Computer Methods and Programs in Biomedicine 154 (2018) 45–56

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

journal homepage: www.elsevier.com/locate/cmpb

Surface electromyography based muscle fatigue detection using

high-resolution time-frequency methods and machine learning
algorithms
P.A. Karthick a,b,∗, Diptasree Maitra Ghosh a, S. Ramakrishnan a
a
Biomedical Engineering Group, Department of Applied Mechanics, Indian Institute of Technology Madras, Chennai, India
b
Montreal Neurological Institute and Hospital, McGill University, Montreal, Canada

a r t i c l e i n f o a b s t r a c t

Article history: Background and objective: Surface electromyography (sEMG) based muscle fatigue research is widely pre-
Received 16 April 2017 ferred in sports science and occupational/rehabilitation studies due to its noninvasiveness. However, these
Revised 22 September 2017
signals are complex, multicomponent and highly nonstationary with large inter-subject variations, partic-
Accepted 29 October 2017
ularly during dynamic contractions. Hence, time-frequency based machine learning methodologies can
improve the design of automated system for these signals.
Keywords: Methods: In this work, the analysis based on high-resolution time-frequency methods, namely, Stockwell
Muscle fatigue analysis transform (S-transform), B-distribution (BD) and extended modified B-distribution (EMBD) are proposed
Surface electromyography to differentiate the dynamic muscle nonfatigue and fatigue conditions. The nonfatigue and fatigue seg-
S-transform
ments of sEMG signals recorded from the biceps brachii of 52 healthy volunteers are preprocessed and
EMBD
subjected to S-transform, BD and EMBD. Twelve features are extracted from each method and promi-
Time-frequency features
SVM nent features are selected using genetic algorithm (GA) and binary particle swarm optimization (BPSO).
Five machine learning algorithms, namely, naïve Bayes, support vector machine (SVM) of polynomial and
radial basis kernel, random forest and rotation forests are used for the classification.
Results: The results show that all the proposed time-frequency distributions (TFDs) are able to show the
nonstationary variations of sEMG signals. Most of the features exhibit statistically significant difference in
the muscle fatigue and nonfatigue conditions. The maximum number of features (66%) is reduced by GA
and BPSO for EMBD and BD-TFD respectively. The combination of EMBD- polynomial kernel based SVM
is found to be most accurate (91% accuracy) in classifying the conditions with the features selected using
GA.
Conclusions: The proposed methods are found to be capable of handling the nonstationary and multi-
component variations of sEMG signals recorded in dynamic fatiguing contractions. Particularly, the com-
bination of EMBD- polynomial kernel based SVM could be used to detect the dynamic muscle fatigue
conditions.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction (sEMG) signals are the electrical activities recorded over the sur-
face of the skin during muscle contractions. The analysis of these
Bioelectric signals represent physiological functions of various signals finds application in various areas such as clinical diagnosis,
tissues/organs associated with the human body. Research on these prosthetic myoelectric control, ergonomics and sports biomechan-
signals aims at extracting useful biomarkers that reflect the func- ics [1]. Recently, this non-invasive sEMG technique is widely used
tional state of the human system. In the recent years, several ad- to analyze the muscle fatigue [2].
vanced signal processing algorithms and machine learning meth- Muscle fatigue is a neuromuscular condition in which mus-
ods have been reported in the literature. However, the analy- cles fail to generate the required force. It occurs in both normal
ses of these biosignals are still complex due to its inherent ran- and abnormal subjects [1,2]. Neuromuscular fatigue is a most com-
dom and nonstationary variations. The surface electromyography mon symptom in neurological disorders such as multiple sclero-
sis, Parkinson disease and stroke. It is reported that around 60% of
neuromuscular patients suffers from severe fatigue [3]. The detec-
∗
Corresponding author. tion and prevention of fatigue help to improve the performance in
E-mail address: [email protected] (P.A. Karthick).

https://doi.org/10.1016/j.cmpb.2017.10.024
0169-2607/© 2017 Elsevier B.V. All rights reserved.
46 P.A. Karthick et al. / Computer Methods and Programs in Biomedicine 154 (2018) 45–56

