See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/45630828

Muscle Fatigue Detection in EMG Using Time–Frequency Methods, ICA and

Neural Networks

Article  in  Journal of Medical Systems · August 2010

DOI: 10.1007/s10916-009-9292-7 · Source: PubMed

CITATIONS READS

56 617

2 authors:

Abdulhamit Subasi M. Kemal Kiymik

Effat University Kahramanmaras Sutcu Imam University
169 PUBLICATIONS   5,006 CITATIONS    29 PUBLICATIONS   937 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Comparison of Different Feature Extraction and Machine Learning Techniques in EEG-Based Wireless BCI System in a Cloud View project

Event Driven ECG acquisition, processing and classification for an effective Diagnosis of the Cardiovascular Diseases View project

All content following this page was uploaded by Abdulhamit Subasi on 07 April 2015.

The user has requested enhancement of the downloaded file.

Dear Author
Here are the proofs of your article.

• You can submit your corrections online or by fax.

• For online submission please insert your corrections in the online correction form.
Always indicate the line number to which the correction refers.
• Please return your proof together with the permission to publish confirmation.
• For fax submission, please ensure that your corrections are clearly legible. Use a
fine black pen and write the correction in the margin, not too close to the edge of the
page.
• Remember to note the journal title, article number, and your name when sending
your response via e-mail, fax or regular mail.
• Check the metadata sheet to make sure that the header information, especially
author names and the corresponding affiliations are correctly shown.
• Check the questions that may have arisen during copy editing and insert your
answers/corrections.
• Check that the text is complete and that all figures, tables and their legends are
included. Also check the accuracy of special characters, equations, and electronic
supplementary material if applicable. If necessary refer to the Edited manuscript.
• The publication of inaccurate data such as dosages and units can have serious
consequences. Please take particular care that all such details are correct.
• Please do not make changes that involve only matters of style. We have generally
introduced forms that follow the journal’s style.
Substantial changes in content, e.g., new results, corrected values, title and
authorship are not allowed without the approval of the responsible editor. In such a
case, please contact the Editorial Office and return his/her consent together with the
proof.
• If we do not receive your corrections within 48 hours, we will send you a reminder.

Please note
Your article will be published Online First approximately one week after receipt of
your corrected proofs. This is the official first publication citable with the DOI.
Further changes are, therefore, not possible.
After online publication, subscribers (personal/institutional) to this journal will have
access to the complete article via the DOI using the URL:
http://dx.doi.org/10.1007/s10916-009-9292-7
If you would like to know when your article has been published online, take advantage
of our free alert service. For registration and further information, go to:
http://www.springerlink.com.
Due to the electronic nature of the procedure, the manuscript and the original figures
will only be returned to you on special request. When you return your corrections,
please inform us, if you would like to have these documents returned.
The printed version will follow in a forthcoming issue.
To: Springer Customer Support 3
E-mail: CorrAdmin3@spi-bpo.com
Fax: +1-703-5620785
SPi SPi Building, Sacsac Bacong Oriental Negros 6216 Philippines

Re: Journal of Medical Systems DOI 10.1007/s10916-009-9292-7

Muscle Fatigue Detection in EMG Using Time?Frequency Methods, ICA and Neural Networks
Subasi · Kiymik

Permission to publish
I have checked the proofs of my article and

‰ I have no corrections. The article is ready to be published without changes.

‰ I have a few corrections. I am enclosing the following pages:
‰ I have made many corrections. Enclosed is the complete article.

Date / signature: _________________________________________________

 Testsite Page 1 of 2
AUTHOR'S PROOF

Metadata of the article that will be visualized in Online

1 Article Title Muscle Fatigue Detection in EMG Using Time–Frequency Methods,

ICA and Neural Networks
2 Article Sub- Title
3 Article Copyright - Springer Science + Business Media, LLC 2009
Year (This will be the copyright line in the final PDF)
4 Journal Name Journal of Medical Systems
5 Family Name Subasi
6 Particle
7 Given Name Abdulhamit
8 Corresponding Suffix
9 Author Organization Kahramanmaras Sutcu Imam University
10 Division Department of Electrical and Electronics Engineering
11 Address Kahramanmaraş 46500, Turkey
12 e-mail asubasi@ksu.edu.tr
13 Family Name Kiymik
14 Particle
15 Given Name M. Kemal
16 Suffix
Author
17 Organization Kahramanmaras Sutcu Imam University
18 Division Department of Electrical and Electronics Engineering
19 Address Kahramanmaraş 46500, Turkey
20 e-mail mkemal@ksu.edu.tr
21 Received 25 January 2009
22 Schedule Revised
23 Accepted 6 April 2009
24 Abstract The electromyography (EMG) signals give information about different features
of muscle function. Real-time measurements of EMG have been used to
observe the dissociation between the electrical and mechanical measures that
occurs with fatigue. The purpose of this study was to detect fatigue of biceps
brachia muscle using time–frequency methods and independent component
analysis (ICA). In order to realize this aim, EMG activity obtained from activated
muscle during a phasic voluntary movement was recorded for 14 healthy young
persons and EMG signals were observed in time–frequency domain for
determination of fatigue. Time–frequency methods are used for the processing
of signals that are non-stationary and time varying. The EMG contains transient
signals related to muscle activity. The proposed method for the detection of
muscle fatigue is automated by using artificial neural networks (ANN). The
results show that ANN with ICA separates EMG signals from fresh and fatigued
muscles, hence providing a visualization of the onset of fatigue over time. The
system is adaptable to different subjects and conditions since the techniques
used are not subject or workload regime specific.
25 Keywords Electromyography (EMG) - Muscle fatigue - Time–frequency methods -

file://C:\Programs\Metadata\temp\MES99292.htm 4/15/2009
 Testsite Page 2 of 2
AUTHOR'S PROOF
separated by ' - ' Independent Component Analysis (ICA) - Artificial Neural Network (ANN)
26 Foot note
information

file://C:\Programs\Metadata\temp\MES99292.htm 4/15/2009
AUTHOR'S PROOF JrnlID 10916_ArtID 9292_Proof# 1 - 15/04/2009

