Page 1 of 26 IET Science, Measurement & Technology

This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

Analysis of ALS and normal EMG signals based on empirical mode

decomposition
a
Vipin K Mishra, aVarun Bajaj, aAnil Kumar, bG. K. Singh
a
Discipline of Electronics and Communication Engineering, PDPM Indian Institute of Information

Technology, Design and Manufacturing Jabalpur, India

b
Department of Electrical Engineering, Indian Institute Technology Roorkee, India

Email:[email protected], [email protected], [email protected],

[email protected]

Abstract—Electromyogram (EMG) signals contain a lot of information about the

neuromuscular diseases like amyotrophic lateral sclerosis (ALS). ALS progressively

degenerates the motor neurons in spinal cord. In this paper, a new technique for the analysis

of normal and ALS EMG signals is proposed. EMG signals are decomposed into narrow

band intrinsic mode functions (IMFs) by using empirical mode decomposition (EMD)

technique. The area of complex plot, two bandwidths namely amplitude modulation

bandwidth (BAM) and frequency modulation bandwidth (BFM), normalised instantaneous

frequency ( IFn ) , spectral momentum of power spectral density (SMPSD) and mean of first

derivative of instantaneous frequency (MFDIF) are extracted from analytic IMFs obtained by

EMD technique. These six features are used as input in least square support vector machine

(LS-SVM) classifier for the classification of ALS and normal EMG signals. Experimental

results and comparative analysis shows that classification performance of proposed method is

better than other existing method in the same database.

Keywords- Electromyogram (EMG) signal, EMG signals analysis, Amyotrophic lateral

sclerosis (ALS), Empirical mode decomposition (EMD).

IET Review Copy Only

 IET Science, Measurement & Technology Page 2 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

1. INTRODUCTION

Electromyography is a technique of collecting and processing electrical signals produced by

the contraction of skeletal muscles. These collected electrical signals are known as

electromyogram (EMG) signal [1]. Electrical signal magnitude is proportional to contraction

potential of the skeletal muscles. EMG signals are being used in many fields such as sports

science, ergonomics, medical research and rehabilitation. In medical science, the EMG

signals can be used for the analysis and identification of diseases related to voluntary muscles

and motor neurons [2]. Brain produces electrical impulses to control the action of muscles.

Motor neuron is a nerve cell, which forms the path for transfer of electrical impulses. Brain

need upper and lower motor neurons to send the signals to a particular muscle in the body [3].

To perform any action, upper motor neurons send signals to trigger lower motor neurons that

further transfer the signals to other part of body. Muscles present in the chest, legs, throat,

face, arms and tongue are controlled by the lower motor neurons [4]. Motor neuron disorder

creates a lot of diseases, especially amyotrophic lateral sclerosis (ALS) [5]. ALS results in

weakness, atrophy, paralysis, respiratory failure and loss of control of brain over muscles.

ALS affects both (upper and lower) motor neurons, due to which muscles cannot be

controlled voluntarily and become weaker and smaller [6] [7]. Autonomic impairment as

dysregulation of blood pressure, heart rate (HR) and increase in the risk of unusual death

have been found in many ALS patients [8]. The EMG signal has been a valuable clinical tool

to assess the changes of ALS. Diagnosis of ALS by visual inspection is a tedious task and

requires trained professionals and clinicians. Therefore, the accurate and automatic

methodology is required for diagnosis of ALS EMG signals.

Parameters extracted from the EMG signals are very useful in diagnostics. Most of the

parameters extracted from EMG signals are based on the time domain, frequency domain,

and time frequency domain. Time domain based features such as root mean square,

IET Review Copy Only

Page 3 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

Spectrogram, kurtosis, entropy and power have been used for classification of ALS and

healthy signals [9]. Mel-frequency cepstral coefficient (MFCC) based features have been

used as input to K-nearest neighbourhood (KNN) classifier for classification of ALS and

healthy EMG signals [10]. Discrete wavelet transform coefficients have been employed to

classify the healthy and ALS signals [11]. Time frequency based features such as

autocorrelation, zero crossing rate, and average value of spectral peak and mean frequency

have been applied to KNN classifier for classification of ALS and healthy EMG signals [12].

The power level of spectrum, frequency shifting to higher region and numbers of onset peak

features extracted from short time Fourier transform have been used for classification of ALS

and normal EMG signals [13].

Time based features, AR spectral measure, AR coefficients, cepstral coefficients feature sets

extracted from motor unit action potentials (MUAPs) have been used with neuro-fuzzy based

classifier to discriminate the normal, myopathy and neuropathy signals [14]. Cepstrum

analysis, which is popular in speech signal processing has been employed for classification of

MUAP based diseases [15]. The on-off timing of skeletal muscles at the time of movement,

continuous wavelet transform has been employed for analysis of EMG signals [16]. Various

features of signal amplitude, area, duration rise time, number of phases have been employed

for classification of EMG signals [17].

