascual, S., Doncarli, C., & Farina, D. (2008). Multi-channelsurface EMG classication using support vector machines and signal-basedwavelet optimization. Biomedical Signal Processing and Control, 3, 169174.

Ma, N., Kumar, D.K., & Pah, N. (2001). Classication of hand direction using multi-channel electromyography by neural network. In Seventh Australian and NewZealand Intelligent Information Systems Conference, Perth, Western Australia.

Parker, P. A., Englehart, K., & Hudgins, B. (2004). The control of upper limbprostheses. In R. Merletti & P. A. Parker (Eds.), Electromyography: Physiologyengineering and non-invasive applications. Wiley-IEEE Press.

Parker, P. A., & Scott, R. N. (1986). Myoelectric control of prostheses. Critical Reviewsin Biomedical Engineering, 13, 283310.

Qian, H., Mao, Y., Xiang, W., & Wang, Z. (2010). Recognition of human activitiesusing SVM multi-class classier. Pattern Recognition Letters, 31, 100111.

Tscharner, V. (2000). Intensity analysis in time-frequency space of surfacemyoelectric signals by wavelets of specied resolution. Journal ofElectromyography and Kinesiology, 10, 433445.

Wang, A., Yuan, W., Liu, J., Yu, Z., & Li, H. (2009). A novel pattern recognitionalgorithm: Combining ART network with SVM to reconstruct a multi-classclassier. Computers and Mathematics with Applications, 57, 19081914.

Wojtczak, P., Amaral, T. G., Dias, O. P., Wolczowski, A., & Kurzynski, M. (2009). Handmovement recognition based on biosignal analysis. Engineering Applications ofArticial Intelligence, 22, 608615.

Identification of EMG signals using discriminant analysis and SVM classifier1 Introduction2 Materials and methods2.1 Material2.2 Methods2.2.1 Pre-processing2.2.2 Discriminant analysis2.2.3 Cross validation (CV)2.2.4 SVM classifier

3 Results4 DiscussionAcknowledgementsReferences