1 J Med Syst
2 DOI 10.1007/s10916-009-9292-7

3 ORIGINAL PAPER

4 Muscle Fatigue Detection in EMG Using Time–Frequency

5 Methods, ICA and Neural Networks
6 Abdulhamit Subasi & M. Kemal Kiymik

F
7 Received: 25 January 2009 / Accepted: 6 April 2009

O
8 # Springer Science + Business Media, LLC 2009

O
10 Abstract The electromyography (EMG) signals give in- Introduction 35

PR
11 formation about different features of muscle function. Real-
12 time measurements of EMG have been used to observe the Works that require repetitive action carried out with upper 36
13 dissociation between the electrical and mechanical meas- limbs are frequent. They usually involve simultaneous
D 37
14 ures that occurs with fatigue. The purpose of this study was manipulations that are done with upper limbs, commonly 38
15 39
TE
to detect fatigue of biceps brachia muscle using time– at a fixed repetition rate. This kind of upper limb activity,
16 frequency methods and independent component analysis which is called as cyclic work or repetitive task, composed 40
17 (ICA). In order to realize this aim, EMG activity obtained of sequences of tasks. The period of the cycle and the 41
EC

18 from activated muscle during a phasic voluntary movement relative external force of each cycle period determine the 42
19 was recorded for 14 healthy young persons and EMG conditions of the external workload while repetitive tasks 43
20 signals were observed in time–frequency domain for are performed [1]. It has been demonstrated that repetitive 44
R

21 determination of fatigue. Time–frequency methods are used tasks make work mostly hazardous [2–4]. The relationship 45
R

22 for the processing of signals that are non-stationary and between the external workload characterised by the afore- 46
23 time varying. The EMG contains transient signals related to mentioned parameters of a repetitive task and musculoskeletal 47
O

24 muscle activity. The proposed method for the detection of load or fatigue has been evidenced in numerous studies, in 48
C

25 muscle fatigue is automated by using artificial neural which individual authors discussed various indices of fatigue 49
26 networks (ANN). The results show that ANN with ICA of the musculoskeletal system [4–10]. 50
N

27 separates EMG signals from fresh and fatigued muscles, The existence of fatigue can be determined analyzing the 51
U

28 hence providing a visualization of the onset of fatigue over EMG signal. We can define the muscle fatigue as decrease in 52
29 time. The system is adaptable to different subjects and the force generating capacity of a muscle. This kind of change 53
30 conditions since the techniques used are not subject or in mechanical performance capacity results in EMG changes. 54
31 workload regime specific. The effect of fatigue is observable in the recorded EMG signal 55
as a change of values of selected EMG parameters. Those 56
32 Keywords Electromyography (EMG) . Muscle fatigue . individual parameters can be achieved as a consequence of 57
33 Time–frequency methods . Independent Component Analysis processing the EMG signal in the time and frequency 58
34 (ICA) . Artificial Neural Network (ANN) domains. Muscle fatigue causes a shift of the EMG power 59
spectrum into lower frequency [4, 11]. Successful analysis of 60
transient surface electromyogram (SEMG) signals for fatigue 61
analysis necessitates proper spectral estimation techniques. 62
A. Subasi (*) : M. K. Kiymik The use of short-time Fourier transform (STFT) stays away 63
Department of Electrical and Electronics Engineering,
from the question of stationary signals by describing the 64
Kahramanmaras Sutcu Imam University,
46500 Kahramanmaraş, Turkey time-interval to be used in the computation. However, there 65
e-mail: asubasi@ksu.edu.tr are limitations in the use of Fourier transform due to the time 66
M. K. Kiymik frequency resolution [5]. Particularly, the signal must be of 67
e-mail: mkemal@ksu.edu.tr time-invariant (stationary) or periodic frequency content 68
AUTHOR'S PROOF
JrnlID 10916_ArtID 9292_Proof# 1 - 15/04/2009

J Med Syst

69 within the analysis window, otherwise, the resulting spec- 176.5 cm, SD 4.5 cm) and mass varied from 53 to 80 kg 118
70 trum will make little physical sense. Unfortunately, the EMG (mean 71 kg, SD 7.7 kg). Subjects were recruited from the 119
71 is a time-varying signal, in particular for contraction levels university setting through requests for volunteers and 120
72 higher than 50% of maximum voluntary contraction (MVC) represented a population of recreationally active young 121
73 [12]. Hence time–frequency methods such as STFT, Wigner– adults, none of which were professional athletes. All 122
74 Ville distribution (WVD) and continuous wavelet transform subjects gave their informed consent to participate in the 123
75 (CWT) can be used for time-varying (non-stationary) signal study. No subject had known symptoms of neuromuscular 124
76 analysis. In this paper, we propose a method via time– disorders. EMG signals were obtained from the right biceps 125
77 frequency resolution to validate the efficiency of the variable brachii in an isometric constant force experiment. The 126
78 to enumerate EMG manifestation of muscle fatigue. These participants performed cyclic contractions of the dominant 127
79 methods overcome the problems of the conventional Fourier arm, holding a 4 kg dumbbell corresponding to 15% of 128
80 spectral variables deriving method by avoiding the spectral their MVC (1 s raising the weight and 1 s bringing down 129
81 estimation. Since it does not require any quasi-stationarity the weight). The duration of the test was limited to 2 min, 130
82 and linear assumptions, time–frequency methods are suitable but subjects were allowed to stop earlier when exhausted. 131

F
83 for non-stationary signals. The methods supply a compact Thirteen subjects performed the task for 2 min. Only one 132
84 and physically meaningful representation of EMG signal [6]. subject had to stop the contraction earlier due to exhaustion. 133

O
85 In the analysis of an upper limb repetitive task, it is Bipolar surface EMG (with BIOPAC MP150 system, 134

O
86 imperative to express muscular fatigue quantitatively. This EMG100C, Gain: 1,000, band-pass: 1–500 Hz; sampling 135
87 makes it possible to differentiate dissimilarity in muscle rate 1,000 Hz) was recorded using surface Ag–AgCl 136

