400 IEEE TRANSACTIONS ON REHABILITATION ENGINEERING, VOL. 6, NO.

4, DECEMBER 1998

EMG Pattern Recognition Based on

Artificial Intelligence Techniques
Sang-Hui Park, Member, IEEE, and Seok-Pil Lee

Abstract— This paper presents an electromyographic (EMG) resulted in some theoretical and practical achievements for
pattern recognition method to identify motion commands for powered prosthetic arms, further advancement such as accurate
the control of a prosthetic arm by evidence accumulation based identification of motion and exact modeling of EMG signals,
on artificial intelligence with multiple parameters. The integral
absolute value, variance, autoregressive (AR) model coefficients, is required to achieve an ultimate goal [1].
linear cepstrum coefficients, and adaptive cepstrum vector are This paper presents an EMG pattern recognition method
extracted as feature parameters from several time segments of for more accurate identification of a motion command. The
EMG signals. Pattern recognition is carried out through the proposed method based on artificial intelligence (AI) [18]
evidence accumulation procedure using the distances measured is able to accommodate expected interindividual differences
with reference parameters. A fuzzy mapping function is designed
to transform the distances for the application of the evidence and requires little computing time in the pattern recognition
accumulation method. Results are presented to support the fea- with the extracted feature parameters. Based on previous
sibility of the suggested approach for EMG pattern recognition. research [1], [6], [15], [19], integral absolute value, difference
Index Terms— Artificial intelligence, electromyographic absolute mean value, variance, autoregressive (AR) model
(EMG), pattern recognition, prosthesis. coefficients, and linear cepstrum coefficients, are extracted
as feature parameters. Also considering the nonstationary
property [9] of EMG signals, the adaptive cepstrum vector
I. INTRODUCTION [7] is extracted as a feature parameter. To evaluate the utility

I T has been proposed that the electromyographic (EMG)

signals from the body’s intact musculature can be used to
identify motion commands for the control of an externally
of the above feature parameters for EMG pattern recognition, a
simple separability measure is provided. The Dempster–Shafer
theory of evidence [20]–[22] is employed as an evidence
powered prosthesis [1]–[5]. Information extracted from EMG accumulation method for the pattern recognition. A fuzzy
signals, represented in a feature vector, is chosen to minimize mapping function is designed for the application of the Demp-
the control error [6]. In order to achieve this, a feature set ster–Shafer theory of evidence. Finally, a series of evidence
must be chosen which maximally separates the desired output accumulation procedures according to each motion and the
classes. The extraction of accurate features from the EMG recognition error rates are provided.
signals is the main kernel of classification systems and is
essential to the motion command identification [7]. But the II. FEATURE PARAMETERS
nonstationarity of the EMG signal makes it difficult to extract
The success of any pattern classification system depends
feature parameters precisely with the block processing station-
almost entirely on the choice of features used to represent
ary model such as an autoregressive (AR) model [8]–[10]. And
the raw signals [6]. It is desirable to use multiple feature
it is very difficult for one feature parameter to reflect the unique
parameters for EMG pattern classification since it is very
feature of the measured EMG signals to a motion command
difficult to extract a feature parameter which reflects the unique
perfectly. Once a feature set has been chosen, a suitable pattern
feature of the measured signals to a motion command per-
classifier can be used to determine class output.
fectly. But the inclusion of an additional feature parameter with
Several approaches to solve the motion command identi-
a small separability may degrade overall pattern recognition
fication problem using EMG signals have been suggested,
performance [23], [24].
e.g., stationary time series modeling (AR model) of the EMG
Considering previous works [1], [6], [7], [13], [15], [19],
[11]–[13], use of a linear discriminant function [14], use of
the following feature parameters based on time and spectral
a learning linear classifier [15], [16], and use of artificial
statistics are chosen to represent the myoelectric pattern.
neural networks [6], [17]. Although previous works have
1) Integrated Absolute Value (IAV): This is an estimate of
mean absolute value of the signal, in segment which
Manuscript received November 11, 1997; revised March 24, 1998 and
August 24, 1998. This work was supported by a Special Fund for University is samples in length, as given by
Research Institute, Korea Research Foundation.
S.-H. Park is with the Department of Electrical Engineering, Yonsei
University, Seoul, Korea. (1)
S.-P. Lee is with the Digital TV Research Center, Daewoo Electronics
Company, Ltd., Seoul, Korea.
Publisher Item Identifier S 1063-6528(98)09177-0. where is the th sample in segment .
1063–6528/98$10.00  1998 IEEE

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on December 03,2024 at 05:08:55 UTC from IEEE Xplore. Restrictions apply.
PARK AND LEE: EMG PATTERN RECOGNITION ON AI TECHNIQUES 401

2) Difference Absolute Mean Value (DAMV): This is the

mean absolute value of the difference between the adja-
cent samples, and as defined by

(2)

3) Variance (VAR): This is an estimate of the variance of

the signal in segment as defined by
(3)
where is the expected value of the signal in
segment .
4) AR Model Coefficients (ARC): This is the feature param-
eter based on spectral statistics and contains information Fig. 1. Schematic diagram.
about the location of peaks of the signal on its spectrum.
5) Linear Cepstrum Coefficients (LCC): This is the feature
parameter based on spectral statistics and comprises the and body weight fluctuations will produce changes in feature
accurate spectrum information of the signals. values over time. A suitable recognition method must be able
6) Adaptive Cepstrum Vector (ACV): This is the enhanced to accommodate the expected individual differences. And it
feature parameter based on spectral statistics and is must generate reasonably accurate results with the extracted
extracted by the algorithm which combines with block feature parameters as well as minimize the required comput-
and adaptive processing. ing time for real time prosthesis control [6]. The evidence
These features are extracted from each time segment to accumulation method was chosen as the recognition method
create the total feature set used to represent the myoelectric for this application. Fig. 1 illustrates the schematic diagram of
pattern. The total number of feature parameters is determined the proposed approach for EMG pattern recognition.
by the number of time segments in the pattern. Although the First of all, the set of feature parameters introduced in
variance in the time structure of the signals is high, waveform Section II is extracted from the sample EMG signals per each
statistics may be stable enough to allow pattern classification. movement. The mean of each feature parameter is stored
The effect of segment length on classification accuracy must be per individual as the reference parameter. And a distance
examined to determine the value which is the best compromise measure between the reference parameters and a set of feature
between class information and feature estimation error. parameters extracted from a test EMG signal is computed. The
Considering the previous work [6], [11], [13], [25], the Euclidean distance [26], [27] is used as the distance measure.
segment length and overlap rate was determined to be 64 ms Then an evidence accumulation procedure is carried out for
and 0.5, respectively, and the order of filter at ARC, LCC, pattern recognition where a fuzzy mapping function is used
and ACV was determined to be 6 in this paper. To evaluate to transform the distances measured for the application of
the feasibility of the above feature parameters for EMG pattern the Dempster–Shafer theory of evidence. Finally the motion
classification, a simple test of separability measure is provided corresponding to the EMG signal is identified.
by the Bhattacharyya distance [26] in the test result. The
Bhattacharyya distance (1/2) is used as an important measure A. Evidence Accumulation
of the separability between distributions The shortcomings of evidence accumulation schemes com-
monly employed in rule based expert systems based on the
MYCIN model [28] have been recognized. The classical
theory and the conventional MYCIN rule have an undesirable
(4) flaw in the combining of opposing pieces of evidence—the
greater the weights of contradicting evidence, the greater
the resulting certainty in their accumulation. In contrast, the
where is the mean of class 1 and 2, respectively,
Dempster–Shafer theory of evidence not only eliminates that
is the covariance of class 1 and 2, respectively.
problem but also provides a graceful certainty estimate as such
contradictory evidence accumulates [20].
III. PATTERN RECOGNITION BASED Four components, e.g., evidence for , evidence against
ON EVIDENCE ACCUMULATION
, neutral evidence , and contradictory evidence ,
There are several factors which must be considered when are used to represent the evidence of an event in the Demp-
choosing a classifier or a recognition method for the present ster–Shafer theory. Each component is a number in the range
application. Due to the nature of the myoelectric signal, it [0, 1]. The accumulation of evidences for a class is illustrated
is reasonable to expect a large variation in the value of in Fig. 2.
a particular feature between individuals. Many factors such The combining operator has closure, commutivity, and
as changes in electrode position, myoelectric signal training, associativity. Fig. 2(a)–(c), etc., represents feature parameters

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on December 03,2024 at 05:08:55 UTC from IEEE Xplore. Restrictions apply.
402 IEEE TRANSACTIONS ON REHABILITATION ENGINEERING, VOL. 6, NO. 4, DECEMBER 1998

Fig. 2. Evidence accumulation procedure.

