Journal of Intelligent & Fuzzy Systems 35 (2018) 5099–5109 5099

DOI:10.3233/JIFS-169794
IOS Press

sEMG signal classification with novel

feature extraction using different machine
learning approaches

PY
Yogendra Narayan*, Lini Mathew and S. Chatterji
Department of Electrical Engineering, National Institute of Technical Teachers Training and Research,
Panjab University, Chandigarh, India

CO
Abstract. Selection of suitable features plays a pivotal role in Electromyography pattern recognition (EMG-PR) based
system designing. Time-domain features are widely used in EMG-PR based application and show improved proficiency in
the development of rehabilitation robotics. Even though, the performance of existing features is not satisfactory. In this study,
we proposed four novel time-domain features obtained by using first-order differentiation of original surface electromyogram
OR
(sEMG) signals feature. Here, sEMG signals were acquired from ten healthy volunteers with the help of myotrace400 device
for six different arm movements. The data acquisition and pre-processing stage were carried out followed by the feature
extraction process for better classification results. Four different classifiers namely, k-nearest neighbors (KNN), Linear
discriminant analysis (LDA), Quadratic discriminant analysis (QDA) and Medium tree (MT) classifiers were utilized for
the performance evaluation of proposed and conventional features. Experimental results demonstrate that proposed features
extracted by using first-order differentiation of sEMG signals feature attained better classification accuracy with MT classifier
TH

as compared to the feature extracted from original sEMG signals with the conventional features. The accuracy of proposed
feature based on first-order differentiation improved up to 6%. The results indicate that proposed features may be considered
for developing the EMG-PR based system designing.

Keywords: sEMG signal, pattern recognition, time domain features, differentiation technique, classification accuracy
AU

1. Introduction Recent development in EMG-PR based system

revealed that sEMG based control strategy has the
The sEMG signals are widely used as an input potential for controlling more DOF as compared to
signal for controlling the rehabilitation robots and conventional control strategies [5]. The EMG-PR
other myoelectric prostheses devices due to their based control strategy consists of sEMG acquisi-
non-invasive nature [1, 2]. The sEMG signals based tion (to obtain more accurate myoelectric signals),
conventional methods work well in some cases but feature extraction (to maintain the discriminating
have a limited degree of freedom (DOF) and user intu- information), classification (to predict one motion
itiveness for controlling the myoelectric prostheses among all motion) and generating the control com-
[3, 4]. mands for interfacing external world devices [6].
The performance of any EMG-PR system depends
∗ Corresponding author. Yogendra Narayan, Electrical Engi- on the attributes of the sEMG features extracted
neering Department, National Institute of Technical Teachers
and characterized by non-stationary properties of
Training and Research, Panjab University, Chandigarh 160019,
India. E-mail: [email protected]. sEMG signals [7]. Due to non-stationary nature of

1064-1246/18/$35.00 © 2018 – IOS Press and the authors. All rights reserved
5100 Y. Narayan et al. / sEMG signal classification with novel feature extraction

sEMG signal, extraction of the robust set of feature 2. Materials and methods
becomes difficult which can easily decode the arm
movements effectively for controlling purpose. Nev- 2.1. Data recording and pre-processing
ertheless, over the years, various efforts have been
made for the extraction of proper sets of features Ten healthy volunteers participated in the sEMG
so that movement classification accuracy can be data acquisition with the help of myotrace400 device
enhanced [8]. As a result, different features from and skin electrodes placed properly on specified
time, frequency and time scale domain have been pro- acupressure points [19]. For ensuring the good
posed for motion detection [9]. Among these, time electrode-skin contact, the subject skin was pre-
domain features can be extracted with less comput- pared with abrasive paste and alcohol wipes so that
ing resources and have a wide range of application impedance between the contacts would be minimum.
over the years [10, 11]. Each volunteer was asked to sit on the comfortable

PY
Traditional EMG-PR methods for the human- chair in a relaxing position and directed to follow the
machine interface have been evolved under the instructions displayed on the computer screen to fur-
presumption that sEMG signals are stationary [12]. ther record the data from different arm movements.
The sEMG signals are assumed stationary if their sta- A rest session was introduced between two consec-
tistical properties do not vary over time [13]. So, utive classes of arm motion for avoiding the mental

CO
the systems that are designed with the conjecture and muscle fatigue. Six consecutive trials were exe-
of stationarity would break down for sEMG signal cuted to acquire the data per volunteer. In each trial,
classification in dynamic portions. In modern EMG- every muscle movement contraction was maintained
PR, different time-frequency analysis methods like for 5 sec with 10 repetitions in order to get 50 sec
fast Fourier transform, discrete wavelet transforms sEMG signal recordings for each motion category.
(DWT) and empirical mode decomposition has been Consequently, a total of 300 sec sEMG recordings
suggested for studying the time-varying properties i.e. (50 sec × 6 trials) were accomplished for each
OR
of sEMG signals [14], Previously, the sEMG sig- movement in the experiment of six-trial on volunteer
nals were analyzed using first-order differentiation participant.
of sEMG signals; d(1) (t) as a part of time domain The volunteers performed the elbow flexion (EF)
methods [15]. The computation of time domain fea- elbow extension (EE), wrist flexion (WF), wrist
tures was achieved on the basis of statistics of d(1) (t), extension (WE), wrist supination (WS) and wrist
TH

