ACEEE Int. J. on Electrical and Power Engineering, Vol. 01, No. 03, Dec 2010
EMG Diagnosis using Neural Network Classifier with Time Domain and AR Features *Er. Gurmanik Kaur1, Dr. A. S. Arora2, and Dr. V. K. Jain2 EIE Deptt., SLIET Longowal, Distt. Sangrur, Punjab, INDIA, 148106. Email: Gurmanik Kaur- mannsliet@gmail.com Dr. A.S.Arora- ajatsliet@yahoo.com Dr.V.K.Jain- vkjain27@yahoo.com
recruited. Different MUAPs will overlap, causing an interference pattern in which the neurophysiologist cannot detect individual MUAP shapes reliably. Traditionally, neurophysiologists assess MUAPs from their shape using an oscilloscope and listening to their audio characteristics. Thus an experienced electrophysiologist can detect abnormalities with reasonable accuracy. However, subjective MUAP assessment, although satisfactory for the detection of unequivocal abnormalities, may not be sufficient to delineate less obvious deviations or mixed patterns of abnormalities [5]. These ambiguous cases call for quantitative MUAP analysis. With the aid of computer technology, today it is possible to analyze EMG signals quantitatively that helps in saving time, standardizes the measurements and enables the extraction of additional features which cannot be easily calculated manually. To further the development of quantitative EMG techniques, the need has emerged for adding automated decision making support to these techniques so that all data is processed in an integrated environment. Towards this goal, Blinowska [6] proposed the use of discriminant analysis for the evaluation of MUAP findings, Coatrieux and associates [7]-[9] applied cluster analysis techniques for the automatic diagnosis of pathology based on MUAP records. Andreassen and co-workers [10]-[12] developed the MUNIN (Muscle and Nerve Inference Network) which employs a causal probabilistic network for the interpretation of EMG findings, Fuglsang-Frederiksen and his group [13], [14] developed a rule-based EMG expert system named KANDID, and Jamieson [15], [16] developed an EMG processing system based on augmented transition networks. In most of these systems, the generation of the input pattern assumes a probabilistic model, with the matching score representing the likelihood that the input pattern was generated from the underlying class [17]. In addition, assumptions are typically made concerning the probability density function of the input data. Recently, artificial neural networks (ANNs) have been proposed as an alternative tool to pattern recognition and classification problems. One of their major advantages is that ANN models make no assumption about the underlying probability density functions of the input data, thus possibly improving the performance of classifiers, especially when the data depart significantly from normality. Other features of artificial neural networks that
Abstract—The shapes of motor unit action potentials (MUAPs) in an electromyographic (EMG) signal provide an important source of information for the diagnosis of neuromuscular disorders. To extract this information from the EMG signals, the first step is identification of the MUAPs composed by the EMG signal, second step is clustering of MUAPs with similar shapes, third step is extraction of the features of MUAP clusters and last step is classification of MUAPs. In this work, the MUAPs are identified by using a data driven segmentation algorithm, statistical pattern recognition technique is used for clustering of MUAPs. Followed by the extraction of time domain and autoregressive (AR) features of the MUAP clusters. Finally, a neural network (NN) classifier is used for classification of MUAPs. A total of 12 EMG signals obtained from 3 normal (NOR), 5 myopathic (MYO) and 4 motor neuron diseased (MND) subjects were analyzed. The success rate for the segmentation technique is 95.90% and for the statistical technique is 93.13%. The classification accuracy of NN is 66.72% with time domain parameters and 75.06 % with AR parameters. Index Terms—Electromyography, motor unit action potentials, neural networks, classification. I. INTRODUCTION Electromyography is the study of the muscular function based on the analysis of electromyographic (EMG) signals that are electrical activities generated by muscles during voluntary, involuntary or stimulated contractions [1]. EMG signals are composed of motor unit action potentials (MUAPs) generated by different motor units (MUs). The term motor unit refers collectively to one motor neuron and the group of muscle fibers it innervates and is the smallest unit of skeletal muscle that can be activated by volitional effort. MUAPs from different MUs tend to have distinct shapes, which remain almost the same for each discharge [2-4]. When a patient maintains low level of muscle contraction, individual MUAPs can be easily recognized. As contraction intensity increases, more motor units are *Corresponding author: Er. Gurmanik Kaur, Research Scholar, EIE Deptt., SLIET, Longowal, Distt. Sangrur, Punjab, INDIA. 148106. E-mail: mannsliet@gmail.com
12 © 2010 ACEEE DOI: 01.IJEPE.01.03.70
ACEEE Int. J. on Electrical and Power Engineering, Vol. 01, No. 03, Dec 2010
B. Segmentation
make them so attractive to investigate are that: 1) they exhibit adaptation or learning, 2) they pursue multiple hypotheses in parallel, 3) they may be fault tolerant, 4) they may process degraded or incomplete data, and 5) they may create complex classification boundaries [18]. In this work, the EMG signals recorded from (NOR), myopathic (MYO) and motor neuron diseased (MND) subjects were segmented using an algorithm that automatically detects the areas of low activity and candidate MUAPs. The possible MUAPs are then clustered using a statistical pattern recognition technique. Furthermore, the time domain and autoregressive (AR) features of MUAP clusters are extracted and are given to a neural network (NN) classifier for their classification into NOR, MYO and MND classes. Finally, the performance of NN classifier with time domain and AR parameters is compared.