Fig. 4. An example of the evidence boundary.

Fig. 3. Fuzzy mapping function.

Fig. 5. Mean of variances.

such as IAV, DAMV, VAR, ARC, LCC, and ACV. The

evidences for the class computed by using multiple feature It is useful in pattern recognition since the difference between
parameters, e.g., (a)–(c), etc. in Fig. 2, are accumulated one the values of the functions at the vicinity of the boundary of
after another in each class. After the above procedure is done each class is large [29]. Then the components of evidence are
per class, the class which has the maximal final value of ef is formed as follows:
chosen as the motion class corresponding to the EMG signal. MAX
(6)
MAX.
B. Fuzzy Mapping In case the distance d is smaller than MAX, only
To apply the Dempster–Shafer theory of evidence to EMG and exist according to the distance since the input feature
pattern recognition, the components of evidence are deter- parameter can be recognized as the evidence for the class. On
mined based on the distance d between the sample parameters the contrary, in the case that is larger than MAX, only
and a series of feature parameters extracted from the test and according to the distance exist since the input feature
EMG signals. A fuzzy mapping function is designed to parameter cannot be recognized as the evidence for the class.
transform the distance and is shown in (5) and Fig. 3 The value of each component of evidence according to each
feature parameter per motion class can be obtained from (6).
MAX, MAX Fig. 4 represents an example of the evidence boundary of three
(5)
MAX. classes in two-dimensional (2-D) distribution.
MAX The evidence for each class exists only inside the bound-
MAX in Fig. 3 is the maximal value of the difference ary and the evidence against each class exists only outside
between the mean and the elements in a distribution of the boundary. X1 and X2 represent the feature parameters
each parameter from the sample EMG signals per class per in Fig. 4. Using the multiple feature parameters defined in
individual. The fuzzy mapping function varies largely at the Section II, the evidence boundary in a six-dimensional (6-D)
vicinity of the point where the value of the distance is MAX. distribution is formed.

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on December 03,2024 at 05:08:55 UTC from IEEE Xplore. Restrictions apply.
PARK AND LEE: EMG PATTERN RECOGNITION ON AI TECHNIQUES 403

(a) (b)

(c) (d)

(e) (f)
Fig. 6. Separability between classes.

IV. TEST RESULTS used to measure the EMG signals. They were located over two
sites: on the bulge of biceps brachii and on the lateral head
Six basic motions, e.g., elbow flexion (FL) and extension of triceps. The 1 bipolar recording was used to measure the
(EX), wrist pronation (PR) and supination (SU), and humeral EMG signals. The measured signals were bandpass-filtered
rotation in (RI) and rotation out (RO), were considered as (10–500 Hz) and amplified by 1000 and then sampled at 1
pattern classes. kHz. A four-pole Butterworth bandpass filter is used for the
Floating point surface electrodes, silver/silver chloride test. The EMG signals were recorded for 50 repetitions of
pregelled disposable electrodes with hypoallergenic tape, were each motion from six subjects ages 21–35, with no known

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on December 03,2024 at 05:08:55 UTC from IEEE Xplore. Restrictions apply.
404 IEEE TRANSACTIONS ON REHABILITATION ENGINEERING, VOL. 6, NO. 4, DECEMBER 1998

TABLE I TABLE II
EVIDENCE ACCUMULATION PROCEDURE; FOR TEST SIGNALS BELONGING TO FL RECOGNITION ERROR RATES: USING THE DISTANCES WITH ARC
(A); WITH LCC (B); WITH ACV (C); USING THE SUM OF THE
DISTANCES WITH MULTIPLE PARAMETERS (D); USING THE EVIDENCE
ACCUMULATION METHOD WITH MULTIPLE PARAMETERS (E)