PR
88 fatigue depending on the circumstances of the external load electrodes placed over dominant hand biceps brachia 137
89 both for continuous static loads and for repetitive ones. As a muscle of subjects. The electrode site was initially cleansed 138
90 consequence, it is crucial to consider how that time should be with sterile alcohol pads by applying a sufficient abrasive
D 139
91 determined. The time of sustaining a load imposed by action to lessen the resistance of the skin and consequently 140
92 141
TE
experimental conditions does not take into consideration improve the SNR. All software implementations were done
93 individual capabilities of participants and it can be either too in MATLAB (The MathWorks Inc., Natick, MA). 142
94 short or too long considering the process of muscle fatigue.
EC

95 As a result, it can be assumed that the issues are of particular Time–frequency representations of signals 143
96 meaning in assessing muscle fatigue, that is, changes in the
97 value of a particular EMG signal parameter in time, time The simultaneous mapping of time and frequency can 144
R

98 after which muscle fatigue takes place. Thus, muscle fatigue never be completely achieved due to the uncertainty 145
R

99 should be assessed with an indicator enabling quantitative principle. Hence, Fourier transform methods are not 146
100 assessment of muscle fatigue [4]. usually suitable approach in the analysis of signals with 147
O

101 The aim of this study was to distinguish fatigue of upper transient components [13]. To capture time and frequency 148
C

102 limb muscles, which involves a function describing changes information, a special type of signal representation is 149
103 in the EMG signal parameter in time and frequency, using needed, which creates a function of two variables, time 150
N

104 artificial neural networks (ANN). The proposed methodology and frequency, that captures transient effects in the signal. 151
U

105 consists of three steps. In the first step, the EMG signal is A more comprehensive representation of the signal can be 152
106 preprocessed using time–frequency methods to extract fea- obtained by using time–frequency representations. The 153
107 tures from EMG. In the second step, the dimension of time–frequency methods have turn out to be a powerful 154
108 extracted features is reduced by using independent component alternative for the analysis of transient signals. The time– 155
109 analysis (ICA). In the third step, an unknown EMG signal is frequency has variable time–frequency resolution over the 156
110 classified as fresh or fatigued using multi-layer perceptron time–frequency plane by providing good time resolution at 157
111 neural network (MLPNN) with Levenberg–Marquardt (L-M) high frequency and good frequency resolution at low 158
112 and gradient descent (GDA) algorithms. frequencies [14]. 159

Short-term Fourier transform: The spectrogram 160

113 Material and methods
The short-time Fourier transform (STFT) is a popular 161
114 Experimental design method of analysing non-stationary signals since it is 162
simple and computationally efficient and yields reliable 163
115 Fourteen healthy, right-hand dominant volunteers (mean time–frequency plots for slowly varying signals. The major 164
116 age 23.3, S.D. 1.6 years), participated in the experiment. drawback is that there is a compromise between time and 165
117 The height of the subjects varied from 167 to 183 cm (mean frequency resolution of the decomposition. For signals 166
AUTHOR'S PROOF JrnlID 10916_ArtID 9292_Proof# 1 - 15/04/2009

J Med Syst

167 containing transient features it may be not possible to Ville distribution can be defined from the time domain 214
168 handle this compromise efficiently [15]. representation of the signal: 215
169 The basic equation for the STFT in the continuous R


1 

170 domain is: WVDX ðt; f Þ ¼ x t þ 2t :x t
2t  e
2j:pf :t  d t ð3Þ

1
R1
j2pft


XSTFT ðt; f Þ ¼ xðt Þ:wðt
t Þe dt ð1Þ where Rx ðt; t Þ ¼ x t þ 2t  x t
2t is the instantaneous 216
217

1
autocorrelation function and * indicates conjugate opera- 218
171
172 where w(τ−t) is the window function and t is the variable tion [16]. This distribution formula satisfies a large number 219
173 that slides the window across the waveform, x(t). There of desirable mathematical properties. In particular the WVD 220
174 are two major problems with the spectrogram: (1) is always real valued and it preserves time and frequency 221
175 selecting an optimal window length for data segments shifts and satisfies the marginal properties. In the last 222
176 which include numerous different features may not be decades, many researches have been made at the efficient 223
177 possible, and (2) the time–frequency tradeoff: shortening suppression of crossterms and enhancement of the frequency 224
178 the data length, N, to get better time resolution will resolution, maintaining the required properties of time– 225

F
179 decrease frequency resolution which is approximately frequency energy distribution. Although the Wigner–Ville 226
1/(NTs). Shortening the data segment could also result in distribution has the merits of a high resolution in both time

O
180 227
181 the loss of low frequencies that are no longer fully and frequency but the interfering terms produced by the 228

O
182 included in the data segment. Consequently, if the window interaction of two signal components make it hard to 229
183 is made smaller to improve the time resolution, then the interpret the result. Hence it is better to reduce the number 230

PR
184 frequency resolution is degraded and vice versa. This time– and the amplitude of the interference. If a time smoothing 231
185 frequency tradeoff has been equated to an uncertainty window g(t) and a frequency smoothing window h(t) are 232
186 principle where the product of frequency resolution applied then the WVD becomes the smoothed-pseudo-
D 233
187 (expressed as bandwidth, B) and time, T, must be greater Wigner–Ville distribution (SPWVD) which is written as 234
RR
TE
188 than some minimum. Specifically:
SPWVDx ðt; f Þ ¼ hðt
t Þgð f
xÞW x ðt; xÞ d t d x ð4Þ
1 where W is the Wigner Ville distribution. The previous 235
236
BT  ð2Þ
EC

4p compromise of the spectrogram between time and frequency 237

resolutions is now replaced by a compromise between the 238
191
189
190 The trade-off between time and frequency resolution joint time–frequency resolution and the level of the 239
R

192 inherent in the STFT, or spectrogram, has motivated a interference terms. That is, if you increase the smoothing in 240
R