The next step is to cut the EMG signal into segments of possible MUAP waveforms and eliminate areas of low activity. We have evaluated three different segmentation technique and it is concluded that the segmentation of EMG signals by finding the peaks of the MUAPs, yielded the best results [19]. The segmentation algorithm calculates a threshold depending on the maximum value and the mean absolute value of the whole EMG signal, The Threshold T is calculated as:
( )
if max{x i } > i
30 L 5 L , then x T = ∑ i ∑ xi L i =1 L i =1 else T = max{xi } 5 i
Where xi are the discrete input values and L is the number of samples in the EMG signal. Peaks over the calculated threshold are considered as candidate MUAPs. Then a window of 120 sampling points (i.e. 6ms at 20 KHz) is centered at the identified peak. If a greater peak is found in the window, the window is centered at the greater peak; otherwise the 120 points are saved as MUAP waveform [20]. The segmented EMG signals of NOR, MYO and MND subjects in segments of 6ms and centered at the maximum peak, are shown in Figure 4, Figure 5 and Figure 6 respectively.
II. MATERIAL AND METHODOLOGY A. Data Acquisition and Pre-Processing Our data contain real time EMG signal obtained from the Department of Computer Science, University of Cyprus, Cyprus. All the EMG signals were acquired from the biceps brochii muscle at upto 30% of the maximum voluntary contraction (MVC) level under isometric conditions., for 5 seconds, using the standard concentric needle electrode, from NOR, MYO and MND subjects. Figure 1, Figure.2 and Figure 3 shows the typical EMG recordings. The EMG signals were analogue band pass filtered at 3-10 KHz, sampled at 20 KHz with 12-bit resolution and then low pass filtered at 8KHz.
400
300
Magnitude(µV)
200
100
0 300
-100 200
Magnitude(µV)
-200
0
20
40
100
60 Sample No.
80
100
120
Figure 4. Segmented EMG signal of a NOR subject in segments of 6ms and centered at the maximum peak.
0
-100
500
-200 0
1000
2000
3000 Sample No.
4000
5000
400 300 200 Magnitude(µV)
Figure 1. Raw EMG signal of a NOR subject. 1000
100 0 -100 -200
500
-300
M agnitude(µV)
-400 0
-500
0
20
40
60 Sample No.
80
100
120
-500
Figure 5. Segmented EMG signal of a MYO subject in segments of 6ms and centered at the maximum peak.
-1000
0
500
1000
1500
2000
2500 3000 Sample No.
3500
4000
4500
5000
1000
Figure 2. Raw EMG signal of a MYO subject.
800
M agnitude(µV)
600 1500
1000
M agnitude(µV)
500
400
200
0
0
-200
-500
-400
0
20
40
60 Sample No.
80
100
120
-1000
0
1000
2000
3000 Sample No.
4000
Figure 6. Segmented EMG signal of a MND subject in segments of 6ms and centered at the maximum peak.
5000
Figure 3. Raw EMG signal of a MND subject.