Optimal wavelet packet method with neural network classifier has been used for the

classification of surface EMG signals [18]. A method based on optimization of mother

wavelet has been used for the classification of EMG signals [19]. Method based on multiscale

entropy of EMG signal has been employed with support vector machine for the classification

of EMG signals [20]. Multiscale principle component analysis has been used with k-NN,

ANN and SVM classifier for the classification of EMG signals [21]. Multi fractal analysis on

EMG signals has been done to classify the different arm movements [22]. A method based on

IET Review Copy Only

 IET Science, Measurement & Technology Page 4 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

pattern recognition technique has been used for the discrimination of electromyographic

interference pattern into normal and myopathy [23].

In this paper, six new features namely area, BAM, BFM, SMPSD, MFDIF and normalized

instantaneous frequency are extracted from analytic IMFs for the classification of ALS and

normal EMG signals. The rest of the paper is organized as follows: The EMG dataset, the

empirical mode decomposition method, features extraction from IMFs, and LS-SVM

classifier are presented in Section 2. Experimental results for the classification of ALS and

normal EMG signals based on extracted features are given in Section 3. Section 4 presents

the discussion. Finally, Section 5 concludes the paper.

2. METHODOLOGY

2.1 DATASET

The dataset EMG signals are available online [24]. In normal group, 10 subjects comprising 6

males and 4 females of age 21-37 were present. No patient of normal group had any history

of neuro muscular disorders. ALS group covers 8 subjects: 4 males and 4 females’ of age

group 35-67 years. 5 subjects having the sign of ALS died few years after the occurrence of

disease. A standard needle electrode of leading off area 0.07 mm2 was used for the signal

acquisition. Sampling rate of EMG signals was 23437.5 Hz and digitized by A/D convertor of

16 bit resolution. Recorded signals were filtered at 2Hz and 10 kHz by high and low pass

filter respectively (i) To monitor the signal quality, video and audio feedbacks were used, (ii)

A standard concentric needle electrode was used, (iii) 2 KHz and 10 KHz range high and low

pass filters were used in EMG amplifier, (iv) Three level of needle insertion was used to

collect the signal from five places of a muscle, (v) The recording was conducted at constant

and low level of contraction. In this study, 87 ALS EMG signals and 133 normal EMG

IET Review Copy Only

Page 5 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

signals are used. The processing steps for the analysis of EMG signals are presented in the

Fig. 1.

EMG signal

Band pass filter

Empirical mode decomposition

…………..
IMF1 IMF2 IMFn

HT HT HT
…………..

Analytic IMF1 Analytic IMF2 Analytic IMFn

Features Extraction

kruskal-wallis statistical test and LS-SVM Classifier

ALS EMG Signal Normal EMG Signal

Fig. 1. Block diagram of the required processing steps for classification of ALS and normal

EMG signals.

2.2 Empirical mode decomposition

EMD is an adaptive signal decomposition method. It decomposes any non-stationary signal

into some set of band limited functions known as intrinsic mode function (IMFs). After

decomposition it maintains the original pattern of data, therefore it is a highly efficient

method [25]. For each IMF to be generated, two conditions must be satisfied: (i) the numbers

IET Review Copy Only

 IET Science, Measurement & Technology Page 6 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

of extrema in data set should be equal to or differ by at most one from the numbers of zero

crossing, and (ii) the mean value of both the envelopes (minima and maxima) should be zero

at any point. EMD algorithm follows these steps to break the signal into IMFs [26-27]:

1) Detect the minima and maxima of g0 ( t ) = x ( t ) .

2) Use cubic spline curve to connect extrema and generate lower and upper envelopes el ( t )

and eu ( t ) respectively.

3) Calculate the mean m(t) of upper and lower envelope and check for zero mean

el ( t ) + eu ( t )
m (t ) =
2

4) Find g1 ( t ) = g0 ( t ) − m ( t ) if mean is not zero.

5) Iterate process 1-4 till g n ( t ) becomes an IMF, (which can be checked by two conditions

illustrated above) c1 ( t ) = g n ( t ) .

6) Find residue r1 ( t ) = x ( t ) − c1 ( t ) . Repeat step 1-5 taking this residue as new signal to

generate other IMFs. Continue the process till the final residue is generated. Residue

contains one local maxima and minima, through this no more IMF can be obtained.

7) Finally, EMD produces n numbers of IMFs c1 ( t ) to c p ( t ) with a residual signal rM ( t ) .

Now, x ( t ) can be shown as

M
x(t ) = ