Pairwise Linear Discriminant Analysis of Electromyographic signals

Oscar De Silva , George Mann and Ray Gosine

Faculty of Engineering and Applied Science Memorial University of Newfoundland, Canada. {oscar.desilva,gmann,rgosine}@mun.ca

Abstract Electromyographic (EMG) signals are used as rich information sources for control of intelligent prosthetics. For efcient classication the machine learning algorithms used should allow the nonlinear nature of the multi class problem. For generalized application they should have the analytical ability to systematically tackle the problem in hand. To meet these requirements a pair-wise Linear Discriminant Analysis(LDA) is performed in a systematic manner on EMG signals captured from forehand muscles. A 6 class classication performance from 4 EMG channels are reported along with the ability of the algorithm to scale and visualize complex multidimensional cases. Index Terms EMG, Linear discriminant analysis,Upper limb prosthetics, Pattern Recognition

I. I NTRODUCTION Powered upper limb prosthetics are devices engineered for the class of amputee patients with partial or complete amputation, related to the arm. The devices should meet the requirements of day to day manipulation tasks amputee patients, thus considered a complex engineering problem both in terms of hardware design and control. Electromyography (EMG) pattern recognition based strategies have emerged as the dominant form of control for these devices during the last decade. Pattern recognition based methods used for prosthetic control has enabled sequential execution of multiple operations by the patient. By exploiting the set of EMG patterns distinct to each discrete task, the method is able to continuously produce accurate control signals for the device. The classiers have simplied from powerful Multi Layer Perceptron (MLP) neural nonlinear methods[1] to simpler Linear Discriminant (LD) methods [2] with comparable accuracy and speed for real time implementations. Products based on EMG pattern recognition has already made its way to active lower arm prosthetic market. EMG based control of prosthetics become complex and highly constrained as the amputation is higher in the arm. As the level of amputation becomes higher the control input sources(muscles,nerves) reduces drastically. This issue is dominant in shoulder and forequarter amputees, requiring much cognitive demand from the patient to carry out simple manipulation tasks [3]. As a solution, complex strategies which employ EMG signals together with other information sources are sought to achieve multi functionality, simultaneous operation and reduced overall cognitive burden in control.

In developing such higher dimensional control methods it is necessary to employ algorithms which are scalable to different muscle groups and has the ability to provide analytical advantage over the conventional classiers available. [4] reports the ability of partitioning schemes (Multiple Binary Classications (MBC)) to meet the analytical advantage for EMG controller development. This study develops on the work by introducing modications to the partitioning scheme with pair wised implementation and linear discriminant functions which is observed to perform accurately in the multidimensional nonlinear setting. The study develops a feature projection based classication method for EMG signals to suit the analysis of algorithms in different muscle groups. The multi class classication is performed using a partitioning method which improves the ability to visualize the separability between tasks. A linear discriminant (LD) classier is employed in the reduced partitioned space to preserve the simplicity present in the well established multi class linear discriminant classiers. This method is found to posses features which would serve as a valuable analytical tool for complex EMG controller development. II. BACKGROUND Classiers in the context of EMG signal processing is widely reported for multi functional control of prosthetics. It was rst applied for transient EMG signal classication using invasive electrodes [1] and later developed for steady state EMG signals from surface electrodes [5]. The pattern recognition process can be separated in to three main steps which includes data segmentation, feature extraction/ dimensionality reduction and classication. Table I summarizes the main methods reported in literature, which suggests that Linear Discriminant(LD) classier on Time domain (TD) feature sets have become the popular choice for EMG signal classication. These have comparable accuracy to the much powerful yet resource demanding nonlinear neural classiers with higher order wavelet packet feature sets. Control methods available for upper extremity prosthetics (UEP) can be broadly identied as hybrid methods and surgical methods. The hybrid methods are developed by utilizing other body signals together with the available EMG signals. Establishing proprioception has been a concept studied for

Reference Hudgins et al.[1] Chan et al.[6] Englehart et al.[5] Huang et al.[7] Kuiken et al.[8]

Year 1993 2000 2002 2005 2009

Feature set/Segmentation TD feature set, transient EMG analysis TD feature set, transient EMG analysis WPT feature set with PCA, continuous steady state EMG analysis TD+AR+RMS feature set, continuous steady state EMG analysis TD feature set, continuous steady state EMG analysis

Classier ANNMLP Adaptive FUZZY LD GMM LD

Channels /Classes 2 channel 4 class classier 2 channel 4 class classier 4 channel 6 class classier 4 channel 6 class classier 12 channel 11 class classier

Classier Error 16-2% 10-1.3% 7.9-1.1% 6.7-1.1% 96.3elbow 86.9-wrist

Delay 250ms <300ms <300ms <300ms elbow 220ms, wrist 380ms

TABLE I S UMMARY OF M YOELECTRIC CLASSIFIER PERFORMANCES

the development of these devices [9] [10]. Hybrid methods reported for UEP control include shoulder position sensors [11] [12], foot mounted interfaces [13], cognitive vision [14] and Mechanomyography [16]. Recent development for UEP control is reported by surgical rerouting of the control nerves from brain to targeted muscle areas in the chest. This method (Targeted muscle Reinnervation (TMR)) allows the use of EMG classiers for task discrimination using the reinnervated muscles instead of the actual physical muscle which is no longer available due to amputation [8]. These developments in TMR prosthetics and hybrid control schemes demand the EMG classiers to provide analytical ability and robustness to handle nonlinearity in the higher dimensional setting. Dimensionality reduction of EMG features, for classication is previously reported in both supervised [17] [4] and unsupervised settings [5] . Principal Componant Analysis (PCA) is reported to improve feature representation in wavelet and time domain feature sets [5]. Multi class Linear Discriminant Analysis (LDA) is reported for forearm EMG signal classication, but multi class implementation of the LDA has forced the use of complex neural methods to achieve classication accuracy [17]. Partitioned feature projection and classication methods are reported for transforming multi class problems to series of binary classication problems [4]. The partitioning was implemented in a One against all scheme with simple threshold selection for class boundaries. Improvement over LD classier was not achieved but improved controllability and intuitive visualization suitable for analysis is reported. This research extends the partitioning scheme to a pair wised implementation and replaces the simple thresholding with statistical representations. The visualization capability in the lower dimensional domain and ability to handle nonlinearities of the approach serves as a valuable analytical tool for different EMG classier development. III. M ETHODOLOGY This study proposes pairwised partitioned feature projection approach for EMG classication. Fishers linear discriminant, also termed Linear Discriminant Analyisis (LDA) is used for supervised feature projection in a partitioned feature

space. It is important to note the difference between the Linear Discriminant classier (LD) and the supervised feature projection method Linear Discriminant Analysis(LDA). The data acquisition, data segmentation and feature extractions steps of the pattern recognition process replicates the work of Englehart et al. [2]. The classier developed for this study includes a pairwise LDA feature projection and a LD classier for binary classication in the reduced feature space. A. Reference Data set Four surface electrodes were placed on wrist exors, extensor, unlnar and radial deviators. These were used to capture resulting myoelectric signals during wrist exion, wrist extension, ulnar deviation, radial deviation and grasp tasks. These tasks were performed sequentially and represented by indexes 0 to 5 in order. This one cycle of 5 tasks were repeated 20 times. The data sets were sampled at 1 KHz using an EMG acquisition system. Although there was a labeled data set with an assigned class per 5 second window, the task was initiated from rest and terminated to rest during the 5 second period. By using thresh holds at the corresponding muscles EMG level the period which the arm was resting was identied. This captured labeled data set was used for training and validation of the developed classier. B. Data Segmentation Data segmentation involves dividing data stream in to packets to be used for feature extraction purposes. The extracted features are then classied in to control signals corresponding to the data packet considered. In order to overcome the perceived delay in prosthetic operation it is necessary to produce control decision streams at a minimum speed of 3.3Hz (300 ms delay). Englehart et al. reports a continuous windowing scheme which uses a moving window to exploit the full process capability of the available signal processor. This produces denser decision streams which can improve classication error with maximum voting post processing operations. In this study the continuous windowing scheme is implemented with experimentally selected parameters.

C. Feature Extraction Four time domain features were extracted from each EMG channel according to the data segmentation scheme selected. This includes Mean Absolute value (MAV), Zero crossings (ZC), Sign changes (SC) and Signal length (SL). This constitutes a 16 dimensional feature vector for the pattern recognition task . The alternate options reported for EMG classications include higher dimensional Wavelet Packet Transform (WPT) features and similar frequency domain features, but for overall simplicity of classication, Time Domain (TD) feature set was selected. 1) Mean Absolute Value: The estimate of the mean abso i of window segment i with a sample length lute value M AV of L takes the form i= M AV 1 L
L

x k
k

(1)

supervised projection is performed considering one class as the positive examples and all other classes as the negative examples. Therefore the original n feature k class problem transforms to k subproblems of bi-classication in a 1 dimensional feature space. The pairwise partitioning method considers only two classes at a time and performs the feature projection and classication. Therefor in pairwise LDA n Ck subproblems of bi classication is performed. In partitioned feature projection, the supervised LDA assumes that the class given feature i |Cj ) are gaussian and of equal variance distributions P (X . Since the variances of the two classes tend to change, a mean variance is used in place. Then the maximum separating hyperplane or in the two class case the maximum separating vector w is given by equation 3. The constants are neglected since we are only interested in the vector direction. w = (y=0 + y=1 )1 (y=1 y=0 ) E. Classication The simplicity in the 1 dimensional 2 class classication problem allows application of the LD classier. This classier is a statistical analytical solution to the classication problem, which can be applied when sufcient knowledge is available |Cj ). Here of the underlying probability distributions P (X the derivation is presented for a general multi class multi dimensional classication case which is easily simplied for binary classication problems. The probability of class given feature distributions |Cj ) are assumed to be multivariate normal distributions P (X with mean j and covariance matrix j . The probability of each class P (Cj ) is assumed to be multinomal distribution with equal probabilities for each class. Given a feature vector , the probability of belonging to class P (Cj |X i ) is found X by Bayes rule. (i) |Cj ) P (X 1 1 j )) exp( (X j )T 1 (X n/ 2 2 2 (i) (i) ) = P (X |Cj ) P (Cj ) P (Cj |X (i) ) P (X (i) )) Cj = argmax(P (Cj |X =
Cj

k is the signal vector from EMG channels at the k th where X sampling time. 2) Zero Crossing: number of times the signal x k crosses zero is used as a simple frequency measure. Increment the i if; zero crossing ZC (xk > 0 and xk+1 < 0) or (xk < 0 and xk+1 > 0) and (|xk xk+1 | ) where is a thresh hold to lter high frequency noise. 3) Slope Sign Changes: is another estimate of frequency by counting the number of derivative sign changes of signal xk in the window length of L. The threshold value is used to lter high frequency noise induced triggers. Increment the i if; slope sign change SC (xk1 > xk and xk < xk+1 ) or (xk1 < xk and xk > xk+1 ) and (|xk xk1 | ) and (|xk xk+1 | ) 4) Signal Length: A measure of signal complexity is i in the window of calculated by the total signal length SC length L L i= 1 |Xk | (2) SL L
k

(3)

D. Feature projection Applying feature projection to the problem effectively reduces the dimensionality of the problem. In this study we use Linear Discriminant Analysis (LDA) for supervised feature projection. LDA reduces a n dimensional feature space to k 1 dimensions where k is the number of classes in the classication problem. For this particular study two partitioning methods were assessed for feature projection and classication. Partitioning method allows to reduce a multi class classication problem in to number of subproblems of binary classication. In a one against all approach ,the

(4) The class with the maximum posterior probability ) is selected as the correct class. The terms P (Cj ) P (Cj |X ) are constants, thus neglected during logarithm and P (X maximization. j C (i) |Cj )|) = argmax(ln|P (X
Cj 1 = argmax(X T 1 j T j ) j Cj

(5)

For a two class classication problem the distribution i ) is a sigmoid function taking the form in equation 6. P (Cj |X For verication of the performance of the simple LDA classier, an algorithmic logistic classier is also implemented with

2 Channel 1 0 2 2 0 2 2 0 2 2 0 2
5 4
State

Channel 3

Channel 2

projected positive example Projected negative example Classifier threshhold P(C1|X) P(X|C1) & P(X|C0)

Channel 4

Actual state Classifier output

3 2 1 0

Probability
0.6
0 5 10 15 20 25 Time(s) 30 35 40 45 50

0.5

0.4

0.3 0.2 Projected Features (w*X)

0.1

0.1

Fig. 2. Fig. 1. Pairwise LDA classier results for rst 50 second period of the validation dataset

The partitioned classier visualization between two classes.