V. CONCLUSIONS
We proposed an EMG pattern recognition method to identify
motion commands for the control of a prosthetic arm by
evidence accumulation with multiple parameters. A series of
evidence accumulation procedure showed that the proposed
method recognized the desired motion efficiently with the
multiple incomplete feature parameters. And the separability
test showed that ACV is more feasible for EMG pattern
classification than the other feature parameters. Also, the
recognition error rates showed that the proposed recognition
method is superior to the other methods for EMG pattern
recognition. This approach to EMG pattern recognition focuses
on generating reasonably accurate results with less computing
Fig. 7. Value of ef according to the procedure of evidence accumulation time using the extracted feature parameters and little subject
for test signals belonging to FL. training, it seems advantageous over other techniques that
require considerable training. Further work is recommended to
neurological or neuromuscular deficits. The recording time is find the optimal feature parameters to be used as inputs to the
1 s for one time. EMG pattern classifier and to enhance the decision algorithm
Fig. 5 shows the mean values of variances of ARC, LCC, for more precise pattern recognition with the accumulated
and ACV per subject. evidences.
As shown in Fig. 5, ACV presented a stable feature for
EMG signals since the variances of ACV are smaller than REFERENCES
those of ARC and LCC. Fig. 6 presents the separability [1] S. Lee, and G. N. Saridis, “The control of a prosthetic arm by EMG
between classes for the tested parameters: IAV, DAMV, VAR, pattern recognition,” IEEE Trans. Automat. Contr., vol. AC-29, pp.
ARC, LCC, and ACV, by the Bhattacharyya distance. As 290–302, 1984.
[2] P. A. Parker and R. N. Scott, “Myoelectric control of prostheses,” CRC
shown in Fig. 6, ACV is superior to the other feature param- Crit. Rev. Biomed. Eng., vol. 13, pp. 283–310, 1986.
eters in separability for EMG pattern classification. [3] R. N. Scott and P. A. Parker, “Myoelectric prosthesis: State of the art,”
J. Med. Eng. Technol., vol. 12, pp. 143–151, 1988.
An example of evidence accumulation procedure according [4] D. Dorcas and R. N. Scott, “A three state myoelectric control,” Med.
to the test EMG signals is shown in Table I. The evidence is Biol. Eng., vol. 4, pp. 367–372, 1966.
represented as in Table I. The variation of the [5] M. Korenberg and E. Morin, “Parallel cascade classification of myo-
electric signals,” in Proc. IEEE Conf. Eng. Med. Biol., Montreal, P.Q.,
value of according to the procedure of evidence accumula- Canada, 1995, pp. 1399–1400.
tion is represented in Fig. 7. The number on the -axis repre- [6] B. Hudgins, P. Parker, and R. N. Scott, “A new strategy for multifunc-
sents the accumulation number in Fig. 7. The maximal value tion myolectric control,” IEEE Trans. Biomed. Eng., vol. BME-40, pp.
82–94, 1993.
of the number is six since there are six feature parameters. [7] S.-P. Lee, J.-S. Kim, and S.-H. Park, “An enhanced feature extraction
As shown in the evidence accumulation procedure, the pro- algorithm for EMG pattern classification,” IEEE Trans. Rehab. Eng.,
posed classifier recognized the desired motion corresponding vol. 4, pp. 439–443, 1996.
[8] C. J. De Luca, “Physiology and mathematics of myoelectric signals,”
to the test EMG signals as the motion with the maximal final IEEE Trans. Biomed. Eng., vol. BME-26, pp. 313–325, 1979.
value of although there are some undesirable evidences [9] E. Shwedyk, R. Balasubraminian, and R. N. Scott, “A nonstationary
model for the electromyogram,” IEEE Trans. Biomed. Eng., vol. BME-
resulted from the inaccurate feature parameters. The recogni- 24, pp. 417–424, 1977.
tion error rates with several methods per motion are shown in [10] T. Kiryu, Y. Saitoh, and K. Ishioka, “Investigation on parametric
Table II. The error rates in Table II are normalized by those analysis of dynamic EMG signals by a muscle-structured simulation
model,” IEEE Trans. Biomed. Eng., vol. 39, pp. 280–288, 1992.
used for ARC since the recognition error rates depend on the [11] D. Graupe, and W. K. Cline, “Functional separation of EMG signals via
experimental environment. ARMA identification methods for prosthesis control purposes,” IEEE

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on December 03,2024 at 05:08:55 UTC from IEEE Xplore. Restrictions apply.
PARK AND LEE: EMG PATTERN RECOGNITION ON AI TECHNIQUES 405