in place of original sEMG signals; x(t). [16]. From pronation (WP) movements of their right arm. For
a statistical point of view, d(1) (t) and polynomial the elbow movement electrodes were placed on the
trend removal techniques are employed for enhancing biceps brachii, triceps brachii muscles whereas for
the signal compatibility and stationarity [17]. How- wrist motion electrodes were attached to pronator
ever, better classification accuracy was achieved with teres, supinator, flexor carpi radialis and extensor
AU

d(1) (t) features over x(t) features in previous studies carpi radialis muscles respectively. Figure 1 shows
[18]. the complete experimental setup and sEMG data
In this work, four-time domain features were pro- acquisition system for different arm movements. The
posed using d(1) (t) technique with an attempt to figure shows the myotrace400 device along with nec-
enhance the performance of EMG-PR based system. essary component and six different hand postures
The first objective of this work was to investigate used during the data recording.
the performance of various classifiers with features A sliding window analysis with a length of
extracted from d(1) (t) and x(t) whereas the second 150 msec was utilized for segmentation of combined
was to propose novel promising features (move- sEMG data sheet. The sliding window has an incre-
ment detector) based on d(1) (t) technique in place ment of 100 msec and overlapping of 50 msec for
of x(t). providing the continuous stream of data [20]. Fig-
This paper is organized as follows: the first section ure 2 shows the systematic operations for sEMG
describes the introduction of sEMG signal, the second signal classification and demonstrates that in the pre-
section represents the data recording, pre-processing processing stage, full wave rectification, amplitude
and involved classification methods, and the third sec- normalization, smoothing, and filtering operation
tion describes the results followed by conclusions in was performed followed by DWT de-noising [21].
the fourth section. After this, suitable features were extracted and d(1) (t)
 Y. Narayan et al. / sEMG signal classification with novel feature extraction 5101

PY
Fig. 1. Data Acquisition system and complete experimental set with different arm postures.

and (vi) waveform length (WL). The existing time

CO
domain features and their corresponding statistics on
the transformed signal are shown in Table 1. WAMP
and MYOP were computed with the threshold value
of 25 mvolt for the amplitude of sEMG signal to
avoid the noises.
The sEMG signals pattern is shown in Fig. 3 in
OR
Fig. 2. Block diagram of sEMG classification.
which data was recorded from six different limb
technique was applied to get novel features for the movements after the full-wave rectification process.
classification of sEMG signal to decode different The figure indicates that amplitude of sEMG signal
movements. For avoiding the noise, a bandpass fil- varies for different movement and it is due to the
ter of 20 to 500 Hz bandwidth was utilized [22]. The variation of the contraction force produced by the
muscles for different movement thereby producing
TH

signal pre-processing circuitry had CMRR greater

than 100, the amplifier gain of 500, input impedance the dissimilar values.
more than 100 M and baseline noise level less than
1␮Volt [23]. 2.3. Proposed features using d(1) (t)
AU

2.2. Feature extraction The pre-processed and de-noised signals were

employed for features extraction to make feature vec-
Time statistics based features were vastly tor for classification purpose. One commonly used
employed in the field of bio-medical robotics which and three latest features that have not formerly been
provides better accuracy as compared to the fre- applied to d(1) (t), were evaluated with four different
quency domain and time-frequency domain features. classification methods. These features are Absolute
Time domain features computed from original value of the Summation of the Square root (ASS),
sEMG signal, x(t), are: (i) myopulse percentage rate Mean value of the Square Root (MSR), Absolute
(MYOP), (ii) root mean square (RMS), (iii) variance value of the Summation of the expth root of data
of EMG (VAR), (iv) simple square integral (SSI) and its Mean (ASM) and modified V-order whose
(v) Mean absolute value (MAV) and (vi) integrated mathematical definitions are as follows:
 
EMG (IEMG), whereas corresponding matched  k 1
time domain features computed from d(1) (t), are: (i) − ASS =  (xn ) 2  (1)
n=1
willison amplitude (WAMP), (ii) difference absolute
standard deviation value (DASDV), (iii) difference where, xn is analysis window data.
variance value (DVARV), (iv) second order moment 1 k 1
(M2 ), (v) difference absolute mean value (DAMV) − MSR = (xn ) 2 (2)
k n=1
5102 Y. Narayan et al. / sEMG signal classification with novel feature extraction

Table 1
Mathematical formulas of existing time domain features obtained from d(1) (t) and x(t)
Feature extracted from d(1) (t) Feature extracted from x(t)
N=1   N  
WAMP = f (|x (t + 1) − x (t)|) ; MYOP = 1
f (|x (t)|) ;
 t=1  N t=1
1&if a ≥ threshold 1&if a ≥ threshold
f (a) = f (a) =
0
& otherwise 0& otherwise
N−1  
N
DASDV = 1
(x (t + 1) − x (t))2 RMS = 1
x(t)2
N−1t=1
N−1 N
 N
t=1
DVARV = 1
(x (t + 1) − x(t))
2 VAR = N−1
1
x(t)2
N−1N−2 t=1 N t=1
M2 = (x (t + 1) − x(t))2 SSI = x(t)2
t=1 N−1 N
t=1
DAMV = N−11
|x (t + 1) − x(t)| MAV = N1 |x(t)|
N−1 t=1 N t=1
WL = t=1
|x (t + 1) − x(t)| IEMG = t=1
|x(t)|

PY
CO
OR
TH
AU

Fig. 3. sEMG signals patterns for six different arm movements.

 
 k (x )exp  The newly proposed time domain features are
 n=1 n 
− ASM =   (3) based on the ASS, MSR, ASM and modified V-order
 k 
feature and obtained by applying the differentiation
where, exp = 0.5 if 0.75 ≥n ≥ 0.25*k, otherwise technique on it. The mathematical definitions of pro-
exp = 0.75 posed novel features are as follows:

1 k
1
5 – Difference Absolute value of the Summation of
− ModifiedV − order = (xn )5
k n=1 the Square root (DASS) is obtained from the
(4) absolute value of the summation of the square
 Y. Narayan et al. / sEMG signal classification with novel feature extraction 5103

1 k
root using a differentiation technique and it is 1
5
expressed by Equation (5) DMV − order = (xn − xn−1 )5 (8)
k n=1

  It is calculated by taking the first-order difference

 k 1 of modified V-order feature. DMV-order is deter-

DASS =  (xn − xn−1 ) 
2 (5)
n=1 mined by the few steps: first, the difference of all
value is calculated by subtracting the successive value
where k is the size of window and xn is correspond-
from previous one. Secondly, the summation of all
ing analysis window data. The feature was computed
value with 5th power and the 5th root of the resultant
in four steps: first subtracting the successive value
are calculated. Finally, the mean value of resultant is
from previous one in the analysis window, second is
determined.
square root calculation for all values, third is com-
puting the summation of all value and fourth is the

PY
2.4. sEMG classification
absolute value of summation.
– Difference Mean value of the Square Root The quality evaluation of the sEMG signal fea-
(DMSR) is the second time domain feature pro- tures was done on the basis of classification rates and
posed which is acquired from the mean value of the ratio of the number of correctly classified sam-

CO
the square root using first-order differentiation ples to the number of misclassified classified samples
method and presented in Equation (6) (CCS/MCS) yielded by classifiers. Pervasively, the
performance of EMG-PR is measured in terms of
classification accuracy as the main index. Another
1 k 1 metric which defined the performance of the classifier
DMSR = (xn − xn−1 ) 2 (6)
k n=1 is the CCS/MCS ratio. Generally, the classification
accuracy can be defined as the ratio of correctly clas-
OR
Where, k is a number of analysis window and xn
sified sample to the total number of samples in a test
are analysis window data corresponding. DMSR is
set.
calculated in three steps: difference of all value is
calculated by subtracting the successive value from Truley classified samples
Accuracy = (9)
the previous one followed by the square root of all Total number of samples
values and mean of their resultants.
TH

The KNN, LDA, QDA and MT classifiers have

– Difference Absolute value of the Summation of been widely applied for the classification task of bio-
the expth root of data and its Mean (DASM) is also medical signals, therefore, KNN, LDA, QDA and MT
extracted by using the first-order differentiation were used for classification of sEMG signals [24–26].
technique and shown in Equation (7) The popular and four proposed features obtained
AU

from the sEMG signals were tested on the classifiers.

The classifiers are trained and tested for combined
 
 k (x − x )exp  data set collected from all subjects. Different fold
 n=1 n n−1 
DASM =   (7) cross validation (FCV) varied from two to ten, were
 k 
carried out for each feature for dividing the data into
Where exp = 0.5 if 0.75 ≥n ≥ 0.25*k, training and testing sets. The accuracy of existing and
Otherwise, exp = 0.75 proposed features was critically compared with each
Based on the sEMG signal, the values of exp can be other for finding the usefulness of proposed features.
either 0.5 or 0.75. The feature is determined in four
steps: the difference of all value found is determined
as calculated in the previous three cases followed by 3. Result and discussion
the addition of all exp root of the window, mean of
resultant value and the absolute value of the resultant The performance evaluation of different classi-
mean. fiers was investigated through the features extracted
from d(1) (t) and x(t). Here, the results are com-
– Difference of modified V-order (DMV-order) is pared in order to evaluate the performance of four
the fourth new feature and defined by Equation proposed novel features in contrast with some con-
(8) ventional features on the basis of four commonly
5104 Y. Narayan et al. / sEMG signal classification with novel feature extraction

Table 2
Average classification accuracies (%) of the existing feature scale based on d(1) (t) and x(t) using KNN classifier
Feature 2FCV 3FCV 4FCV 5FCV 6FCV 7FCV 8FCV 9FCV 10FCV Mean (±SD)
WAMP d(1) (t) 71.1 74.2 75.3 76.7 76.6 72.2 76.2 77.7 77.7 75.3 (±2.4)
DASDV d(1) (t) 78.3 79.8 78.9 78.7 79.9 77.8 75.6 78.6 79.8 78.6 (±1.3)
DVARV d(1) (t) 77.9 77.9 75.6 79.5 79.7 81.6 80.1 79.7 81.0 79.2 (±1.8)
M2 d(1) (t) 75.1 77.1 77.6 78.7 78.1 79.6 79.2 79.3 78.8 78.2 (±1.4)
DAMV d(1) (t) 77.9 48.3 77.1 79.5 76.5 78.8 78.9 78.8 78.8 75.0 (±10.0)
WL d(1) (t) 74.4 76.8 77.1 76.6 76.9 76.4 75.6 75.6 76 76.2 (±0.9)
Overall average classification accuracy of all feature based on d(1) (t) 77.1 (±3.0)
RMS x(t) 75.8 76.9 76.5 77.7 77.5 76.5 76.8 76.9 77.0 76.8 (±0.6)
VAR x(t) 75.3 75.5 76.3 76.5 75.3 75.6 76.5 76.6 77.2 76.0 (±0.7)
SSI x(t) 74.3 75.6 72.8 77.1 75.1 74.8 77.5 75.9 74.9 75.3 (±1.4)
MAV x(t) 71.8 76.0 76.1 76.9 75.7 76.4 75.8 75.2 75.5 75.4 (±1.5)
IEMG x(t) 64.8 64.2 64.8 64.6 65.6 64.9 66.8 65.7 66.2 65.3 (±0.8)

PY
MYOP x(t) 69.2 68.1 67.9 66.9 67.2 67.8 67.5 69.0 67.7 67.9 (±0.8)
Overall average classification accuracy of all feature based on x(t) 72.8 (±1.0)

Table 3
Average classification accuracies (%) of the existing feature scale based on d(1) (t) and x(t) using LDA classifier

CO
Feature 2FCV 3FCV 4FCV 5FCV 6FCV 7FCV 8FCV 9FCV 10FCV Mean (±SD)
WAMP d(1) (t) 70.9 71.0 70.5 70.5 70.6 71.6 70.8 70.1 70.4 70.7 (±0.4)
DASDV d(1) (t) 75.2 75.9 75.8 76.4 75.2 74.0 77.1 77.2 76.2 75.9 (±1.0)
DVARV d(1) (t) 71.9 74.3 72.9 73.9 75.2 75.9 72.9 74.8 72.1 73.8 (±1.4)
M2 d(1) (t) 72.8 73.5 74.2 74.5 76.2 73.2 73.7 74.1 73.7 74.0 (±1.0)
DAMV d(1) (t) 70.8 70.3 70.8 70.4 70.9 70.0 70.8 70.7 70.1 70.5 (±0.3)
WL d(1) (t) 68.3 68.1 67.3 68.5 68.8 68.7 68.6 68.3 68.6 68.4 (±0.5)
OR
Overall average classification accuracy of all feature based on d(1) (t) 72.2 (±0.8)
MYOP x(t) 67.2 67.7 66.7 67.4 67.3 67.0 66.8 67.1 66.8 67.1 (±0.3)
RMS x(t) 70.9 72.7 71.9 72.5 72.7 71.9 72.1 71.9 72.2 72.1 (±0.6)
VAR x(t) 70.5 69.5 70.3 69.4 69.7 70.2 69.9 69.4 69.6 69.8 (±0.4)
SSI x(t) 70.1 68.6 68.7 69.7 69.4 69.5 70.2 69.7 69.6 69.5 (±0.5)
MAV x(t) 67.8 68.1 68.5 68.2 68.5 68.4 68.6 68.2 68.1 68.3 (±0.3)
IEMG x(t) 62.0 61.8 61.3 61.5 61.3 61.4 61.9 61.1 61.5 61.5 (±0.3)
TH

Overall average classification accuracy of all feature based on x(t) 69.6 (±0.4)

used classification methods viz. KNN, LDA, QDA as compared to the features extracted from x(t). Out
and MT classifiers. The conventional time domain of all classifiers, the MT classifier demonstrated the
AU

features extracted from the original sEMG signal; best performance and the average classification accu-
x(t), are: (i) MYOP, (ii) RMS, (iii) VAR, (iv) SSI, racy of 77.5% were achieved in case of conventional
(v) MAV, and (vi) IEMG and their corresponding features extracted with d(1) (t) technique. The accu-
first-order differentiation of time domain features racy was increased by 6% in comparison with x(t)
extracted from the original sEMG signal; d(1) (t), are and the KNN classifier was found the second best
: (i) WAMP, (ii) DASDV, (iii) DVARV,(iv) M2, (v) method. The WAMP feature demonstrated the best
DAMV, and (vi) WL. Now the accuracy of these con- accuracy with the average classification accuracy
ventional features are compared with the accuracy of 78.6% using MT classifier followed by DAMV,
of four proposed novel features namely as DASS, DVARV, DASDV, M2 and WL with 78.1%, 78%,
DASM, DMSR, and DMV-order procured from the 77.6%, 77.5% and 74.9% respectively. Therefore, the
features namely as ASS, ASM, MSR and MV-order features from d(1) (t) are suggested instead of x(t)
respectively. features. The results also clarified that the features
Tables 2–5 showed the average classification accu- extracted from d(1) (t), raised the accuracy by 6%
racies of conventional features extracted from d(1) (t) because it makes the signal more stationary compared
and x(t). The computation of accuracy was done by to x(t). Figure 4 shows the average classification accu-
averaging the classification accuracy of all FCV var- racies of KNN, LDA, QDA and MT classifier for
ied from two to ten. The result indicates that the individual conventional features which is a graphical
features extracted from d(1) (t) have better accuracy representation of Tables 2–5.
 Y. Narayan et al. / sEMG signal classification with novel feature extraction 5105

Table 4
Average classification accuracies (%) of the existing feature scale based on d(1) (t) and x(t) using QDA classifier
Feature 2FCV 3FCV 4FCV 5FCV 6FCV 7FCV 8FCV 9FCV 10FCV Mean (±SD)
WAMP d(1) (t) 69.8 71.1 69.9 71.5 70.1 69.9 70.2 69.5 69.9 70.2 (±0.6)
DASDV d(1) (t) 74.4 74.5 73.9 74.2 74.0 74.9 74.3 74.4 74.4 74.3 (±0.3)
DVARV d(1) (t) 76.8 76.1 75.8 75.8 76.4 76.0 76.2 76.4 74.6 76.0 (±0.6)
M2 d(1) (t) 74.8 75.5 75.0 77.6 75.6 74.6 74.3 74.4 75.3 75.2 (±1.0)
DAMV d(1) (t) 71.6 72.2 72.0 72.6 70.9 71.1 72.1 71.7 72.9 71.9 (±0.7)
WL d(1) (t) 68.9 68.6 68.2 68.2 68.5 68.2 68.3 68.7 68.9 68.5 (±0.3)
Overall average classification accuracy of all feature based on d(1) (t) 72.7 (±0.6)
MYOP x(t) 66.7 66.7 66.5 66.5 66.8 66.5 66.5 66.6 66.7 66.6 (±0.1)
RMS x(t) 69.3 70.1 70.7 70.9 70.7 70.9 70.3 72.4 72.1 70.8 (±1.0)
VAR x(t) 61.0 70.3 71.7 71.6 71.1 71.4 71.9 71.1 71.6 70.2 (±3.5)
SSI x(t) 68.6 71.2 71.0 71.7 71.8 71.1 71.6 72.0 71.5 71.2 (±1.0)
MAV x(t) 67.8 68.1 68.7 68.2 68.0 67.8 68.8 68.3 68.7 68.3 (±0.4)

PY
IEMG x(t) 63.3 64.1 63.0 63.0 62.8 63.2 63.3 63.1 63.5 63.3 (±0.4)
Overall average classification accuracy of all feature based on x(t) 68.4 (±1.1)

Table 5
Average classification accuracies (%) of the existing feature scale based on d(1) (t) and x(t) using MT classifier

CO
Feature 2FCV 3FCV 4FCV 5FCV 6FCV 7FCV 8FCV 9FCV 10FCV Mean (±SD)
WAMP d(1) (t) 77.8 78.0 79.8 78.8 77.9 77.7 79.8 77.9 79.6 78.6 (±1.0)
DASDV d(1) (t) 75.1 78.2 78.6 72.3 77.6 79.6 79.5 78.7 78.5 77.6 (±2.4)
DVARV d(1) (t) 78.3 76.6 77.5 78.6 78.2 76.6 79.8 77.7 79.2 78.0 (±1.1)
M2 d(1) (t) 76.3 77.4 78.1 78.7 77.7 77.4 77.5 77.9 78.0 77.5 (±0.5)
DAMV d(1) (t) 76.6 77.9 78.6 77.9 77.5 79.5 77.6 78.7 78.4 78.1 (±0.8)
WL d(1) (t) 75.3 73.8 76.6 75.1 74.0 74.8 74.7 73.9 75.5 74.9 (±1.0)
OR
Overall average classification accuracy of all feature based on d(1) (t) 77.5 (±1.1)
MYOP x(t) 65.9 65.9 66.1 65.9 66.9 67.1 67.7 66.7 66.9 66.6 (±0.6)
RMS x(t) 72.7 75.1 74.9 74.4 74.2 73.4 75.2 74.2 74.8 74.3 (±0.8)
VAR x(t) 76.0 75.2 76.4 76.5 75.2 75.8 72.3 71.9 73.5 74.8 (±1.8)
SSI x(t) 72.1 74.9 74.2 74.9 74.8 76.1 76.5 74.1 74.2 74.6 (±1.3)
MAV x(t) 71.5 74.6 71.5 72.1 71.9 73.7 73.8 74.9 74.3 73.1 (±1.4)
IEMG x(t) 63.9 66.6 65.8 64.3 65.5 64.4 65.6 67.9 69.6 66.0 (±1.8)
TH

Overall average classification accuracy of all feature based on x(t) 71.6 (±1.3)
AU

Fig. 4. Individual classification accuracies of the different classifier for conventional features.

Tables 6–9 exhibited the average classification indicated that MT classifier was again found best with
accuracies of proposed features extracted from d(1) (t) 84.9% average classification accuracy by utilizing
and x(t) with the same classifiers. The results the proposed feature obtained with d(1) (t) technique
5106 Y. Narayan et al. / sEMG signal classification with novel feature extraction

Table 6
Average classification accuracies (%) of the proposed feature based on d(1) (t) and x(t) using KNN classifier
Feature 2FCV 3FCV 4FCV 5FCV 6FCV 7FCV 8FCV 9FCV 10FCV Mean (±SD)
DASS d(1) (t) 85.1 85.7 86.6 87.9 88.7 88.2 87.5 86.3 82.1 86.5 (±2.0)
DMSR d(1) (t) 84.4 85.5 85.2 85.8 86.3 86.5 84.9 85.9 86.6 85.6 (±0.8)
DASM d(1) (t) 84.6 87.3 85.5 85.4 87.3 87.1 86.1 85.7 88.3 84.4 (±1.2)
DMV-order d(1) (t) 78.3 78.0 79.3 79.0 79.1 77.9 79.7 78.5 77.7 78.7 (±0.6)
Overall average classification accuracy of all feature based on d(1) (t) 84.3 (±1.2)
ASS x(t) 79.5 81.2 81.7 83.1 82.4 83.3 84.2 83.1 82.8 82.4 (±1.4)
MSR x(t) 79.1 81.4 81.5 82.7 82.7 82.9 83.3 80.6 82.8 81.9 (±1.4)
ASM x(t) 83.1 82.8 83.2 83.3 83.6 82.6 82.8 83.8 83.9 83.2 (±0.5)
MV-order x(t) 75.7 75.1 76.4 75.4 77.5 76.0 76.0 76.6 76.6 76.1 (±0.7)
Overall average classification accuracy of all feature based on x(t) 80.9 (±1.3)