sports related scenarios and aids in the financial growth of indus- classification of sEMG signals under various neuromuscular condi-
tries [4,5]. The characteristics of sEMG signals are found to largely tions [21].
vary with fatigue conditions due to metabolic, structural and ener- In this work, sEMG signals are recorded from the biceps brachii
getic changes in muscle [6]. muscle during dynamic fatiguing contractions. The nonfatigue and
The frequency and amplitude characteristics of sEMG signals fatigue segments are subjected to S-transform, BD and EMBD tech-
are random, multicomponent and nonstationary due to the physi- niques and then twelve time-frequency features are extracted from
ological parameters such as motor unit recruitment patterns, firing each TFD. Then the feature selection and machine learning algo-
rate and anisotropic nature of volume conductors. However, these rithms are applied to differentiate muscle nonfatigue and fatigue
signals are considered as stationary in isometric contractions. Sev- conditions. The performance of these TFDs is compared using clas-
eral measures such as root mean square value, zero crossing and sifier performance.
median frequency have also been employed for these sEMG anal-
ysis [1,7]. However, these measures do not account for the time- 2. Material and methods
varying frequency characteristics of sEMG signals [8].
Several researches have been carried out on the sEMG signals 2.1. Material (Experimental protocol and signal acquisition)
detected under isometric contractions. However, dynamic muscle
contractions better represent the real life activities [8]. The sta- The sEMG signals are recorded using Biopac MP36 data acquisi-
tionarity assumption does not hold for the signals recorded un- tion system. The signal to noise ratio and common mode rejection
der dynamic contractions because the frequency components vary ratio of this system are 89 dB and 110 dB respectively. The signals
continuously with time. Changes in muscle force, muscle shape, are acquired at the sampling rate of 10,0 0 0 Hz with the gain of
elbow-joint angle, and movement of the innervation zone with 10 0 0 [22,23].
respect to the recording electrodes results in the high degree of Fifty two untrained healthy subjects with no history of neuro-
nonstationarity [8,9]. Short time Fourier transform (STFT) has been logical and neuromuscular disorders participated in this study. The
widely used to characterize the time varying frequency compo- participants are instructed to take 12 h of complete rest before the
nents of sEMG signals. However, this method suffers from limita- start of experiment. The nature of the study is well explained to
tions such as stationarity requirement and the uncertainty prin- the subjects and informed consent is obtained. This experiment
ciple [2]. These nonstationary signals are typically analyzed using adhere the tenet of Declaration of Helsinki. Two Ag–AgCl disc type
time-scale representation and/or time-frequency distribution tech- disposable electrodes are placed with an inter electrode distance
niques. of 3 cm over the muscle belly after the skin preparation. The dif-
Continuous wavelet transform (CWT) represents the signal in ferential configuration set-up has been used for this study. The ref-
the time-scale space and it has a wide range of application in the erence electrode is placed at the elbow joint [18].
analysis of biomedical signals. It can also be seen as an exten- The subjects are instructed to stand erect on the wooden plat-
sion of STFT. A recently developed time-frequency method called form to prevent electric shocks. Then, the subjects are asked to
Stockwell transform (S-transform) is actually a combination of perform continuous biceps curl exercise with 6 kg dumbbell load
CWT and STFT. S-transform is an effective time-frequency tech- using their dominant hand until they experiences fatigue or un-
nique, which preserves phase information and provides good time- able to continue the exercise. The subjects are asked to maintain
frequency resolution using movable and scalable Gaussian window the curl speed at their comfortable pace. sEMG signals are acquired
[10,11]. Recently, this technique has been utilized for the analysis from the dominant hand for the entire duration of exercise [18,21].
of electroencephalography and electrocardiography signals in order The sEMG signals associated with the muscle nonfatigue and fa-
to track changes in signal amplitude, frequency and phase [12,13]. tigue conditions are extracted from the first and last curl of the ex-
However, the applicability of S-transform has not been explored for ercise. The peak of first and last curl is identified using the moving
the sEMG based muscle fatigue analysis. average root mean square value and a segment of 250 ms is ex-
All time-frequency techniques that belong to Cohen class repre- tracted before and after the peak for the analysis [24]. The motion
sent the signals in time-frequency axis. Choi–Williams and Born– artifacts, power line interference and high frequency noises are re-
Jordan distributions are smoothed version of Wigner-Ville distribu- moved using band pass filter (10–400 Hz) and notch filter (50 Hz)
tion (WVD) which are widely used to analyze the nonstationary [18].
property of the sEMG signals [14–16]. B-distribution (BD), Mod-
ified B-distribution (MBD) and extended modified B-distribution 2.2. Methods
(EMBD) belongs to reduced interference Cohen class TFDs [17]. Re-
cently, BD and MBD based techniques have been used for the sEMG This study consists of computation of time-frequency spectrum,
analysis [18]. EMBD based TFD has been reported to perform bet- feature extraction, feature selection and classification of muscle fa-
ter than MBD in the representation of slow and fast variations of tigue and nonfatigue conditions.
frequency contents associated with heart sound signals [19]. Fur-
ther, this technique has been extensively used to study the mul- 2.2.1. Time-frequency representation of sEMG signals
ticomponent nonstationary nature of new born seizure EEG and Fourier spectrum represents the frequency components present
fetal movement signals [20]. The focus of this work is to intro- in the signals. However, it does not provide information about how
duce these high-resolution time-frequency methods such as EMBD the frequency component varies with time. The time-frequency
and S-transform for the muscle fatigue analysis and to compare the methods show the nonstationary nature of biomedical signals and
classification performances using machine learning techniques. it has become an ideal tool for this analysis [14].
In addition to conventionally used time-frequency features such
as instantaneous mean and median frequency, other attributes 2.2.1.1. Stockwell transform. The Stockwell transform (S-transform)
such as time-frequency based statistical and complexity features addresses the issue of fixed window of STFT. It uses scalable lo-
are also used to provide deeper insight about the muscle mech- calized Gaussian window. The dilation process associated with this
anisms in fatigue condition. Several classification algorithms have transform aids in providing good resolution. The mathematical ex-
been reported for the differentiation of muscle nonfatigue and fa- pression of S-transform for a signal x(t) is defined as [11]
tigue conditions. The K-nearest neighbor, naïve Bayes and genetic ∞
algorithm based support vector machine have been used for the S (t, f ) = ∫ x(τ )g(t − τ , σ )e− j2π f τ dτ (1)
−∞
 P.A. Karthick et al. / Computer Methods and Programs in Biomedicine 154 (2018) 45–56 47

where g(t − τ , σ ) is a scaled Gaussian window. It can be visualized When the signal is represented in ambiguity domain, it can be
that this is a local spectrum calculated at time t = τ , obtained by seen that the actual frequency (auto terms) components will be
multiplying the given signal x(t) with a Guassian window concen- concentrated around the origin and the cross-terms will be located
trated at t = τ and taking the Fourier transform of the product. The away from the origin. When g(v, τ ) = 1, or g(v, τ ) is a all-pass fil-
scalable Gaussian window is defined as ter, and then the Eq. (8) became Wigner-Ville distribution which
1 −(t−τ )2 is a fundamental TFD of Cohen class family. In this scenario, the
g(t − τ , σ ) = √ e 2σ 2 (2) resultant TFD presents almost all the cross-terms generated due to
σ 2π
the multicomponent nature of the signals and thereby it leads to
By substituting σ = |1f | then the Eq. (2) becomes the misunderstanding about the real phenomena. In order to re-
| f | −(t−τ2)2 f 2 duce these cross-terms, the two dimensional low pass filter is de-
g(t − τ , 1/ f ) = √ e (3) veloped in ambiguity domain.
2π
The general form of Cohen class TFD in discrete domain is given
When the f increases, the Gaussian window g(t − τ , σ ) becomes by
more localized on t due to the compression. The time integral of 
ST gives the frequency spectrum of the signal and it is expressed ρz [n, k] = 2 G[n, m]∗n (z[n + m]∗z∗ [n − m] )e− j2π km/M (10)
as |m| <M
2

∞
X ( f ) = ∫ S (t, f )dt (4) where z[n] and G[n, m] is a discrete version of z[n] and G(t, τ )
−∞ respectively.
Then the inverse S-transform can be obtained by taking inverse
Fourier transform of Eq. (4) 2.2.1.3. B-distribution based TFD. Several researches have been car-
  ried out in the design and development of kernel function g(v, τ ).
∞ ∞
B-distribution (BD), Modified B-distribution and Extended modified
x(t ) = ∫ ∫ S (t, f )dt ei2π f t df (5)
−∞ −∞ B-distribution are belonged to high-resolution Cohen class TFD.
The B-distribution based TFD is developed in order to provide
Hence, the S-transform can be represented in Fourier domain
both high resolution and cross-term suppression. It is a two di-
and it is given by
∞
 2π a2
 mensional function in Doppler-lag domain centered on a origin
S (t, f ) = ∫ X ( f + a )e f2 ei2π f a da (6) with sharp cut-off edges. Thus it retains most of actual frequency
−∞ components of signals and attenuates most of the cross-terms. The
The mathematical expression of S-transform in discrete domain mathematical expression of B-distribution in ambiguity domain is
is given as given by [14]
(β + jπ v )
 g(v, τ ) = |τ |β
N−1
2π 2 m2 i2π mn (11)
S[n, k] = e− k2 X [k + m]e− N (7) 21−2β (2β )
m=0
where β is smoothening parameter that controls the volume of the
2 2 ∞
− 2π 2m
where n is a time sample, k is a frequency sample and e k is kernel in 2D- g(v, τ ) domain and (z ) = ∫ t z−1 e−t dt (Re(z) > 0). It
−∞
a Fourier transform of Gaussian window g(t − τ , σ ) [11]. is reported that the value β = 0.01 is suitable for the better cross-
term suppression and good time-frequency resolution [25]. Hence,
2.2.1.2. Cohen class TFD. Cohen class TFDs are particularly devel-
the value of α = 0.01 is adapted for this study.
oped for the representation of nonstationary signals and it does
The discrete time expression of BD based kernel is
not have any stationarity assumptions. It is a bilinear transform