maximum likelihood parameter estimation using the gradient decent method. 1 (i) ) = P (Cj |X 1 + eT X T (6) 2 2 2 1 1 2 = , 2 In the one against all partitioning scheme the probability (i) ) is directly estimated, but in the pairdistribution P (Cj |X wise partitioning scheme only the feature given probability distribution between two classes, j and k are estimated (i) , k ). Thus the joint probability is found by mulP (Cj |X (i) , k ) over all k . tiplying pairwise probabilities P (Cj |X F. Validation The last 10 trials(50%) of data were used as the validation data set. The classier calculates posteriori probabilities of each feature set to select the class with maximum posteriori (i) ). Since a dense decision stream is obprobability P (Cj |X tained using the continuous windowing scheme, the decision stream is combined with a maximum voting post processing to form 1 decision for the period of allowed time delay. IV. R ESULTS The recorded dataset of 4 EMG channels were processed using a time domain feature extraction in a continuous windowing scheme, with 64ms window length and 16ms time delay for processing. Performance of 3 classiers were studied, 1) LD classifer without feature projection and same with 2) simple partitioned LDA feature projection and 3) pairwise LDA feature projection. All three methods were implemented on the same training and validation data sets. Table 2 summarize the performance of the classiers implemented i.e. LD- Linear Discriminant

classier, PLDA1- LDA classier with One against all partitioning scheme, PLDA2- LDA classier with pairwise classication scheme, PLDA3- LDA classier with pairwise classication scheme (trained by gradient decent method) .The pairwise feature projection scheme out performs the other methods in classication accuracy. The LD classier(a) was observed to achieve faster training ,thus a candidate choice for robust and simple implementations. The pairwise LDA feature projection (c) takes longer for training but achieves better accuracy and insight in to the problem under study. The longer time periods are acceptable during training because its usually carried out ofine. The processing time for validation datasets were similar in all three methods. Therefor the methods can be implemented in a realtime embedded setting.
Classier LD PLDA 1 PLDA 2 PLDA 3 Validation Accuracy 88.6% 8.5% 90.3% 91.4% Training time (sec) 3.0 78 67 500 Validation time (sec) 3.3 5 4 4

TABLE II P ERFORMANCE OF THE IMPLEMENTED CLASSIFIERS FOR EMG TASK CLASSIFICATION .

The proposed method of combining the LD classier with a pairwise LDA feature projection has allowed visualization of distributions between class pairs. This provides further information for feature selection and class selection in developing custom control methods for upper arm prosthetic cases. A rather interesting result was identied by analyzing the reduced dimensional space. All class given probabilities were resembling very closely to gaussian distributions except class 0 which was rest. This resembled a Gamma distribution.