PY
Table 7
Average classification accuracies (%) of the proposed feature based on d(1) (t) and x(t) using LDA classifier
Feature 2FCV 3FCV 4FCV 5FCV 6FCV 7FCV 8FCV 9FCV 10FCV Mean (±SD)
DASS d(1) (t) 76.8 78.1 77.2 77.2 77.0 75.5 77.3 77.4 77.9 77.2 (±0.7)
DMSR d(1) (t) 78.3 78.6 78.0 78.3 78.1 78.1 78.3 78.1 79.0 78.3 (±0.3)

CO
DASM d(1) (t) 78.8 78.9 78.9 78.8 79.3 79.7 78.8 75.9 78.6 78.6 (±1.1)
DMV-order d(1) (t) 74.8 73.5 76.4 75.1 74.9 75.8 74.7 75.6 77.0 75.3 (±1.0)
Overall average classification accuracy of all feature based on d(1) (t) 77.5 (±0.8)
ASS x(t) 73.7 73.4 73.9 74.1 73.9 74.5 73.4 73.9 74.0 73.9 (±0.3)
MSR x(t) 75.9 74.3 74.1 74.1 73.5 73.8 73.8 74.0 74.1 74.2 (±0.7)
ASM x(t) 73.0 73.7 73.8 73.2 72.6 73.6 73.7 73.8 73.6 73.4 (±0.4)
MV-order x(t) 69.4 69.0 69.1 69.8 70.6 69.9 70.0 69.7 69.1 69.6 (±0.5)
Overall average classification accuracy of all feature based on x(t) 72.8 (±0.5)
OR

Table 8
Average classification accuracies (%) of the proposed feature based on d(1) (t) and x(t) using QDA classifier
Feature 2FCV 3FCV 4FCV 5FCV 6FCV 7FCV 8FCV 9FCV 10FCV Mean (±SD)
DASS d(1) (t) 79.1 78.8 78.6 78.7 79.1 78.1 78.5 78.3 79.0 78.7 (±0.4)
d(1) (t)
TH

DMSR 78.1 75.8 76.6 75.4 75.9 76.2 75.8 76.3 76.6 76.3 (±0.8)
DASM d(1) (t) 80.1 80.4 80.8 79.0 79.1 78.8 79.6 80.1 79.3 79.7 (±0.7)
DMV-order d(1) (t) 76.3 76.0 75.2 75.8 75.8 75.2 75.3 76.1 76.7 75.8 (±0.5)
Overall average classification accuracy of all feature based on d(1) (t) 77.6 (±0.6)
ASS x(t) 74.3 74.4 73.3 73.5 73.2 72.6 73.9 73.2 74.8 73.7 (±0.7)
MSR x(t) 75.9 74.0 75.2 72.9 73.0 73.4 73.9 73.5 72.8 73.9 (±1.1)
AU

ASM x(t) 75.7 75.1 73.9 73.1 74.5 74.6 74.8 74.3 74.2 74.5 (±0.7)
MV-order x(t) 72.1 72.0 72.2 72.6 72.1 72.2 72.2 71.8 72.4 72.2 (±0.2)
Overall average classification accuracy of all feature based on x(t) 73.6 (±0.7)

Table 9
Average classification accuracies (%) of the proposed feature based on and x(t) using MT classifier
Feature 2FCV 3FCV 4FCV 5FCV 6FCV 7FCV 8FCV 9FCV 10FCV Mean (±SD)
DASS d(1) (t) 86.3 86.2 86.5 84.7 86.0 86.5 88.5 88.1 87.9 86.7 (±1.2)
DMSR d(1) (t) 43.1 82.8 83.5 86.4 88.4 86.2 87.3 86.2 86.9 85.8 (±1.9)
DASM d(1) (t) 89.6 86.2 84.4 85.9 89.0 88.5 87.6 90.1 88.7 87.8 (±1.9)
DMV-order d(1) (t) 78.5 79.6 78.6 80.0 79.1 79.1 78.0 78.9 78.9 79.0 (±0.6)
Overall average classification accuracy of all feature based on d(1) (t) 84.9 (±1.4)
ASS x(t) 79.6 78.9 78.8 78.6 79.6 80.5 80.9 80.7 82.0 80.0 (±1.1)
MSR x(t) 76.2 80.1 77.3 80.2 80.5 81.4 81.2 81.7 81.4 80.0 (±1.9)
ASM x(t) 80.4 81.8 81.1 80.8 80.7 83.7 82.8 82.6 81.2 81.7 (±1.1)
MV-order x(t) 72.6 72.9 72.1 72.2 73.0 73.7 73.9 72.9 71.9 72.8 (±0.7)
Overall average classification accuracy of all feature based on x(t) 78.6 (±1.2)
 Y. Narayan et al. / sEMG signal classification with novel feature extraction 5107

PY
Fig. 5. Individual classification accuracies of the different classifier for original and proposed features.