193 number of other time–frequency methods as well as the time and/or frequency you will get a poorer resolution in 241
194 time-scale approaches. In spite of these types of limi- time and/or frequency [15–20]. 242
O

195 tations, the STFT has been used effectively in a wide The WVD has a number of limitations. Most serious of 243
C

196 range of problems, mainly those where only high these is the production of cross products: the demonstration of 244
197 frequency components are of interest and frequency energies at time–frequency values where they do not exist. 245
N

198 resolution is not crucial [16]. These phantom energies have been the most important 246
U

motivator for the development of other distributions that 247

199 The Wigner–Ville distribution apply various filters to the instantaneous autocorrelation 248
function to mitigate the damage done by the cross products. 249
200 Different kind of approaches has been developed to Besides, the Wigner–Ville distribution can have negative 250
201 eliminate some of the limitations of the spectrogram. The regions that have no meaning. The Wigner–Ville distribution 251
202 first of these was the Wigner–Ville distribution (WVD) also has poor noise properties. In essence the noise is 252
203 which is also one of the most studied and best understood distributed across all time and frequency including cross 253
204 of the many time–frequency methods. The approach was products of the noise, even if in some cases, the cross products 254
205 actually developed by Wigner for use in physics, but later and noise influences can be reduced by using a window. In 255
206 applied to signal processing by Ville, hence the dual name. such situations, the required window function is applied to the 256
207 The Wigner–Ville distribution is a special case of a wide lag dimension of the instant autocorrelation function similar to 257
208 variety of similar transformations known under the heading the way it was applied to the time function [16]. 258
209 of Cohen’s class of distributions. For an extensive summary
210 of these distributions see [16–17]. The continuous wavelet transform 259
211 The Wigner–Ville distribution uses an approach that
212 harkens back to the early use of the autocorrelation function Wavelets handle the time–frequency resolution compromise 260
213 for calculating the power spectrum. Hence the Wigner– in a different way to the STFT. Rather than having uniform 261
AUTHOR'S PROOF
JrnlID 10916_ArtID 9292_Proof# 1 - 15/04/2009

J Med Syst

262 time and frequency resolution across the time–frequency carried out over a range of wavelet sizes: although its shape 309
263 plane, wavelets present good time resolution at high remains the same, the time span of the wavelet varies. Since 310
264 frequencies and good frequency resolution at low frequencies. the net area of a wavelet is always zero by design, a 311
265 This property can be extremely valuable in the detection waveform that is constant over the length of the wavelet 312
266 of short-time transients, such as high-frequency waves. would give rise to zero coefficients. Wavelet coefficients 313
267 In other cases a multi-resolution decomposition is not respond to changes in the waveform, more robustly to 314
268 suitable for representing a signal, particularly when good changes on the same scale as the wavelet, and most 315
269 frequency (or time) resolution is required across the strongly, to changes that is similar to the wavelet. Even 316
270 frequency range [16]. though a redundant transformation, it is often easier to 317
271 The basic idea underlying wavelet analysis consists of analyze or recognize patterns using the CWT [16, 21]. 318
272 expressing a signal as a linear combination of a particular
273 set of functions, obtained by shifting and dilating one single Independent component analysis (ICA) 319
274 function called a mother wavelet. Then the correlation
275 between the resulting wavelet shape and the signal is Makeig et al. [22] presented the first applications of ICA to 320

F
276 calculated. This coefficient is a measure of how much of biomedical time series analysis. The significance of using ICA 321
277 the wavelet at that particular scale and time point is to distinguish a source, xi, from mixtures, yi, is that the activity 322

O
278 included in the signal. The translated and dilated wavelet of each source is statistically independent of the other sources 323

O
279 is defined in terms of the mother wavelet as [23]; i.e., the mutual information between any two sources, xi 324
  and xj, is zero. General ICA model can be written as 325

PR
1 t
b
y a;b ðt Þ ¼ pffiffiffiffiffiffi y ð5Þ Y ¼ AX; ð7Þ
j aj a
where A is an unknown full-rank matrix, called the mixing
D 326
327
280
281 for positive a. Conventionally, a is 1 for the mother wavelet matrix, and X is the independent component (IC) data matrix, 328
282 and increasing a>1 dilates the wavelet, expanding the
TE
and Y is the measured variable data matrix. The crucial 329
283 interval over which it takes non-zero values. The factor p1ffiffiffiffi problem of ICA is to estimate the independent component 330
jaj
284 ensures that all of the scaled wavelets have a squared norm matrix X or to estimate the mixing matrix A from the 331
EC

285 of 1 so the wavelet transforms at different scales are directly measured data matrix Y without any knowledge of X or A. 332
286 comparable. When the wavelet is dilated for large scales its The practical problem of ICA is to calculate a separating 333
287 scale content becomes more localised except the time matrix W so that the components of the reconstructed data 334
R

288 period it covers becomes less localised. This is why matrix Xb become as independent of each other as possible, 335
R

289 wavelets have good time resolution and poor scale given as 336
290
O

resolution for low scales, but poor time resolution and

291 good scale resolution at high scales. The wavelet operates b ¼ WY
X ð8Þ
C

292 as a smoothing function in the time and scale domains in

293 The requirement of independence of components X b is 339
337
338
N

the same way as windows smooth the STFT. The

294 continuous wavelet transform (CWT) of a signal x(t) is equivalent to the assumption of their non-Gaussian distribu- 340
U

295 formally defined at time t and scale a by the formula tion. As a result, the independent components are evaluated 341
Z1   optimizing some measure of the non-Gaussianity of the 342
1 t
b vector WY. In this work, a fast fixed-point algorithm [24, 25] 343
W ða; bÞ ¼ xðt Þ pffiffiffi y d t; ð6Þ
a a is used to minimize or maximize the fourth order cumulant to 344