13 © 2010 ACEEE DOI: 01.IJEPE.01.03.70
ACEEE Int. J. on Electrical and Power Engineering, Vol. 01, No. 03, Dec 2010
C. MUAP Clustering In this stage, MUAP clusters are automatically detected and for each cluster the average or template shape is determined. We have used statistical pattern recognition technique for clustering of similar MUAPs. In this technique the euclidian distance is used to identify and group similar MUAP waveforms. The group average is continuously calculated and is used for the classification of MUAPs using a constant threshold [21]. The implementation steps are: Step 1: Start with the first waveform x as input (the first member of the class). Step 2: Calculate the vector length of x and the distance between it and all the other segmented waveforms y as:
500
150
400 100
300 200
50
100 0
0 -50
-100 -200
-100
-300 -150
-400 -500
0
20
40
60
80
∑
20
40
-200
0
20
40
60
80
100
120
150
100
50
0
-50
-100
-150
lx =
120
200
N
where N = 120
xi2
100
0
60
80
100
120
Figure 8. Clustered EMG signals of a MYO subject.
i =1
and
300
d xy =
∑ (x
− yi )
i
2
120
0
20
40
60
80
100
-200
120
20 0
100
120
Figure 7. Clustered EMG signal of a NOR subject.
14 © 2010 ACEEE DOI: 01.IJEPE.01.03.70
20
40
60
80
100
120
0
20
40
60
80
100
120
In this work, we have extracted the time domain and AR features of the MUAP clusters. To extract time domain features, all clustered MUAP waveforms are expanded to 25ms on the original signal where the position of identified peak was marked during segmentation. The rationale is that the MUAP duration is in most of the cases longer than 6ms and the signal expansion is therefore, necessary for a correct measurement of parameters [22]. The following time domain parameters are computed from the MUAP waveforms: Spike duration: measured from the first to the last positive peak. Amplitude: Amplitude difference between minimum positive and maximum negative peak. Area: Rectified MUAPs integrated over the calculated duration Phases: Number of baseline crossings where amplitude exceeds ±25µV, plus one. Turns: Number of positive and negative peaks where the difference from the preceding and following turn exceeds 25µV. In the autoregressive (AR) model a current signal x(n)
40
80
0
D. Feature Extraction
60
60
-200
120
Figure 9. Clustered EMG signal of a MND subject.
80
40
100
-150
100
20
80
-100
120
0
60
0
140
-20
40
-50
0
100
20
50
50
80
-150
0
100
-50
60
-100
150
-50
40
-100
200
100
20
-50
250
150
0
0
-50
-200
200
-100
50
0
-150
250
0
100
50
300
50
150
100
average and go to step 1 with group average as input. else if number of group members > 2, then form a new class. else waveform is superimposed, go to step 1 with y as input. This process continues where it stopped comparing the last encountered waveform with all the remaining until all waveforms are processed. The threshold value of 0.3 in Step 4 is critical because a smaller value may split a MUAP class with high waveform variability in two or more subclasses, whereas a greater threshold value may merge resembling MUAP classes. Figure 7, Figure 8 and Figure 9 illustrates the clustered EMG signals of NOR, MYO and MND subjects respectively.
100
200
150
Step 3: Find the waveform y with the minimum distance d min . The waveform y having minimum distance with the x has the greatest similarity with x and remove it from the input data. Step 4: if d min / l x < 0.3 then group, calculate group
150
250
200
i =1
-100
300
250
N
ACEEE Int. J. on Electrical and Power Engineering, Vol. 01, No. 03, Dec 2010
technique is 93.13%. Table 2 presents the results of clustering for each of the three MUAP classes.
is described as linear combination of previous samples x(nk) weighted by a coefficient [23]. A common strategy to calculate the AR coefficients is to use the Burgâ&#x20AC;&#x2122;s algorithm. It provides an iterative and fast method to figure out the parameters of the AR-model adaptively. We have used this method in our work to find the AR coefficients ao to a2 of a 3rd order AR model.