β
and it suffers from spurious terms called cross-terms or artifacts |2m|
which present in the midway between the actual frequency com- G[n, m] = ∗sincm (12)
cosh2 n
ponents of multicomponent signals. The general form of Cohen
class TFD for a signal x(t) is given by [14] where G[n, m] is discrete version of G(t, τ ).
∞ ∞
ρz (t, f ) = ∫ ∫ g(v, τ )Az (v, τ )e j2π (vt− f τ ) dvdτ (8) 2.2.1.4. Extended modified B-distribution based TFD. EMBD based
−∞ −∞ TFD addresses the drawback of B-distribution kernel. When τ = 0,
where g(v, τ ) is weighing or smoothening function in Doppler the ambiguity function of B-distribution become null and thus it
(v)-lag (τ ) domain or ambiguity domain and Az (v,τ ) is a ambigu- attenuates most of actual frequency components of the signals.
ity function. The mathematical expression of ambiguity function is Further, the B-distribution based kernel became lag independent or
given by Doppler dependent for 0 < β  1. In this scenario, the cross-terms
∞ τ  τ presents in the Doppler domain will be smoothened where as it
Az (v, τ ) = ∫ z(t + )z ∗ t − e− j2π vt dt (9) retains the cross terms that are presented in the lag direction. In
−∞ 2 2
order to avoid kernel null function around the origin of ambiguity
where z(t) is an analytic associate of the considered signal x(t) and
domain and to perform filtering in both Doppler and lag directions,
it can be represented as z(t) = x(t) + iH{x(t)}. In addition, the time-
∞ EMBD kernel has been proposed. The mathematical expression of
lag kernel is obtained by g(v, τ ) = ∫ G(t, τ )e− j2π vt dt. EMBD is defined as [19,26]
−∞
The Eq. (8) indicates that the TFD is first obtained by smoothing |(α + jπ v )|2 |(β + jπ τ )|2
the ambiguity function using the kernel function (g(v, τ )) and then g( v, τ ) = (13)
 2 (α )  2 (β )
taking two dimensional Fourier transform over frequency and time
axis. This process is equivalent to performing two-dimensional fil- where α and β are smoothing parameters. This formulation of
tering on the ambiguity function. Hence, the careful selection of EMBD allows one to control the kernel function in both Doppler
kernel (two dimensional filter) in the ambiguity domain leads to and lag domain separately. The time-lag kernel in discrete domain
reduction or removal of the cross-terms generated due the non- is given as
stationary and multicomponent nature. The kernel function deter- cosh−2α m cosh−2β n
mines the properties of TFDs. G[m, n] =  −2α m
 −2β n
(14)
n cosh n cosh
48 P.A. Karthick et al. / Computer Methods and Programs in Biomedicine 154 (2018) 45–56

2.2.2. Time-frequency features 2.2.2.5. Spectral entropy. Time-frequency based spectral entropy
The high-resolution time-frequency method captures even measures the randomness of the signal energy distribution. The
small variations of the frequency components that occur with higher value of spectral entropy denotes that the signal energy
respect to its time. However, all TFD points cannot be consid- is uniformly spread in the time-frequency plane while the lower
ered for the classification as it increases computational effort. value indicates that the signal energy more concentrated in the
Hence, it is necessary to identify suitable features that best de- specific regions of time-frequency plane. The mathematical expres-
scribe these small variations. Twelve time-frequency features are sion of spectral entropy is given by [20]
extracted from these TFDs and the mathematical expressions are


N 
L
P [n, k] P [n, k]
given below. SE = −   log2   (20)
n=1 k=1 n k P [n, k] n k P [n, k]

2.2.2.1. Median frequency (MDF). Median frequency is a frequency 2.2.2.6. Renyi entropy (RE). Renyi entropy extracted from the time-
at which the total power is divided in to two equal halves. In or- frequency distribution measures the complexity of signals. A
der to measure the variation of median frequency in time-axis, the higher degree of nonstationary, multicomponent signals results in
time-frequency based MDF is defined as [8] higher RE. For example, a multicomponent signal will yield a larger
  RE than the singles component signal. The time-frequency based
1
M1 L L
P [n, k] = P [n, k] = P [n, k] (15) RE is given by [20]
k=1 k=M1
2
k=1

α
1  N L
P [n, k]
RE = log2   (21)
1−α k P [n, k]
MDF = fM1 (16) n=1 k=1 n

where α should be odd integer and α > 2. We have chosen α = 3

where M1 is a frequency bin associated with the median frequency for this study [28].
( fM1 ) at the time instant n, P[n, k] is the power spectrum of nth
time sample at the frequency bin k, and L is a total length of fre- 2.2.2.7. Singular value decomposition (SVD) based entropy. SVD
quency bin. decomposes the time-frequency spectrum into two subspaces
namely, signal subspace and orthogonal alternate subspace. The
2.2.2.2. Mean frequency (MNF). Mean power frequency is defined decomposition of time-frequency spectrum P[n, k] can be written
as the sum of product of frequencies and its corresponding power as
divided by the total sum of powers. The time-frequency based P [n, k]NxM = USV H (22)
mean frequency is given by [8]
where, U is a NxN unitary matrix, S is a NxM diagonal matrix with
L non-negative numbers and VH is a complex conjugate of V a uni-
k=1 fk P [n, k]
MNF = L (17) tary matrix of MxM size. The diagonal elements of S is called sin-
k=1 P [n, k]
gular values. The entropy based on the singular values measures
where fk is the frequency value of power spectrum at the fre- the complexity of time-frequency spectrum. It provides informa-
quency bin k, and fL is the maximum frequency corresponding to tion about magnitude and number of non-zero elements of the sin-
Lth bin. gular vectors and is given by [20,28]