Features KNN
Table 10
CCS/MCS value based on the conventional time domain and
proposed features using KNN, LDA, QDA and MT classifiers
LDA QDA MT
CO Another important parameter to evaluate the per-
formance of conventional and proposed time-domain
features is CCS/MCS ratio which simply reflects the
robustness of methods for correctly distinguishing
ASS 5.46 4.28 5.72 5.67 the different motion classes of the sEMG signals.
MSR 4.27 4.48 4.12 5.24
In this context, if the ratio is high then the fea-
OR
ASM 6.36 6.17 5.98 6.74
MV-order 4.08 3.87 3.78 4.12 ture will improve the classification accuracy to a
Feature set 1 30.68 28.72 27.42 35.66 higher value. In addition, the CCS/MCS ratios were
DASS 6.15 5.27 4.98 6.34
computed for the different combination of ASS,
DMSR 6.92 6.35 5.78 6.97
DASM 6.64 6.15 6.32 6.75 ASM, MSR and MV-order as well as DASS, DASS,
DMV-order 4.48 4.77 4.58 4.82 DMSR, and DMV-order. The features combination
TH

Feature set 2 40.36 38.15 38.45 45.54 of ASS, MSR, ASM, and MV-order were named as
4-TD 6.78 6.87 6.77 6.75
Feature Set 1 whereas the proposed features com-
bination of DASS, DMSR, DASM and DMV-order
were named as Feature Set 2. Similarly, a conjunc-
and accuracy raised more than 6% as compared to tion of the four commonly used time domain features
AU

x(t). Similarly, the second best classification method (SSC+RMS+WL+MAV) were also examined and
was KNN with the average classification accuracy of referred as 4-TD features. The CCS/MCS ratios were
78.6%. The DASM feature revealed the best accuracy calculated for KNN, LDA, QDA and MT classifiers
with the average classification accuracy of 87.8% by for different combinational features. In Table 10, the
utilizing MT classifier followed by DASS, DMSR CCS/MCS ratio for DMSR feature obtained as 6.97
and DMV-order with 86.7%, 85.8% and 79% respec- which was higher as compared to the 4-TD feature set
tively. The results clearly indicated the usefulness of whereas DASM feature provides an equivalent result
proposed features because the average classification for the CCS/MCS ratio to the 4-TD feature set. While
accuracy was enhanced more than 6% by employing the feature set 1 yielded the good CCS/MCS ratio
the d(1) (t) technique in contrast with x(t). Therefore, which was 35.66 whereas the feature set 2 produced
it is worth mentioning here that proposed features the best CCS/MCS ratio of 45.54. The MT classifier
are effective enough for improving the classification performed the best among the LDA, QDA, and MT
accuracy with MT classifier. Figure 5 shows the aver- classifier. So, we can further conclude that among all
age classification accuracies of KNN, LDA, QDA and methods, MT classifier showed the best classification
MT classifier with proposed features extracted from accuracy with proposed features based on d(1) (t) tech-
d(1) (t) and x(t). The figure indicates that proposed fea- nique. It can be easily demonstrated with the help of
tures have better classification accuracy as compared Fig. 6 showing the relationship between CCS/MCS
to any other features. ratio Vs different features for MT classifier. It is a
5108 Y. Narayan et al. / sEMG signal classification with novel feature extraction

PY
higher performance of MT classifier for the proposed
CO
Fig. 6. CCS/MCS ratio of different features with MT classifier.

graphical representation of Table 10 indicating the feature as well as more prominent than conventional
features. The proposed features can be utilized in var-
feature set as compared to the conventional feature ious EMG applications like human-machine interface
set. design and disease identification.
OR
It is clearly verified from the above discussion that In near future, sEMG data from amputee’s persons
d(1) (t) technique promotes the classification accuracy. can be recorded with the combination of conven-
Another worth mentioning point is that different com- tional and newly proposed features for developing the
bination of proposed novel feature always yielded rehabilitation robotics system along with enhanced
better result as compared to the single feature. So the properties. So, in the field of pattern recognition, the
TH

features combination of DASS, DMSR, DASM and better predictions will be desired for higher DOF
DMV-order should be used for better performance robotic system design based on sEMG signals.

References
AU

4. Conclusion
[1] I. Batzianoulis, S. El-Khoury, E. Pirondini, M. Coscia,
Classification accuracy has immersed a lot of S. Micera and A. Billard, EMG-based decoding of grasp
research attention in the field of rehabilitation gestures in reaching-to-grasping motions, Robotics and
Autonomous Systems 91 (2017), 59–70.
robotics for developing the assistive technology. This [2] D. Karabulut, F. Ortes, Y.Z. Arslan and M.A. Adli, Com-
study suggested four novel time-domain features parative evaluation of EMG signal features for myoelectric
extracted from the sEMG signal of healthy sub- controlled human arm prosthetics, Biocybernetics and
jects for improving the classification performance. Biomedical Engineering 37(2) (2017), 326–335.
[3] J. Ma, N.V. Thakor and F. Matsuno, Hand and wrist
The performance evaluation of proposed features movement control of myoelectric prosthesis based on syn-
was carried out by using classification accuracy and ergy, IEEE Transactions on Human-Machine Systems 45(1)
CCS/MCS metrics. The results suggest that pro- (2015), 74–83.
posed movement detectors have good potential for [4] L.J. Hargrove, G. Li, K.B. Englehart and B.S. Hudgins,
Principal components analysis preprocessing for improved
the improvement of EMG-PR system with MT clas- classification accuracies in pattern-recognition-based myo-
sification method. Notwithstanding, the results of electric control, IEEE Transactions on Biomedical Engi-
this study may hold for elderly or amputees per- neering 56(5) (2009), 1407–1414.
[5] L. Pan, D. Zhang, X. Sheng and X. Zhu, Improving myoelec-
sons, Finally, EMG feature extracted by using the
tric control for amputees through transcranial direct current
first-order differentiation of sEMG signals are found stimulation, IEEE Transactions on Biomedical Engineering
more dominant as compared to original sEMG signal 62(8) (2015), 1927–1936.
 Y. Narayan et al. / sEMG signal classification with novel feature extraction 5109