1
perform ICA (available in The FastICA MATLAB Package 345
296
297 where b acts to translate the function across x(t) just as t [26]). Since statistical independence is more restrictive than 346
298 does in the equations above [15]. The wavelet coefficients, uncorrelation, the measured variables are first transformed 347
299 W(a, b), express the correlation between the waveform and into uncorrelated variables with unit variances. This pretreat- 348
300 the wavelet at a variety of translations and scales: the ment is called as sphering or prewhitening [27, 28]. 349
301 similarity between the waveform and the wavelet at a given
302 combination of scale and position, a, b. In other words, the Artificial neural network models 350
303 coefficients present the amplitudes of a series of wavelets,
304 over a range of scales and translations, that should be added Artificial neural networks (ANNs) are computing systems 351
305 together to reconstruct the original signal. From this point made up of large number of firmly interconnected adaptive 352
306 of view, wavelet analysis can be thought of as a search over processing elements (neurons) that are able to perform 353
307 the waveform of interest for activity that most obviously massively parallel computations for data processing and 354
308 approximates the shape of the wavelet. This search is knowledge representation. ANNs can be trained to recog- 355
AUTHOR'S PROOF JrnlID 10916_ArtID 9292_Proof# 1 - 15/04/2009

J Med Syst

356 nize patterns and the nonlinear models developed during (GDA) training algorithms were used. The L-M algorithm 379
357 training allow neural networks to generalize their conclu- combines the best features of the Gauss–Newton technique 380
358 sions and to make application to patterns not previously and the steepest-descent algorithm, but avoids many of their 381
359 encountered. There are many different types and architec- limitations. In particular, it generally does not suffer from the 382
360 tures of neural networks varying fundamentally in the way problem of slow convergence [32]. Effective training 383
361 they learn; the details of which are well documented in the algorithm and better-understood system behavior are the 384
362 literature [29–31]. In this paper, neural network relevant to advantages of this type of neural network. Selection of 385
363 the application being considered will be employed for network input parameters and performance of classifier are 386
364 designing classifiers; namely the MLPNN. The architecture important in muscle fatigue detection. The efficiency of this 387
365 of MLPNN may contain two or more layers. A simple two- technique can be explained by using the result of experi- 388
366 layer ANN consists only of an input layer containing the ments [33]. This paper clearly demonstrates that our method 389
367 input variables to the problem, and output layer containing is applicable for detecting muscle fatigue. The qualities of 390
368 the solution of the problem. This type of networks is a the method are that it is simple to apply, and it does not 391
369 satisfactory approximator for linear problems. However, for require high computation power. The method can be used as 392

F
370 approximating nonlinear systems, additional intermediate a standalone tool, but it can be implemented as a building 393
371 (hidden) processing layers are employed to handle the block for computer-assisted EMG fatigue detection. 394

O
372 problem's nonlinearity and complexity. Although it depends

O
373 on complexity of the function or the process being
374 modeled, one hidden layer may be sufficient to map an Results and discussion 395

PR
375 arbitrary function to any degree of accuracy. Hence, three-
376 layer architecture ANNs was adopted for the present study. In this study, we propose the use of the Time–Frequency 396
377 For solving pattern classification problem MLPNN methods with ICA as an alternative method for the
D 397
378 employing Levenberg–Marquardt (L-M) and gradient descent classification of the surface electromyography signal for 398
TE

a b
EC

0.5

0.4
R

0.3
Amplitude (V)

0.2
O

0.1

0
C

-0.1
N

-0.2
U

0 50 100 150 200 250

Time(msec)

c d

Fig. 1 a Fresh EMG signal, b STFT, c SPWVD, d CWT of the signal

AUTHOR'S PROOF
JrnlID 10916_ArtID 9292_Proof# 1 - 15/04/2009

J Med Syst

399 studying local muscle fatigue during sustained isometric taken for how to divide data into a training set and a test 422
400 constant force muscle contractions. The time-varying set. In this study, 55% of overall data were used for training 423
401 characteristic of the method enables us to accommodate and the rest of them (45% of overall data) were used for 424
402 non-stationary EMG data in higher-level contraction. The testing. The objective of the modelling phase in this 425
403 reliability and accuracy of the time–frequency methods are application was to develop classifiers that are able to 426
404 compared in terms of their robustness in determination of identify any input combination as belonging to either one of 427
405 muscle fatigue. The results suggest that time–frequency the two classes: fresh or fatigue. For developing the neural 428
406 methods are better choice for the spectral estimation network classifiers, 600 examples were randomly taken 429
407 process commonly used to quantify surface EMG signal from the 1,100 examples and used for training the neural 430
408 manifestations of muscle fatigue. Although much more networks. The remaining 500 examples were kept aside and 431
409 sophisticated classification methods for muscle fatigue detec- used for testing the validity of the developed models. The 432
410 tion exist according to the results obtained, we considered our class distribution of the samples in the training, validation 433
411 choice adequate for a given problem. The efficiency of this and test data set is summarized in Table 1. 434
412 technique can be explained by using the result of experiments. The MLPNN was designed with different features of EMG 435

F
413 In this study, EMG recordings were preprocessed by using signal in the input layer; and the output layer consisted of one 436
414 STFT, SPWVD and CWT (Figs. 1 and 2); ICA is used for node representing whether fatigue detected or not. A value of 437

O
415 dimension reduction. Then, these signals are applied to “0” was used when the experimental investigation indicated a 438

O
416 MLPNN. fresh and “1” for fatigued muscle. Depending on the nature of 439
the method, we used different nodes in the input layer for each 440