N
CM = − s̄i logs̄i (23)
2.2.2.3. Normalized spectral moment (NSM). It is a ratio of spectral
i=1
moment of order (−1) to the spectral moment of higher order such
as 2, 3 4 and 5. The spectral moment of order (−1) is sensitive to where s̄i is the normalized singular value of the matrix P[n, k]NxM

the changes in the ultralow and low frequencies of sEMG power and it is calculated using s̄i = si / si [21,23].
i
spectrum and it attributes to the increased negative after poten-
tials during fatigue condition. Further, the higher order of spectral 2.2.2.8. Time-frequency based statistical features. The time domain
moment is sensitive to the higher frequency components and it is based statistical properties such as mean, variance, coefficient of
related to the change in conduction velocity [27]. The mathemati- variation, skewness, and kurtosis characterize the probability dis-
cal expression of NSM is given by tribution of signals. In order to fully exploit the time-frequency
kmax spectrum, the 1d time domain features are extended to 2d and
k=k
fk−1 P [n, k] the corresponding time-frequency statistical features are given by
NSM p = k min (18)
max
k=kmin
fkp P [n, k] [20,28]
1 
where NSMp is a spectral moment of order p. Mean = P [n, k] (24)
MN
N M

1 
2.2.2.4. Concentration measure. It determines the concentration of V ariance = (P[n, k] − μ )2 (25)
energy distribution of signals in the TFD domain and it is given by MN
N M
[20,28] 
1
 2 Skewness = (P[n, k] − μ )3 (26)

N 
L (MN − 1 )μ 3
N M
1/2
Mp = |P[n, k]| (19)

1
n=1 k=1 Kurtosis = (P[n, k] − μ )4 (27)
(MN − 1 )μ 4
N M
where L and M are total number of time samples and FFT points
σ
respectively. A larger value of concentration measure indicates that Coe f f icient o f variation (CoV ) = (28)
the power is distributed in the entire time-frequency plane while μ
the smaller value implies that the power is concentrated only on where μ and σ are time-frequency mean and standard deviation
certain areas of time-frequency plane. respectively.
 P.A. Karthick et al. / Computer Methods and Programs in Biomedicine 154 (2018) 45–56 49

2.2.3. Feature selection Table 1

Demographics of subjects.
Genetic algorithm (GA) and binary particle swarm optimization
(BPSO) are employed as feature selection techniques to remove the Parameter Unit Mean ± SD
redundant and irrelevant features. Age Years 26.12 ± 3.12
Genetic algorithm (GA) is a stochastic iterative search method, Weight kg 69.4 ± 11.56
used to select the best features from the large set of feature vec- Height m 1.63 ± 0.28
tors. This algorithm has been developed from the inspiration of bi- Body mass index kg/m2 24.39 ± 3.1

ological evolution process. The initial population has been created

randomly and their fitness value has been evaluated. Based on the
fitness value, a new offspring will be produced and used for the 2.2.4.2. Support vector machine. SVMs are supervised machine
optimization of entire population. The new offsprings are gener- learned methods, extensively used in classification, regression and
ated using the operations such as mutation or cross over operation. density estimation. It is considered to be one of the most ro-
The best feature set will be obtained when the process iterates and bust and accurate machine learning algorithms. SVM classifier uses
reaches the maximum number of generation or when the process a discriminate hyperplane to differentiate the members of two
converges to the expected solution [29]. classes. It performs nonlinear mapping of input vectors to a very
Particle swarm optimization is an another optimization tech- high dimensional feature space. An optimal hyperplane maximizes
nique based on the social behavior of bird flocking. It is developed the margin or the distance between the nearest training data of
to optimize the problems that could be solved using real numbers. two different classes. It is a quadratic optimization problem. De-
Recently, this approach is widely used as a feature selection tool. In cision boundary of the SVMs can be linear or nonlinear based on
this technique, particles are composed of cells called position. Ac- the separability of training data points. In the real life applications,
tually, swarm composed of these particles separates in the solution most of the extracted features are not linearly separable. A nonlin-
space randomly. Each particle in the swarm is a candidate solution ear decision boundary or hyperplane can be obtained with a low
of the solution set. Based on the best value of each particle (pbest )in increase in computational complexity by using kernel trick [21,32].
the swarm and the swarm itself (gbest ), the particle updates it ve- In this work, the kernels such as polynomial and radial basis func-
locity and position in order to obtain the optimum solution. The tions are considered.
velocity and position of the particle are given by [30,31]
    2.2.4.3. Random forest. Random forest is an ensemble learning
vi, j (t +1 ) = wvi, j (t ) + c1 R1 pbesti, j − xi, j (t ) + c2 R2 gbesti, j − xi, j (t ) technique. It is a powerful classifier widely used in several studies.
(29) It is constructed using multiple distinct decision trees and the class
that is predicted by the majority of tress will be considered for
the final decision. In this ensemble technique, each decision tree
xi, j (t + 1 ) = xi, j (t ) + vi, j (t + 1 ) (30) is trained with different subsets of the training data using random
samples drawn with replacement (Bootstrap aggregating) from the
where i and j are the index and position of a particle, t is an iter- original training set. Therefore, some instance, may be used more
ation number, vi,j (t) is the velocity of ith particle in swarm on jth than once in the training process of these classifiers. This phe-
index of position in particle, vmin ≤ vi,j (t) ≤ vmax and xi,j (t) is the nomenon increases the stability and robustness of the classifier for
position. R1 and R2 are the random numbers follows uniform dis- the data that are with slight variations [33,34].
tribution between 0 and 1. c1 and c2 are the acceleration numbers
and w is an inertia weight. 2.2.4.4. Rotation forest. Rotation forest is also an ensemble classi-
In the binary particle swarm optimization (BPSO) technique, the fication method. In this method, the feature set is divided into a
probability of the particle being 1 or 0 is determined by the ve- number of subsets and principal components are extracted from
locity value using sigmoid function. The position of the particle is each subset. Then the same number of rotations is performed to
updated using the following expression [30] generate new features for a decision tree. Hence, both diversity
   and individual accuracy can be achieved in this ensemble classi-
0 i f rand () ≥ S vi, j (t + 1 )
xi, j (t + 1 ) =   (31) fier [34,35].
1 i f rand () < S vi, j (t + 1 )

where rand() is the random numbers drawn from the uniform dis- 3. Results and discussion
tribution between 0 and 1. S(.) is the sigmoid function.
The representative sEMG signal segments of muscle nonfatigue
2.2.4. Classification and fatigue conditions are shown in Fig. 1. The amplitude and fre-
Five classifiers, namely, naïve Bayes classifier and support vector quency variations of the recorded signals are dependent on sev-
machine with polynomial and radial basis kernel, random forest eral muscle physiological parameters such as firing rate, motor unit
and rotation forest are considered in this work. recruitment patterns, muscle fiber conduction velocity, different
types of muscle fibers and volume conductors. The demographic
details of the participants are presented in Table 1.
2.2.4.1. Naïve Bayes classifier. It is a simple probabilistic classifier
Fig. 2 shows the time-frequency spectrum of the representative
which works based on Bayes theorem. It assumes the features are
sEMG signals obtained from the time-frequency methods, namely,
independent. NB classifier estimates model parameters based on
STFT (a, b), S-transform (c, d), BD-TFD (e, f) and EMBD-TFD (g,
the distribution of training data sets. It computes the probability
h). It shows the temporal variations of instantaneous power spec-
of new data and then selects the most probable class. The mathe-
tral components of the signals. These variations are dependent on
matical expression of NB classifier is given by [21]
recruitment and decruitment of motor units, changes in muscle
   