Table II MUAP Clustering Success Rate MUAP classes
Success rate (%)
NOR
93.13 (192/204)
MYO
93.10 (270/290)
MND
92.10 (175/190)
Total
93.13 (637/684)
E. Classification In order to classify the clustered MUAPs into NOR, MYO and MND classes, a NN classifier with back propagation (BP) training algorithm is employed. The back propagation algorithm works in much the same way as the name suggests: After propagating an input through the network, the error is calculated is propagated back through the network while the weights are adjusted in order to make the error smaller. Taking into account the problem under consideration, NN architecture with three layers is used, as this model is documented as able to draw the boundaries of arbitrarily complex decision regions [17].
After clustering of MUAPs, time domain and AR features of the MUAPs are extracted. It is observed that standard deviation (SD) of all the time domain parameters of the MYO 1 signal is zero, because in that case we have only one class of MUAPs. Moreover, the SD of all the AR parameters of a signal will be zero, if the total numbers of classes are less than two. Finally the extracted features are given to a NN classifier for training and classification of MUAPs. Table 3 presents the results of classifications.
III. RESULTS EMG data recorded from 3 NOR, 5 MYO and 4 MND subjects were analyzed using the methodology described in section II. Following the pre-processing, EMG signals are segmented by identifying the peaks of the MUAPs. Table 1 presents the success rate for each of the three MUAP classes.
Table III Comparison of the Results of NN Classifier with Time Domain and AR Features.
Table I MUAP Detection Success Rate MUAP classes
NOR
96.07 (196/204)
MYO
95.86 (278/290)
MND
95.78 (182/190)
Total
95.90 (656/684)
% Classification accuracy
Time domain features
66.72
AR features
75.06
It is observed from the results that the classification of MUAPs is better with AR features than time domain features. Moreover, AR parameters have two important advantages over time domain parameters: 1) variations in the positioning of the electrodes on the surface of the muscle do not severely affect the AR-coefficients. 2) the amount of information to be presented to the classifier is greatly reduced. Therefore, the total processing time is also reduced.
The success rate is the percentage ratio of the correctly identified MUAPs by the segmentation algorithm and the number of true MUAPs identified by manual observation. The lowest success rate (95.78%) for the MND group is attributed to the more complex and variable waveform shapes. The segmented MUAPs are clustered by using statistical pattern recognition technique. Sometimes due to waveform variability, MUAP classes coming from the same MU, although they looked simiIar, were not grouped together. Merging of these classes can be achieved using a pattern recognition technique with a greater constant threshold and the averaged class waveforms as input. The total success rate obtained by using this
CONCLUSIONS In conclusion, the methodology described in this work make possible the development of a fully automatic EMG signal analysis system which is accurate, simple, fast and reliable enough to be used in routine clinical environment. This work can provide a good understanding of EMG analysis procedures to the researchers to identify neuromuscular diseases. Future work will evaluate the algorithms developed in this study on EMG data recorded from more muscles and more subjects. In addition, this 15
Š 2010 ACEEE DOI: 01.IJEPE.01.03.70
Features Success rate (%)
ACEEE Int. J. on Electrical and Power Engineering, Vol. 01, No. 03, Dec 2010