PR
417 Development of ANN model method (STFT, SPWVD and CWT). We used 150 nodes for 441
STFT, 160 for SPWVD and 175 for CWT. The preliminary 442
418 The adequate functioning of MLPNN depends on the sizes architecture of the network was examined using one and two
D 443
419 of the training set and test set. During training, the input hidden layers with a variable number of hidden nodes in each. 444
420 445
TE
and desired data will be repeatedly presented to the It was found that one hidden layer is adequate for the problem
421 network. When using neural network, decision must be at hand. Thus the sought network will contain three layers of 446
EC

a b
0.6
R

0.4
R
Amplitude (V)

0.2
O

0
C

-0.2
N
U

-0.4

0 50 100 150 200 250

Time(msec)

c d

Fig. 2 a Fatigued EMG signal, b STFT, c SPWVD, d CWT of the signal

 AUTHOR'S PROOF JrnlID 10916_ArtID 9292_Proof# 1 - 15/04/2009

J Med Syst

t1.1 Table 1 Class distribution of the samples in the training and test data Table 2 Comparison of time–frequency methods for muscle fatigue t2.1
sets detection using L-M algorithm

t1.2 Class Training set Test set Total Preprocessing method Accuracy Specificity Sensitivity t2.2

t1.3 Normal 300 250 550 STFT (%) 88.5 87 90 t2.3

t1.4 Fatigued 300 250 550 SPWVD (%) 90 87 93 t2.4
t1.5 Total 600 500 1,100 CWT (%) 91 91.5 90.5 t2.5

447 nodes. The training procedure started with one hidden node in Experimental results 486
448 the hidden layer, followed by training on the training data, and
449 then by testing on the validation data to examine the network’s In this paper, different techniques based on time–frequency 487
450 prediction performance on cases never used in its develop- methods and neural networks are proposed for detecting 488

F
451 ment. Then, the same procedure was run repeatedly each time muscle fatigue. Conventional techniques related to frequency 489
452 the network was expanded by adding one more node to the and amplitude analysis do not work in this case. The signal 490

O
453 hidden layer, until the best architecture and set of connection amplitude for all subjects is falling near the end of the test 491

O
454 weights were obtained. Using the backpropagation (L-M and showing that the subjects are putting less force. This shows 492
455 GDA) algorithm for training, a training rate of 0.001 and that while fatigue is obviously present according to the 493

PR
456 momentum coefficient of 0.9 was found optimum for subjects’ own experience it cannot be detected due to the 494
457 training the network with various topologies. The selection condition of constant force being not applicable. Time– 495
458 of the optimal network was based on monitoring the frequency based decomposition is used to isolate muscle 496
459 variation of error and some accuracy parameters as the
D
activity from fourteen volunteer subjects related to the same 497
460 498
TE
network was expanded in the hidden layer size and for phasic voluntary movements. The proposed method for
461 each training cycle. The sum of squares of error represent- detecting fatigue by observing the behaviour of the time– 499
462 ing the sum of square of deviations of MLPNN solution frequency spectrum with ICA can be automated using neural 500
EC

463 (output) from the true (target) values for both the training networks. An MLPNN has been used to visualize the variation 501
464 and test sets was used for selecting the optimal network. of the time–frequency methods and aid the detection of 502
465 The optimum network configuration found was 65 neurons fatigue. The creation of the input dataset for the MLPNN has 503
R

466 for the hidden layer. In the hidden layer sigmoidal function been performed by creating vectors whose elements are 504
R

467 and in the output layer linear function was used. extracted by using STFT, SPWVD and CWT. 505
The training set provided to the MLPNN was represen- 506
O

468 Evaluation of performance tative of the whole space of concern so that the trained 507
C

MLPNN had the ability of generalization. After training 508

469 Additionally, because the problem involves classification MLPNN, it was determined that the network adequately 509
N

470 into two classes, accuracy, sensitivity and specificity were classified data. We achieved a classification rate of higher 510
U

471 used as a performance measure. These parameters were than 90% by using artificial neural network as a classifier. 511
472 obtained separately for both the training and testing sets Depending on output neuron had a value of “0” or “1”, the 512
473 each time a new network topology was examined. EMG recording was classified as fresh and fatigued. 513
474 Computer programs that we have written for the training Tables 2 (for L-M algorithm) and 3 (for GDA algorithm) 514
475 algorithm based on backpropagation of error, L-M and show a summary of the performance measures. It is obvious 515
476 GDA were used to develop the MLPNNs. The test from Table 2 that, by using L-M algorithm, the highest 516
477 performance of the MLPNN was determined by the success rate was obtained for the non-fatigued group 517
478 computation of the following statistical parameters:

479 Specificity Number of correctly classified non-fatigued Table 3 Comparison of time–frequency methods for muscle fatigue t3.1
480 subjects/number of total non-fatigued subjects. detection using GDA algorithm

481 Sensitivity Number of correctly classified fatigued subjects/ Preprocessing method Accuracy Specificity Sensitivity t3.2
482 number of total fatigued subjects. STFT (%) 87.5 85 90 t3.3
SPWVD (%) 89 86 92 t3.4
483 Total classification accuracy Number of correctly classified
CWT (%) 89 88 90 t3.5
484
485 subjects/number of total subjects.
AUTHOR'S PROOF
JrnlID 10916_ArtID 9292_Proof# 1 - 15/04/2009

J Med Syst

518 (sensitivity) using SPWVD with ICA (93%). Also it can be network that classifies subjects as having fresh or fatigued 566
519 seen from Table 2 the wavelet transform method has the muscle provides a valuable diagnostic decision support tool 567
520 best classification accuracy. The average classification for muscle fatigue detection. In this study, time–frequency 568
521 success rate for CWT was 91%, for SPWVD was 90% methods have been used as a feature extraction method and 569
522 and for STFT was 88.5%. On the other hand, by using ICA has been used to reduce the dimension of feature 570
523 GDA algorithm, the highest success rate was obtained for vectors. Then extracted features of EMG signals have been 571
524 the non-fatigued group (sensitivity) using SPWVD with used as an input to MLPNN that could be used to detect 572
525 ICA (92%). The average classification success rate for muscle fatigue. This process is realized by online data 573
526 CWT and SPWVD was 89% and for STFT was 87.5%. acquisition system. Depending on these methods, classifiers 574
have been developed and trained. The success and 575
527 Discussion performance of the time–frequency methods is proven to 576
detect fatigue under dynamic conditions. The proposed 577
528 The aim of this study is to validate the use of wavelets in techniques can be used to detect the presence of fatigue and 578
529 detecting muscle fatigue in comparison with the SPWVD act against it at an early stage, possibly by adjusting 579