  P A/v j P vj length, arm angle and movement of electrode with respect to the
P v j /A = (32) innervation zone that are concomitant with dynamic contractions
P (A )
[1]. Fig. 2 (a), (c), (e) and (g) show the time-frequency spectrums
where, vj is the hypothesis in the set of hypothesis ν and A is the that are corresponding to a nonfatigue segment of muscle contrac-
set of attributes which describes the data. tion. More frequency components are found in these spectrums in
50 P.A. Karthick et al. / Computer Methods and Programs in Biomedicine 154 (2018) 45–56

Fig. 1. Representative (a) nonfatigue and (b) fatigue segment of sEMG signals.

comparison with spectrums of fatigue condition (Figs. 2(b), (d), (f) Table 2
Comparison of statistical significance of all features in muscle nonfatigue and
and (h)). It is also observed that the instantaneous power spec-
fatigue conditions.
trum shifts towards lower frequencies in fatigue condition. This is
attributed to the reduction in the propagation velocity of action SNo Time-frequency p-value
features
potential along the muscle fibers [8]. S-transform BD-TFD EMBD-TFD
Further, notable variations are observed in the spectral compo-
1 Median frequency 1.75E−18∗∗ 8.13E−13∗∗ 2.60E−18∗∗
nents of STFT, S-transform, BD and EMBD based TFDs. The resolu- 2 Mean frequency 1.75E−18∗∗ 9.93E−15∗∗ 1.477E−19∗∗
tion is found to be relatively poor for the STFT based TFDs and is 3 NSM 0.008 0.12 0.88
due to the fixed size of the window. In fact, it has been reported 4 Concentration 0.00165∗ 0.023∗ 0.003∗
that this technique may not be appropriate in capturing the highly measure
5 Spectral entropy 1.87E−18∗∗ 5.16E−10∗∗ 0.24
varying nonstationary components of sEMG signals recorded un- 6 Renyi entropy 4.55E−12∗∗ 1.98E−6∗∗ 0.82
der dynamic contractions [8]. Therefore, this method is not con- 7 SVD based entropy 3.50E−10∗∗ 2.53E−11∗∗ 8.55E−10∗∗
sidered for the classification. In the case of S-transform, it can be 8 Mean 2.910E−6∗∗ 1.48E−4∗∗ 4.53E−4∗∗
seen that the signal components smear in both time and frequency 9 Variance 0.079 0.086 0.10
10 Skewness 6.41E−14∗∗ 2.34E−20∗∗ 1.73E−17∗∗
axis. It is found that the cross-terms are not completely removed
11 Kurtosis 2.12E−13∗∗ 3.57E−20∗∗ 1.73E−17∗∗
by B-distribution kernel and artifacts are observed in the resultant 12 CoV 3.80E−6∗∗ 1.27E−4∗∗ 6.25E−5∗∗
time-frequency spectrum. Interestingly, the energy concentration is
∗
Statistically significant (p < .05)
also lower in BD-TFD. It is due to the fact that BD kernel func- ∗∗
Statistically highly significant (p < .001)
tion becomes null at origin and attenuating most of the actual sig-
nal components. The time varying frequency components of sEMG
signals are better represented in the time- frequency spectrum of
EMBD technique. This is due to the presence of separable low pass tral entropy, Renyi entropy, CoV, mean and median frequencies are
filter in EMBD kernel that efficiently attenuates the cross-terms in- chosen by GA from the S-transform. In addition to these features,
terference of the signal components in both time and frequency time-frequency mean and SVD based entropy have been selected
axis [19]. for BD-TFD. In the case of EMBD technique, only four features
Most of the extracted features are found to have statistically namely SVD based entropy, kurtosis, mean frequency and median
significant difference in nonfaigue and fatigue conditions and it is frequency are chosen by GA. It is observed that the number of fea-
presented in Table 2. From the p-values, it is found that median tures reduced by 58%, 41% and 66% for S-transform, BD-TFD and
frequency, mean frequency, SVD based entropy, and the statistical EMBD-TFD respectively. It is found that only four features namely
properties such as time-frequency mean, skewness, kurtosis and Shannon entropy, Renyi entropy, concentration measure and CoV
coefficient of variation exhibit statistically highly significant differ- are selected by BPSO from the BD-TFD. Apart from this features,
ence (p < .001). Further, it can be seen that NSM and variance of two more features (median frequency and skewness) are chosen
all TFDs and spectral and Renyi entropy of EMBD technique show for the S-tranform. For the EMBD-TFD, six features, namely Shan-
least significance difference in muscle fatigue and nonfatigue con- non entropy, SVD based entropy, time-frequency mean, CoV, mean
ditions. Other feature, concentration measure exhibits significant frequency and NSM are chosen by using BPSO. It is also observed
difference in these two conditions (p < .05). that highest number of features is reduced for BD-TFD. It is impor-
In this study, 52 healthy subjects participated and totally twelve tant to note that the conventional mean and median frequencies
features are extracted from the TFDs. In order to reduce the di- are selected for majority of time-frequency methods.
mension of the feature set, GA and BPSO are employed. The al- Figs. 3 and 4 shows the median frequency and mean frequency
gorithm selects the prominent features that demarcate the mus- respectively, computed from the considered TFDs. It is found that
cle nonfatigue and fatigue conditions. Five features, namely, spec- the both median frequency and mean frequency is higher in non-
fatigue condition and these results are in agreement with the liter-
 P.A. Karthick et al. / Computer Methods and Programs in Biomedicine 154 (2018) 45–56 51

Fig. 2. TFD representation of the sEMG signal in nonfatigue (a, c, e, g) and fatigue (b, d, f, h) conditions; STFT (a,b), S-transform (c, d), B-distribution (d, f) and EMBD
technique (g, h).