[6] O.W. Samuel, H. Zhou, X. Li, H. Wang, H. Zhang, A.K. San- [17] T. Lorrain, N. Jiang and D. Farina, Influence of the train-
gaiah and G. Li, Pattern recognition of electromyography ing set on the accuracy of surface EMG classification in
signals based on novel time domain features for amputees’ dynamic contractions for the control of multifunction pros-
limb motion classification, Computers & Electrical Engi- theses, Journal of Neuroengineering and Rehabilitation
neering (2017). 8(1) (2011), 25.
[7] R.N. Khushaba, M. Takruri, J.V. Miro and S. [18] A. Phinyomark, F. Quaine, S. Charbonnier, C. Serviere,
Kodagoda.Towards limb position invariant myoelec- F. Tarpin-Bernard and Y. Laurillau, Feature extraction of
tric pattern recognition using time-dependent spectral the first order difference of EMG time series for EMG
features, Neural Networks 55 (2014), 42–58. pattern recognition, Computer Methods and Programs in
[8] K. Xing, P. Yang, J. Huang, Y. Wang and Q. Zhu, A real-time Biomedicine 117(2) (2014), 247–256.
EMG pattern recognition method for virtual myoelectric [19] H.S. Ryait, A.S. Arora and R. Agarwal, SEMG signal
hand control, Neurocomputing 136 (2014), 345–355. analysis at acupressure points for elbow movement, Jour-
[9] A. Phinyomark, F. Quaine, S. Charbonnier, C. Serviere, F. nal of Electromyography and Kinesiology 21(5) (2011),
Tarpin-Bernard and Y. Laurillau, EMG feature evaluation 868–876.
for improving myoelectric pattern recognition robust- [20] X. Li, S. Chen, H. Zhang, O.W. Samuel, H. Wang, P.

PY
ness, Expert Systems with Applications 40(12) (2013), Fang, X. Zhang and G. Li, Corrigendum to Corrigendum to
4832–4840. Towards reducing the impacts of unwanted movements on
[10] P. Geethanjali and K.K. Ray, A low-cost real-time research identification of motion intentions, [J Electromyogr Kine-
platform for EMG pattern recognition-based prosthetic siol 28 (2016) 90–98]. Journal of Electromyography and
hand, IEEE/ASME Transactions on Mechatronics 20(4) Kinesiology: Official Journal of the International Society of
(2015), 1948–1955. Electrophysiological Kinesiology 33 (2017), 118.
[11] A. Phinyomark, P. Phukpattaranont and C. Limsakul, A [21] S. Gupta and S. Kumar, Variational level set formulation and

CO
review of control methods for electric power wheelchairs filtering techniques on ct images, International Journal of
based on electromyography signals with special emphasis Engineering Science and Technology (IJEST) 4(07) (2012),
on pattern recognition, IETE Technical Review 28(4) (2011), 3509–3513.
316–326. [22] K. Shallu, Sumit and E. Aggarwal, A study on adaptive
[12] K. Nazarpour, A.R. Sharafat and S.M.P. Firoozabadi, Appli- wavelet technique for speckle noise removal, Commu-
cation of higher order statistics to surface electromyogram nication and Computing Systems: Proceedings of the
signal classification, IEEE Transactions on Biomedical International Conference on Communication and Com-
Engineering 54(10) (2007), 1762–1769. puting Systems (ICCCS 2016), Gurgaon, India, 9-11
OR
[13] M.A. Mañanas, M. Guillen, J.A. Fiz, J. Morera and P. September, 2016. CRC Press, 2017.
Caminal, Analysis of stationarity and statistical changes in [23] Y. Narayan, L. Mathew and S. Chatterji, sEMG signal clas-
myographic signals from respiratory muscles. Engineering sification using Discrete Wavelet Transform and Decision
in Medicine and Biology Society, 2000, Proceedings of the Tree classifier, International Journal of Control Theory and
22nd Annual International Conference of the IEEE. Vol. 2. Applications 10(6) (2017), 511–517.
IEEE, 2000. [24] KS. Kim, H.H. Choi, C.S. Moon and C.W. Mun, Com-
[14] A. Phinyomark, P. Phukpattaranont and C. Limsakul, The parison of k-nearest neighbor, quadratic discriminant and
TH

usefulness of wavelet transform to reduce noise in the linear discriminant analysis in classification of electromyo-
SEMG signal. EMG methods for evaluating muscle and gram signals based on the wrist-motion directions, Current
nerve function. InTech, 2012. Applied Physics 11(3) (2011), 740–745.
[15] R.H. Chowdhury, M. BI Reaz, M.A.B. Mohd Ali, A.A.A. [25] T. Lajnef, S. Chaibi, P. Ruby, P.-E. Aguera, J.-B. Eichenlaub,
Bakar, K. Chellappan and T.G. Chang, Surface electromyo- M. Samet, A. Kachouri and K. Jerbi, Learning machines and
graphy signal processing and classification techniques, sleeping brains: Automatic sleep stage classification using
AU

Sensors 13(9) (2013), 12431–12466. decision-tree multi-class support vector machines, Journal
[16] P. McCool, L. Petropoulakis, J.J. Soraghan and N. Chat- of Neuroscience Methods 250 (2015), 94–105.
lani, Improved pattern recognition classification accuracy [26] E. Gokgoz and A. Subasi, Comparison of decision tree
for surface myoelectric signals using spectral enhance- algorithms for EMG signal classification using DWT,
ment, Biomedical Signal Processing and Control 18 (2015), Biomedical Signal Processing and Control 18 (2015),
61–68. 138–144.