F
530 and STFT. In order to examine the effect of time–frequency comfort related parameters that can affect the presence of 580
531 methods on classifications efficiency, tests were carried out fatigue. With specificity and sensitivity values both above 581

O
532 using ICA for dimensionality reduction. Average efficiency 90%, the MLPNN classification with L-M algorithm may 582

O
533 obtained for each method when EMG signals were be used as an important decision support system in the 583
534 classified using MLPNN structures. muscle fatigue detection. 584

PR
535 In recent years, number of studies [8–12] has used time– 585
536 frequency methods to process non-stationary EMG signals. Acknowledgement This research has been supported by the 586
Scientific & Technological Research Council of Turkey (TUBITAK 587
537 The Wigner–Ville has a number of advantages over the Project no: 105E039).
D 588
538 STFT, but also has a number of limitations. It produces
539
TE
extremely good picture of the time–frequency structure
540 [14]. It also has favourable marginals and conditional
References 589
541 moments. The marginals narrate the summation over time
EC

542 or frequency to the signal energy at that time or frequency

1. Mathiassen, S. E., The influence of exercise/rest schedule on the 590
543 [16]. Unlike the STFT, the CWT does not assume signal physiological and psychophysical response to isometric shoulder– 591
544 stationarity and uses different length windows (wavelets) to neck exercise. Eur. J. Appl. Physiol. 67:528–539, 1993. 592
R

545 accomplish a multiresolution analysis. For instance, the doi:10.1007/BF00241650. 593

594
R

2. Barnahart, S., Demers, P. A., Miller, M., Longstreth, W. T., and

546 CWT estimates the amount of power at each possible
Rosenstock, L., Carpal tunnel syndrome among ski manufacturing 595
547 frequency for each time instant, resulting in a time–
O

workers. Scand. J. Work Environ. Health. 17:46–52, 1991. 596

548 frequency representation of the signal. The CWT presents 3. Nussbaum, M., Clarc, L. L., Lanza, M. A., and Rice, K. M., 597
C

549 good resolution in both time and frequency, thus making it Fatigue and endurance limits during intermittent overhead work. 598
Am. Ind. Hyg. Assoc. J. 62:446–456, 2001. doi:10.1202/0002- 599
550 useful for non-stationary signal processing [8, 35, 36].
N

8894(2001)062<0446:FAELDI>2.0.CO;2. 600
551 These findings suggested that STFT, SPWVD and CWT 4. Roman-Liu, D., Tokarski, T., and Wojcik, K., Quantitative 601
U

552 methods give almost similar information regarding the physio- assessment of upper limb muscle fatigue depending on the 602
553 logical mechanisms underlying fatigue during dynamic muscle conditions of repetitive task load.. J. Electromyogr. Kinesiol. 603
14:671–682, 2004. doi:10.1016/j.jelekin.2004.04.002. 604
554 actions. Consequently, time–frequency methods are acceptable
5. Moshou, D., Hostens, I., Papaioannou, G., and Ramon, H., Dynamic 605
555 for determining the muscle fatigue in EMG. Although there muscle fatigue detection using self-organizing maps.. Appl. Soft 606
556 were small differences between the results, the testing Comput. 5:391–398, 2005. doi:10.1016/j.asoc.2004.09.001. 607
557 performance of the neural network classification systems are 6. Xie, H., and Wang, Z., Mean frequency derived via Hilbert– 608
Huang transform with application to fatigue EMG signal analysis. 609
558 found to be satisfactory for both L-M and GDA algorithms.
Comput. Methods Programs Biomed. 8 (2)114–120, 2006. 610
559 Hence, we think that this system can be used in muscle fatigue doi:10.1016/j.cmpb.2006.02.009. 611
560 detection in the future after it is developed. This application 7. Hedayatpour, N., Arendt-Nielsen, L., and Farina, D., Non-uniform 612
561 brings objectivity to the evaluation of EMG signals and its electromyographic activity during fatigue and recovery of the 613
vastus medialis and lateralis muscles.. J. Electromyogr. Kinesiol. 614
562 automated nature makes it easy to be used in practical way.
18 (3)390–396, 2008. doi:10.1016/j.jelekin.2006.12.004. 615
8. Beck, T. W., Housh, T. J., Johnson, G. O., Weir, J. P., Cramer, J. 616
T., Coburn, J. W., and Malek, M. H., Comparison of Fourier and 617
563 Conclusions wavelet transform procedures for examining the mechanomyo- 618
graphic and electromyographic frequency domain responses 619
during fatiguing isokinetic muscle actions of the biceps brachii.. 620
564 Determination of muscle fatigue is a difficult task requiring J. Electromyogr. Kinesiol. 15:190–199, 2005. doi:10.1016/j. 621
565 examination of the muscle, an EMG. An artificial neural jelekin.2004.08.007. 622
AUTHOR'S PROOF JrnlID 10916_ArtID 9292_Proof# 1 - 15/04/2009