ature [1,21]. The results indicate that the proposed time-frequency 30% for S-transform, BD-TFD and EMBD-TFD respectively. Similarly,
approach such as S-transform and EMBD-TFD based analysis could a high percentage difference (40%) is observed in the case of S-
be used for muscle fatigue application. Reduction in these fea- transform based mean frequency.
ture values may be attributed to the participation of more slow The performances of the TFDs with the full features pertaining
twitch muscle fibers and synchronization of motor units during fa- to the five classifiers are presented in Table 3. With the all fea-
tigue condition. The percentage differences between the median tures, S-transform performs better than both BD and EMBD-TFDs.
frequencies in these two conditions are found to be 38%, 26% and It is found that the highest accuracy of above 90% is achieved in S-
52 P.A. Karthick et al. / Computer Methods and Programs in Biomedicine 154 (2018) 45–56

Fig. 3. Median frequency extracted from (a) S-transform, (b) BD-TFD and (b) EMBD-TFD.

Table 3
Classification performance indices of TFDs with the full features.

Time-frequency method Classifier Sensitivity (%) Specificity (%) Accuracy (%)

S-transform Naïve Bayes 76.92 59.62 68.27

SVM-polynomial 92.31 86.54 89.42
SVM-Radial basis 98.08 82.69 90.38
Random forest 88.90 92.00 90.45
Rotation forest 89.14 93.86 91.5
BD-TFD Naïve Bayes 76.92 67.31 72.12
SVM-polynomial 87.44 85.64 86.54
SVM-Radial basis 88.46 76.92 82.69
Random forest 84.92 86.34 85.63
Rotation forest 84.24 91.50 87.87
EMBD-TFD Naïve Bayes 75.00 59.62 67.31
SVM-polynomial 86.54 92.31 89.43
SVM-Radial basis 82.69 80.77 81.73
Random forest 82.14 87.46 84.80
Rotation forest 88.62 90.20 89.41

-Highest accuracies among different TFDs are represented in bold.

 P.A. Karthick et al. / Computer Methods and Programs in Biomedicine 154 (2018) 45–56 53

Fig. 4. Mean frequency extracted from (a) S-transform, (b) BD-TFD and (b) EMBD-TFD.

transform - rotation forest classifier combination. Naive Bayes clas- tures. It may be due to the attenuation of actual signal components
sifier results in poor performance for all the TFDs resulting in an by the BD-kernel in the resultant time-frequency spectrum.
average accuracy of 69%. Table 6 shows a brief summary of studies related to muscle
Tables 4 and 5 represent the classification accuracies of TFDs for fatigue classification in both isometric and dynamic contractions.
the features selected by GA and BPSO respectively. It can be seen Popular nonstationary methods such as STFT, smoothed-pseudo
that there is a notable improvement in the classification accuracy WVD and continuous wavelet transform (CWT) were used for the
with the reduced number of features. The considerable improve- characterization of sEMG signals in muscle fatigue condition. It was
ment in naïve Bayes classifier’s performance indicates the effec- shown that the CWT coefficients along with multilayer percep-
tiveness of the feature selection algorithms. Surprisingly, the com- tron neural network (MPNN) ‘provided a classification accuracy of
bination of BPSO selected features and SVM radial basis kernel ex- 91% [36]. In another study, a pseudo-wavelet function was devel-
hibits poor performance across all TFDs. It can also be seen from oped and an optimal scale was obtained using GA which specif-
the classification accuracies that GA perform better than BPSO. A ically improved the separation of muscle fatigue and nonfatigue
maximum accuracy of 91.39% is achieved in EMBD-TFD and SVM- conditions. This automated system had a classification accuracy of
polynomial kernel combination with only four selected features. 88.4% [37]. Recently, multiple time-window (MTW) features and
Although the performance of S-transform and EMBD is similar, the classifiers such as naïve Bayes, SVMs, k-nearest neighbor (k-NN),
number of features that have been selected for the classification is and linear discriminant analysis (LDA) were used to differenti-
slightly more in the case of S- transform. It is also observed that ate the sEMG signals in fatigue conditions. The features were ex-
BD-TFD approach exhibits poor performance with the selected fea- tracted from different window functions such as rectangular win-
dows, Hamming windows, trapezoidal windows, and Slepian win-
54 P.A. Karthick et al. / Computer Methods and Programs in Biomedicine 154 (2018) 45–56

Table 4
Classification performance indices of TFDs with features selected using GA.

Time-frequency method Classifier Sensitivity (%) Specificity (%) Accuracy (%)

S-transform Naïve Bayes 89.42 87.50 88.46

SVM-polynomial 90.38 88.54 89.46
SVM-Radial basis 92.31 84.62 88.47
Random forest 85.74 91.66 88.70
Rotation forest 88.50 88.50 88.50
BD-TFD Naïve Bayes 86.54 73.04 79.79
SVM-polynomial 84.62 84.62 84.62
SVM-Radial basis 86.54 75.00 80.77
Random forest 87.00 90.00 88.50
Rotation forest 83.12 93.34 88.23
EMBD-TFD Naïve Bayes 92.31 84.62 88.46
SVM-polynomial 92.31 90.46 91.39
SVM-Radial basis 86.54 86.46 87.50
Random forest 86.82 86.48 86.65
Rotation forest 88.70 90.20 89.45

-Highest accuracies among different TFDs are represented in bold.

Table 5
Classification performance indices of TFDs with features selected using BPSO.

Time-frequency method Classifier Sensitivity (%) Specificity (%) Accuracy (%)

S-transform Naïve Bayes 75.75 88.15 81.95

SVM-polynomial 83.28 86.12 84.70
SVM-Radial basis 63.50 63.50 63.50
Random forest 85.55 89.75 87.65
Rotation forest 83.86 86.84 85.35
BD-TFD Naïve Bayes 64.10 94.30 79.20
SVM-polynomial 77.42 78.38 77.90
SVM-Radial basis 58.46 64.14 61.30
Random forest 75.88 78.20 77.04
Rotation forest 70.75 84.65 74.48
EMBD-TFD Naïve Bayes 72.50 94.32 83.41
SVM-polynomial 90.20 88.60 89.40
SVM-Radial basis 68.55 70.00 69.28
Random forest 85.48 89.82 87.65
Rotation forest 90.23 88.64 89.44

-Highest accuracies among different TFDs are represented in bold

Table 6
Comparison of accuracies with previous studies on muscle fatigue classification.