1600 1400 1200 1000 800 600 400 200 0 0.1 1200

acquisition was supported by School of Human Kinetics and Recreation, Memorial University of Newfoundland. R EFERENCES
[1] B. Hudgins, P. Parker, and R. N. Scott, A new strategy for multifunction myoelectric control., IEEE transactions on bio-medical engineering, vol. 40, pp. 8294, Jan. 1993. [2] K. Englehart and B. Hudgins, A robust, real-time control scheme for multifunction myoelectric control., IEEE transactions on bio-medical engineering, vol. 50, pp. 84854, July 2003. [3] P. Parker, K. Englehart, and B. Hudgins, Myoelectric signal processing for control of powered limb prostheses., Journal of electromyography and kinesiology : ofcial journal of the International Society of Electrophysiological Kinesiology, vol. 16, pp. 5418, Dec. 2006. [4] L. J. Hargrove, E. J. Scheme, K. B. Englehart, and B. S. Hudgins, Multiple binary classications via linear discriminant analysis for improved controllability of a powered prosthesis., IEEE transactions on neural systems and rehabilitation engineering : a publication of the IEEE Engineering in Medicine and Biology Society, vol. 18, pp. 4957, Feb. 2010. [5] K. Englehart, B. Hudgins, and P. a. Parker, A wavelet-based continuous classication scheme for multifunction myoelectric control., IEEE transactions on bio-medical engineering, vol. 48, pp. 30211, Mar. 2001. [6] F. H. Chan, Y. S. Yang, F. K. Lam, Y. T. Zhang, and P. a. Parker, Fuzzy EMG classication for prosthesis control., IEEE transactions on rehabilitation engineering : a publication of the IEEE Engineering in Medicine and Biology Society, vol. 8, pp. 30511, Sept. 2000. [7] Y. Huang, K. B. Englehart, S. Member, B. Hudgins, S. Member, and A. D. C. Chan, A Gaussian Mixture Model Based Classicatio Scheme for Myoelectric Control of Powered Upper Limb Prostheses, IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, vol. 52, no. 11, pp. 18011811, 2005. [8] T. A. Kuiken, B. A. Lock, R. D. Lipschutz, L. A. Miller, K. A. Stubbleeld, and K. B. Englehart, TargetedMuscle Reinnervation for RealtimeMyoelectric Control ofMultifunction Articial Arms, Journal of American Medical Association, vol. 301, no. 6, pp. 619628, 2011. [9] D. Simpson et al., The choice of control system for the multimovement prosthesis: extended physiological proprioception (epp), The control of upper-extremity prostheses and orthoses, pp. 146150, 1974. [10] J. Doubler and D. Childress, An analysis of extended physiological proprioception as a prosthesis-control technique., Journal of Rehabilitation Research and Development, vol. 21, no. 1, p. 5, 1984. [11] P. J. Kyberd and A. S. e. a. Poulton, The ToMPAW Modular Prosthesis: A Platform for Research in Upper-Limb Prosthetics, JPO Journal of Prosthetics and Orthotics, vol. 19, pp. 1521, Jan. 2007. [12] Y. Losier, K. Englehart, and B. Hudgins, A control system for a powered prosthesis using positional and myoelectric inputs from the shoulder complex., Conference proceedings : ... Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Conference, vol. 2007, pp. 613841, Jan. 2007. [13] M. C. Carrozza, A. Persichetti, C. Laschi, F. Vecchi, R. Lazzarini, P. Vacalebri, and P. Dario, A Wearable Biomechatronic Interface for Controlling Robots with Voluntary Foot Movements, IEEE/ASME Transactions on Mechatronics, vol. 12, pp. 111, Feb. 2007. [14] S. Dosen, C. Cipriani, M. Kosti c, M. Controzzi, M. C. Carrozza, and D. B. Popovi c, Cognitive vision system for control of dexterous prosthetic hands: experimental evaluation., Journal of neuroengineering and rehabilitation, vol. 7, p. 42, Jan. 2010. [15] L. P. Kenney, I. Lisitsa, P. Bowker, G. H. Heath, and D. Howard, Dimensional change in muscle as a control signal for powered upper limb prostheses: a pilot study., Medical engineering & physics, vol. 21, pp. 58997, Oct. 1999. [16] K. Shima and T. Tsuji, An MMG-based human-assisting manipulator using acceleration sensors, 2009 IEEE International Conference on Systems, Man and Cybernetics, pp. 24332438, Oct. 2009. [17] J.-U. Chu, I. Moon, Y.-J. Lee, S.-K. Kim, and M.-S. Mun, A Supervised Feature-Projection-Based Real-Time EMG Pattern Recognition for Multifunction Myoelectric Hand Control, IEEE/ASME Transactions on Mechatronics, vol. 12, pp. 282290, June 2007.

0.1

0.2

0.3

0.4

0.5

0.6

0.2

0.15

0.1

0.05

0.05

0.1

0.15

0.2

Class 1 & 2

Class 1 & 3

1000

800

600

400

200

0 0.6

0.5

0.4

0.3

0.2

0.1

0.1

0.4

0.3

0.2

0.1

0.1

Class 1 & 4

Class 1 & 5

Fig. 3. Pairwise feature projection and Feature given class probability estimation results for Class 1. (6 C2 =)15 such pairs were analyzed in the pairwise partitioning scheme for the 6 class problem

This observation supports the assumptions of the LD classier for most classes which assumes gaussian class given feature distributions. V. C ONCLUSION A partitioned feature projection based EMG classier is proposed in the study in order to handle nonlinearity in EMG signals and to provide analytical ability for EMG controller development. Both one against all and a pairwise partitioning scheme were implemented. A test dataset of forearm muscle signals were processed with the proposed methods and the Linear discriminant classier. The pairwise feature projection based linear classier performed better than the widely reported method. Additionally it provides visualization capability of probability distributions in the reduced dimensional space. This can be used both for analysis and as a visual tool for clinical technicians in the training process of prosthetics. The pairwise LDA classier is selected for further studies which involve hybrid controller development for upper extremity amputees. It is expected to apply the method for shoulder and chest muscle groups for shoulder movement discrimination. The analytical advantage and the accuracy of the method is achieved without relying on complex algorithmic classication schemes. Rather the simplicity of linear discriminant classication was preserved in the developed method. ACKNOWLEDGMENT This research is supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) and Memorial University of Newfoundland, Canada. The data