J Med Syst

623 9. Hostens, I., Seghers, J., Spaepen, A., and Ramon, H., Validation logr. Clin. Neurophysiol. 86:23–35, 1993. doi:10.1016/0013-4694 670
624 of the wavelet spectral estimation technique in Biceps Brachii and (93)90064-3. 671
625 Brachioradialis fatigue assessment during prolonged low-level 23. Lee, T. W., Girolami, M., and Sejnowski, T. J., Independent 672
626 static and dynamic contractions. J. Electromyogr. Kinesiol. component analysis using an extended infomax algorithm for 673
627 14:205–215, 2004. doi:10.1016/S1050-6411(03)00101-9. mixed sub-Gaussian and super-Gaussian sources. Neural Comput. 674
628 10. Kumar, D. K., Pah, N. D., and Bradley, A., Wavelet Analysis of 11:606–633, 1999. doi:10.1162/089976699300016719. 675
629 Surface Electromyography to Determine Muscle Fatigue. IEEE 24. Hyvarinen, A., Fast and robust fixed-point algorithms for 676
630 Trans. Neural Syst. Rehabil. Eng. 11 (4)400–406, 2003. independent component analysis. IEEE Trans. Neural Netw. 677
631 doi:10.1109/TNSRE.2003.819901. 10:626–634, 1999. doi:10.1109/72.761722. 678
632 11. Linssen, W. H. J. P., Stegeman, D. F., Joosten, E. M. G., van horst, M. 25. Hyvarinen, A., and Oja, E., Independent component analysis: 679
633 A., Binkhorst, R. A., and Notermans, S. L. H., Variability and algorithms and applications. Neural Netw. 13 (4–5)411–430, 680
634 interrelationship of surface EMG parameters during local muscle 2000. doi:10.1016/S0893-6080(00)00026-5. 681
635 fatigue. Muscle Nerve. 16 (8)849–856, 1993. doi:10.1002/mus. 26. The FastICA MATLAB Package: Available from http://www.cis. 682
636 880160808. hut.fi/projects/ica/fastica. 683
637 12. Karlsson, S., Yu, J., and Akay, M., Enhancement of spectral 27. Widodo, A., and Yang, B. S., Application of nonlinear feature 684
638 analysis of myoelectric signals during static contractions using extraction and support vector machines for fault diagnosis of 685
639 wavelet methods.. IEEE Trans. Biomed. Eng. 46:670–684, 1999. induction motors. Expert Syst. Appl. 33:241–250, 2007. 686
640 doi:10.1109/10.764944. doi:10.1016/j.eswa.2006.04.020. 687

F
641 13. Akay, M., Introduction: wavelets in biomedical engineering. Ann. 28. Widodo, A., Yang, B. S., and Han, T., Combination of 688
642 689

O
Biomed. Eng. 23:531–542, 1995. doi:10.1007/BF02584453. independent component analysis and support vector machines
643 14. Cohen, L., Time–frequency distributions: a review. Proc. IEEE. for intelligent faults diagnosis of induction motors. Expert Syst. 690
644 691

O
77:941–981, 1989. doi:10.1109/5.30749. Appl. 32:299–312, 2007. doi:10.1016/j.eswa.2005.11.031.
645 15. Marchant, B. P., Time–frequency analysis for biosystem engineer- 29. Haykin, S., Neural networks: a comprehensive foundation. 692
646 693

PR
ing. Biosystems Eng. 85 (3)261–281, 2003. doi:10.1016/S1537- Macmillan, New York, 1994.
647 5110(03)00063-1. 30. Basheer, I. A., and Hajmeer, M., Artificial neural networks: 694
648 16. Semmlow, J. L., Biosignal and biomedical image processing, fundamentals, computing, design, and application. J. Microbiol. 695
649 MATLAB-based applications. Marcel Dekker, New York, 2004. Methods. 43:3–31, 2000. doi:10.1016/S0167-7012(00)00201-3. 696
650 17. Boudreaux-Bartels, G. F., and Murry, R., Time–frequency signal 31.
D Chaudhuri, B. B., and Bhattacharya, U., Efficient training and 697
651 representations for biomedical signals. In: Bronzino , J. (Ed.), The improved performance of multilayer perceptron in pattern classi- 698
652 699
TE
biomedical engineering handbookCRC, Boca Raton, 1995. fication. Neurocomputing. 34:11–27, 2000. doi:10.1016/S0925-
653 18. Neto, E. P. S., Custaud, M. A., Frutoso, J., Somody, L., Gharib, 2312(00)00305-2. 700
654 C., and Fortrat, J. O., Smoothed pseudo Wigner–Ville distribution 32. Hagan, M. T., and Menhaj, M. B., Training feedforward networks 701
655 as an alternative to Fourier transform in rats.. Auton. Neurosci. with the Marquardt algorithm. IEEE Trans. Neural Netw. 5 (6) 702
EC

656 Basic Clin. 87:258–267, 2001. doi:10.1016/S1566-0702(00) 989–993, 1994. doi:10.1109/72.329697. 703
657 00211-3. 33. Subasi, A., and Ercelebi, E., Classification of EEG signals using 704
658 19. Narasimhan, S. V., and Nayak, M. B., Improved Wigner–Ville neural network and logistic regression. Comput. Methods Programs 705
R

659 distribution performance by signal decomposition and modified Biomed. 78 (2)87–99, 2005. doi:10.1016/j.cmpb.2004.10.009. 706
660 group delay. Signal Processing. 83:2523–2538, 2003. 34. Bonato, P., Gagliati, G., and Knaflitz, M., Analysis of myoelectric 707 Q1
R

661 doi:10.1016/j.sigpro.2003.07.011. signals recorded during dynamic contractions.. IEEE Eng. Med. 708
662 709
O

20. Drai, R., Khelil, M., and Benchaala, A., Time frequency and Biol. 15:102–111, 1996. doi:10.1109/51.544517.
663 wavelet transform applied to selected problems in ultrasonics. 35. Karlsson, J. S., Gerdle, B., and Akay, M., Analyzing surface 710
664 711
C

NDE. NDT Int. 35:567–572, 2002. doi:10.1016/S0963-8695(02) myoelectric signals recorded during isokinetic contractions. IEEE
665 00041-5. Eng. Med. Biol. 20:97–105, 2001. doi:10.1109/51.982281. 712
N

666 21. Rioul, O., Vetterli, M., Wavelets and signal processing., IEEE SP 36. Karlsson, S., Yu, J., and Akay, M., Time–frequency analysis of 713
667 Magazine, 14–38. 1991. myoelectric signals during dynamic contractions: a comparative 714
U

668 22. Makeig, S., and Inlow, M., Lapses in alertness: coherence of study. IEEE Trans. Biomed. Eng. 47:228–238, 2000. doi:10.1109/ 715
669 fluctuations in performance and EEG spectrum. Electroencepha- 10.821766. 716
AUTHOR'S PROOF
AUTHOR QUERY

AUTHOR PLEASE ANSWER QUERY.

Q1. This reference was not cited anywhere in the article. Please provide a citation or, alternatively,
delete this item from the list.

F
O
O
PR
D
TE
EC
R
R
O
C
N
U

View publication stats