Classifiers-feature
References Contraction type Methods selection Max. classification accuracy (%)

Subasi and Kiymik [36] Isometric STFT, Smoothed MPNN-ICA 91

WVD and CWT
Al-Mulla et al. [37] Isometric GA tuned Wavelet LDA 88.41
function
Venugopal et al. [21] Isometric MTW features k-NN-IG ranking 93
Makaram and Swaminathan [38] Dynamic Multiscale Renyi SVM-Binary bat 86.7
Entropy algorithm
Kiran Marri and Swaminathan [39] Dynamic Multifractal k-NN, NB and LR-IG 84
features
Kiran Marri and Swaminathan [40] Dynamic Multifractal k-NN and LR-WR 88
singularity sum, IG and GA
spectrum based
features
In this work Dynamic S-transform, BD SVM, NB, Random 91.39
and EMBD and Rotation
forest-GA and
BPSO

dows and were utilized for the classification. Classification accu- algorithm was investigated in classifying the sEMG signals under
racy of about 93% was achieved using the combination of MTW dynamic fatiguing muscle contractions. Using the multiscale Renyi
features and k-NN algorithm [21]. It is important to note that all entropy measures and the combination of SVM and binary bat al-
these classification systems are based on isometric muscle contrac- gorithm, a fatigue detection accuracy of more than 86% was ob-
tion. tained [38]. In another research study, the fatigue and the non-
Only a few studies are reported for the automated muscle fa- fatigue segment of sEMG signals were analyzed by using multi-
tigue detection based on real-time dynamic muscle contractions fractal features. In this approach, the dimensionality of multifractal
[38–40]. Recently, the effectiveness of binary bat feature selection features was reduced by IG based ranking and fed to the classi-
 P.A. Karthick et al. / Computer Methods and Programs in Biomedicine 154 (2018) 45–56 55

fiers such as k-NN, naïve Bayes and logistic regression (LR). It was References
shown that LR performed better with a classification accuracy of
84% [39]. In addition to this, twenty five features were extracted [1] R. Merletti, P.A. Parker, Electromyography: Physiology, Engineering, and
Non-Invasive Applications, John Wiley & Sons, 2004.
from the multifractal spectrum and their discriminative power was [2] M. Cifrek, V. Medved, S. Tonković, S. Ostojić, Surface EMG based muscle fatigue
examined using feature selection algorithms such as Wilcoxon rank evaluation in biomechanics, Clin. Biomech. 24 (2009) 327–340.
sum, information gain (IG) and GA techniques and classifiers such [3] M.J. Zwarts, G. Bleijenberg, M. van Engelen, Clinical neurophysiology of fatigue,
Clin. Neurophysiol. 119 (2008) 2–10.
as k-NN and LR. An accuracy of 88% was achieved with k-NN and [4] V.P. Singh, D.K. Kumar, B. Polus, S. Fraser, Strategies to identify changes in
GA combination [40]. SEMG due to muscle fatigue during cycling, J. Med. Eng. Technol. 31 (2007)
In this work, high-resolution time-frequency methods and fea- 144–151.
[5] J. Zhang, T.E. Lockhart, R. Soangra, Classifying lower extremity muscle fatigue
tures are utilized to classify the dynamic muscle fatigue and
during walking using machine learning and inertial sensors, Ann. Biomed. Eng.
nonfatigue conditions. It is found that the combination of high- 42 (2014) 600–612.
resolution EMBD-TFD and SVM-polynomial kernel classifies these [6] M.R Al-Mulla, M. R., F. Sepulveda, M. Colley, A review of non-invasive tech-
niques to detect and predict localised muscle fatigue, Sensors 11 (2011)
nonstationary sEMG signals with an accuracy of 91.39%. It is impor-
3545–3594.
tant to note that this accuracy is relatively higher than other stud- [7] N.A Dimitrova, T.I. Arabadzhiev, J.Y. Hogrel, G.V. Dimitrov, Fatigue analysis of
ies that are based on dynamic muscle fatigue classification [38– interference EMG signals obtained from biceps brachii during isometric vol-
40]. Further, only four features are used to obtain such accuracy untary contraction at various force levels, J. Electromyogr. Kinesiol. 19 (2009)
252–258.
and hence, it is computationally effective. It is well-known that [8] P. Bonato, S.H. Roy, M. Knaflitz, C.J. De Luca, Time-frequency parameters of the
the degree of nonstationarity of sEMG signals is higher in dynamic surface myoelectric signal for assessing muscle fatigue during cyclic dynamic
contractions than isometric contractions. Therefore, these high- contractions, IEEE Trans. Biomed. Eng. 48 (2001) 745–753.
[9] D.R. Rogers, D.T. MacIsaac, EMG-based muscle fatigue assessment during dy-
resolution time-frequency methods could also be used to study the namic contractions using principal component analysis, J. Electromyogr. Kine-
isometric muscle activities in order to improve the classification siol. 21 (2011) 811–818.
accuracy. [10] S. Assous, B. Boashash, Evaluation of the modified S-transform for time-fre-
quency synchrony analysis and source localisation, J. Adv. Sig. Process. 1 (2012)
1–18.
4. Conclusion [11] R.G. Stockwell, L. Mansinha, R.P. Lowe, Localization of the complex spectrum:
the S transform, IEEE Trans. Sig. Process. 44 (1996) 998–1001.
[12] A. Yan, W Zhou, Q. Yuan, S. Yuan, Q. Wu, X. Zhao, J. Wang, Automatic seizure
High-resolution time-frequency methods are considered to be a detection using Stockwell transform and boosting algorithm for long-term EEG,
powerful tool for analyzing the multicomponent and nonstationary Epilepsy Behav. 45 (2015) 8–14.
[13] N. Ortigosa, C. Fernández, A. Galbis, Ó. Cano, Phase information of time-fre-
variations of signals. The objective of this work is to introduce the
quency transforms as a key feature for classification of atrial fibrillation
time-frequency methods such as S-transform, BD and EMBD and episodes, Physiol. Meas. 36 (2015) 409–424.
its features for dynamic muscle fatigue analysis and to compare [14] B. Boashash, Time Frequency Signal Analysis and Processing A Comprehensive
Reference, Elsevier, 2003.
their classification performance using machine learning algorithms.
[15] M. Knaflitz, F. Molinari, Assessment of muscle fatigue during biking, IEEE
The sEMG signals of nonfatigue and fatigue conditions are sub- Trans. Neural Syst. Rehabil. Eng. 11 (2003) 17–23.
jected to S-tranform, BD and EMBD-TFDs. Twelve features are ex- [16] M. Gonzalez-Izal, A. Malanda, I. Navarro-Amezqueta, E.M. Gorostiaga, F. Mallor,
tracted from each TFD and the dominant attributes are chosen us- J. Ibanez, M. Izquierdo, EMG spectral indices and muscle power fatigue during
dynamic contractions, J. Electromyogr. Kinesiol. 20 (2010) 233–240.
ing GA and BPSO. These selected features are given to the five dif- [17] B. Boashash, G. Azemi, J. O’Toole, Time-frequency processing of nonstationary
ferent classifiers and their performances are compared against the signals: Advanced TFD design to aid diagnosis with highlights from medical
full features. The results demonstrate that all the proposed time- applications, IEEE Sig. Process. Mag. (2013) 108–119.
[18] P.A. Karthick, S. Ramakrishnan, Surface electromyography based muscle fatigue
frequency techniques are able to represent the nonstationary vari- progression analysis using modified B distribution time–frequency features,
ations of sEMG signals. Most of the extracted features have statis- Biomed. Sig. Process. Control 26 (2016) 42–51.
tically significant difference in muscle fatigue and non-fatigue con- [19] B. Boashash, T. Ben-Jabeur, Design of a high-resolution separable-kernel
quadratic TFD for improving newborn health outcomes using fetal movement
ditions. Further, the number of features is reduced by 58%, 41% and detection, Information Science, Signal Processing and their Applications, 2012.
66% for S-transform, BD and EMBD TFD respectively, with GA. The [20] B. Boashash, G. Azemi, N. Ali Khan, Principles of time–frequency feature ex-
other feature selection technique, BPSO reduced 50% of features traction for change detection in non-stationary signals: applications to new-
born EEG abnormality detection, Pattern Recognit. 48 (2014) 616–627.
for S-Transform and EMBD-TFD and 66% of features for BD-TFD.
[21] G. Venugopal, M. Navaneethakrishna, S. Ramakrishnan, Extraction and analysis
The combination of EMBD-SVM polynomial kernel gives a maxi- of multiple time window features associated with muscle fatigue conditions
mum accuracy of 91% for the GA selected features. Thus EMBD- using sEMG signals, Expert Syst. Appl. 41 (2014) 2652–2659.
[22] P.A. Karthick, G. Venugopal, S. Ramakrishnan, Analysis of muscle fatigue pro-
SVM polynomial kernel approach could be used for dynamic mus-
gression using cyclostationary property of surface electromyography signals, J.
cle fatigue application such as sports biomechanics, work place and Med. Syst. 40 (2016).
ergonomics. [23] P.A. Karthick, S. Ramakrishnan, Muscle fatigue analysis using surface EMG sig-
nals and time–frequency based medium-to-low band power ratio, Electron.
Lett. 52 (2015) 185–186.
Conflicts of interest [24] D.K. Kumar, V.P. Singh, I. Cosic, S.L. Giudice, J. Hawley, Strategies to identify
changes in SEMG due to muscle fatigue during cycling, in: IEEE Engineering in
Medicine and Biology, 2005, pp. 6683–6686.
The authors declare that there is no conflict of interest related [25] B. Barkat, B. Boashash, A high-resolution quadratic time-frequency distribu-
to this paper. tion for multicomponent signals analysis, IEEE Trans Sig. Process. 49 (2001)
2232–2239.
[26] B. Boashash, M.S. Khlif, T. Ben-Jabeur, C.E. East, P.B. Colditz, Passive detection
Acknowledgment of accelerometer-recorded fetal movements using a time–frequency signal pro-
cessing approach, Digit. Sig. Process. 25 (2014) 134–155.
[27] G.V. Dimitrov, T.I. Arabadzhiev, K.N. Mileva, J.L. Bowtell, N. Crichton, N.A. Dim-
The authors would like to thank the participants for their co- itrova, Muscle fatigue during dynamic contractions assessed by new spectral
operation, and Indian Institute of Technology Madras for providing indices, Med. Sci. Sports Exerc. 38 (2006) 1971–1979.
the instrumentation facilities to carry out this research work. [28] B. Boashash, S. Ouelha, Automatic signal abnormality detection using time-fre-
quency features and machine learning: a newborn EEG seizure case study,
Knowl. Based Syst. 106 (2016) 38–50.
Supplementary materials [29] G. Wang, Z. Yan, X. Hu, H. Xie, Z. Wang, Classification of surface EMG signals
using harmonic wavelet packet transform, Physiol. Meas. 27 (2006) 1255–1267.
[30] I. Babaoglu, O. Findik, E. Ülker, A comparison of feature selection models utiliz-
Supplementary material associated with this article can be ing binary particle swarm optimization and genetic algorithm in determining
found, in the online version, at doi:10.1016/j.cmpb.2017.10.024.
56 P.A. Karthick et al. / Computer Methods and Programs in Biomedicine 154 (2018) 45–56