[12] S. Andreassen, F. Jensen, S. K. Andersen, B. Falck, U. Kjaerulff, M. Woldbye, A. R. Sorensen, A. Rosenfalck, and J. Frank, “MUNIN-An expert EMG assistant,” in Computer Aided Electromygraphy and Expert Systems, J. E. Desmedt, Ed. New York : Elsevier Science Publisher B.V., pp. 255277, 1989. . [13] A. FugIsang-Frederiksen and S. M. Jeppesen, “A rule-based EMG expert system for diagnosing neuromuscular disorders,” in Computer Aided Electromyography and Expert Systems, J. E. Desmedt, Ed. New York: Elsevier Science Publishers B.V., pp. 289-296, 1989. [14] A. Fuglsang-Frederiksen, J. Ronager, and S, Vingtoft, “A plan-test-diagnose expert system for EMG: KANDID,” J.Neurolog. Sci., vol.98 (suppl.), p. 150, 1990. [15] P. W. Jamieson, “Computerized interpretation of electromyographic data.” Electroencephalogr. Clin. Neurophysiol., vol. 75, p. 392, 1990. [16] P. W. Jamieson, “A model for diagnosing and explaining multiple disorders,” Comput. Biomed. Res., vol. 24, pp. 307320, 1991. [17] R. P. Lippmann, “An introduction to computing with neural nets,” IEEE ASSP Mag. pp. 4-22, 1987. [18] Constantinos S. Pattichis, Christos N. Schizas and Lefkos T. Middleton, “Neural Network Models in EMG Diagnosis,” IEEE Trans Biomed Eng., vol. 42, 1995. [19] Gurmanik kaur, A. S. Arora and V. K. Jain, “Comparison of the techniques used for segmentation of EMG signals,” in Proc. WSEAS Int. Conf. on Mathematical and Computational Methods, Baltimore, USA, pp. 124-129, 2009. [20] Christodoulos I. Christodoulou and Constantinos S. Pattichis, “Unsupervised Pattern Recognition for the Classification of EMG Signals,” IEEE Transactions on Biomedical Engg., vol. 46, pp. 169-178,1999. [21] C. I. Christodoulou and C. S. Pattichis, “A new technique for the classification and decomposition of EMG signals,” in Proc. IEEE Int. Conf. on Neural Networks, Perth, Western Australia, vol. 5, pp. 2303–2308, Nov. 1995. C.D. Katsis, D.I. Fotiadis, A. Likas and I. Sarmas, Automatic discovery of the number of MUAP clusters and superimposed MUAP decomposition in electromyograms,” in Proc. of the 4th Annual IEEE Conf on Information Technology Applications in Biomedicine, UK., pp. 177-180, 2003.Marple S. L. Jr., Digital Spectral analysis with applications,Prentice-Hall,USA,1987.
system may be integrated into a diagnostic system for neuromuscular diseases based on neural network where EMG, muscle biopsy, biochemical and molecular genetic findings, and clinical data may be combined to provide a diagnosis. REFERENCES [1] A. O. Andrade, Decomposition and analysis of electromyographic signals, Ph. D. Thesis, University of Reading, UK, 2005. [2] DeLuca CJ, Towards understanding the EMG signal, 4th ed., Baltimore: Williams & Wilkinson; 1978. [3] Krarup C, “Pitfalls in electrodiagnosis,” J Neurophysiol, vol. 81, pp. 1115-1126, 1999. [4] McGill KC, “Optimal resolution of superimposed action potentials,” IEEE Trans Biomed Engg, vol. 49, pp. 640-650, 2002. [5] Richfield EK, Cohen BA, Albers JW, “Review of quantitative and automated needle electromyographic analyses,” IEEE Trans Biomed Eng, pp. 506-514, 1981. [6] K. J. Blinowska, I. Hausmanowa-Petrusewicz, A. MillerLarsson and J. Zachara, “The analysis of single EMG potentials by means of multivariate methods,” Electromyogr. Clin. Neurophysiol, vol. 20, pp. 105-123, 1980. [7] J. L. Coatrieux, P. Toulouse, B. Rouvrais, and R. Le Bars, “Automatic classification of electromyographic signals,” EEG Clin. Neurophysiol., vol. 55, pp. 333-341, 1983. [8] B. Rouvrais, P. Toulouse, J. L. Coatrieux. and R. Le Bars, “A possible method of automatic electromyographic analysis and diagnosis on line,” Electmmyogr. Clin. Neurophysiol.. vol. 23, pp. 457-470, 1983. [9] P. Toulouse, J. L. Coatrieux. and B. Le Marec, “An attempt to differentiate female relatives of Duchenne type dystrophy from healthy subjects using an automatic EMG analysis,” J. Neurolog. Sci., vol. 67, pp. 45-55, 1985. [10] S. K. Andersen, S. Andreassen, and M. Woldbye, “Knowledge representation for diagnosis and test planning in the domain of EMG.” in Proc. 7th Eur. Conf. Artijcicial 1ntell., Brighton, U.K., pp. 357-368, 1986. [11] S. Andreassen, S. K. Andersen, F. V. Jensen, M. Woldbye, A. Rosen- falck, B. Falck, U. Kjaerluff, and A. R. Sorensen, “MUNIN-An expert system for EMG,” Electroenceph,ph. Clin. Neurophvsiol., vol. 66, 1987.
16 © 2010 ACEEE DOI: 01.IJEPE.01.03.70