Trans. Syst. Man Cyber., vol. SMC-5, pp. 252–259, 1975. [26] K. Fukunaga, Introduction to Statistical Pattern Recognition, 2nd ed.
[12] D. Graupe, J. Magnussen, and A. A. Beex, “A microprocessor system for San Diego, CA: Academic, 1990.
multifunctional control of upper-limb prostheses via myoelectric signal [27] D. O’Shaughnessy, Speech Communication-Human and Machine.
identification,” IEEE Trans. Automat. Contr., vol. AC-23, pp. 538–544, Reading, MA: Addison-Wesley, 1987.
1978. [28] B. G. Buchanan and E. H. Shortliffe, Rule-Based Expert Systems.
[13] G. Hefftner, W. Zucchini, and G. G. Jaros, “The electromyogram (EMG) Reading, MA: Addison-Wesley, 1984.
as a control signal for fuctional neuromuscular stimulation—Part I: [29] S-P. Lee, S-H. Park, J-S. Kim, and I-J. Kim, “EMG pattern recognition
Autoregressive modeling as a means of EMG signature discrimination,” based on evidence accumulation for prosthesis control,” in Proc. IEEE
IEEE Trans. Biomed. Eng., vol. BME-35, pp. 230–237, 1988. Conf. Eng. Med. Biol., Amsterdam, The Netherlands, paper #5.7.3, 1996.
[14] R. W. Wirta, D. R. Taylor, and F. R. Finley, “Pattern recognition arm
prostheses: A historical perspective-A final report,” Bull. Prothes. Res.,
Fall 1978, pp. 8–35.
[15] G. N. Saridis and T. Gootee, “EMG pattern analysis and classification
for a prosthetic arm,” IEEE Trans. Biomed. Eng., vol. BME-29, pp. Sang-Hui Park (M’88) received the B.S., M.S., and
403–409, 1982. Ph.D. degrees in electrical engineering from Yonsei
[16] G. N. Saridis and M. A. Newman, “Upper limb EMG statistical University, Seoul, Korea, in 1962, 1964, and 1971,
analysis,” in Proc. MIDCON 1979, Chicago, IL, 1979. respectively.
[17] N. Lan, H-Q. Feng, and P. E. Crago, “Neural network generation of From 1964 to 1969, he was an Assistant Pro-
muscle stimulation patterns for control of arm movements,” IEEE Trans. fessor at Soodo Engineering College. He has been
Rehab. Eng., vol. 2, pp. 213–224, Dec. 1994. the Director of the Institute of Medical Instrument
[18] P. H. Winston, Artificial Intelligence, 3rd ed. Reading, MA: Addison- Technology at Yonsei University. Since 1970, he has
Wesley, 1992. been with Yonsei University, where he is currently a
[19] L. R. Rabiner and R. W. Schafer, Digital Processing of Speech Signals. Professor in the Electrical Engineering Department.
Englewood Cliffs, NJ: Prentice-Hall, 1978. His research interests are biomedical signal and
[20] B. P. Zeigler, “Some properties of modified Dempster-Shafer operators image processing, neural network, and rehabilitation engineering.
in rule-based inference systems,” Int. J. General Syst., vol. 14, pp. Dr. Park is a member of KOSOMBE, IEEK, KIEE, and AKS.
345–356, 1988.
[21] R. R. Yager, “On the Dempster-Shafer framework and new combination
rules,” Memo MII-504, Machine Intelligence Institute, Iona College,
New Rochelle, New York, 1985.
[22] , “On the relationships of methods of aggregation of evidence in Seok-Pil Lee received the B.S., M.S., and Ph.D.
expert systems,” Cybern. Syst., vol. 16, pp. 1–21, 1985. degrees in electrical engineering from Yonsei Uni-
[23] C. Berzuini, “Evaluation of the effectiveness of EMG parameters in the versity, Seoul, Korea, in 1990, 1992, and 1997,
study of neurogenic dieases-A statistical approach using clinical and respectively.
simulated data,” IEEE Trans. Biomed. Eng., vol. BME-32, Jan. 1985. Since July 1997, he has been working as a Senior
[24] D. Lowe and A. R. Webb, “Optimized feature extraction and Bayes Research Engineer in Digital TV Research Center of
decision in feed-forward classifier networks,” IEEE Trans. Pattern Anal. DAEWOO Electronics Co., Ltd., Seoul, Korea. His
Machine Intell., vol. 13, pp. 355–364, 1991. research interests are AI, pattern recognition, neural
[25] D. Groupe, J. Salahi, and Z. Desong, “Stochastic analysis of myo- network, digital signal processing, and multimedia
electric temporal signatures for multifunctional single-site activation of communication.
prostheses and orthoses,” J. Biomed. Eng., vol. 7, pp. 18–29, 1985. Dr. Lee is a member of IEEK, KIEE.

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on December 03,2024 at 05:08:55 UTC from IEEE Xplore. Restrictions apply.