coronary artery disease using support vector machine, Expert Syst. Appl. 37 [36] A. Subasi, M.K. Kiymik, Muscle fatigue detection in EMG using time–frequency
(2010) 3177–3183. methods, ICA and neural networks, J. Med. Syst. 34 (2010) 777–785.
[31] A. Subasi, J. Kevric, M.A. Canbaz, Epileptic seizure detection using hybrid ma- [37] M.R. Al-Mulla, F. Sepulveda, M. Colley, Evolved pseudo-wavelet function to op-
chine learning methods, Neural Comput. Appl. (2017) 1–9. timally decompose sEMG for automated classification of localized muscle fa-
[32] E. Avci, A new intelligent diagnosis system for the heart valve diseases by us- tigue, Med. Eng. Phys. 33 (2011) 411–417.
ing genetic-SVM classifier, Expert Syst. Appl. 36 (2009) 10618–10626. [38] N. Makaram, R. Swaminathan, A binary bat approach for identification of fa-
[33] F. Akram, S. M.Han, T.S. Kim, An efficient word typing P300-BCI system us- tigue condition from semg signals, International Conference on Swarm, Evolu-
ing a modified T9 interface and random forest classifier, Comput. Bio. Med. 56 tionary, and Memetic Computing, Springer, 2014.
(2015) 30–36. [39] K. Marri, R. Swaminathan, Classification of muscle fatigue using surface elec-
[34] E. Aličković, A. Subasi, Breast cancer diagnosis using GA feature selection and tromyography signals and multifractals, 12th International Conference on
rotation forest, Neural Comput. Appl. 28 (2017) 753–763. Fuzzy Systems and Knowledge Discovery, 2015.
[35] X. Du, S. Dua, R.U. Acharya, C.K. Chua, Classification of epilepsy using high- [40] K. Marri, R. Swaminathan, Classification of muscle fatigue in dynamic contrac-
-order spectra features and principle component analysis, J. Med. Syst. 36 tion using surface electromyography signals and multifractal singularity spec-
(2012) 1731–1743. tral analysis, J. Dyn. Syst. Meas. Control 138 (2016) (2016) 111008.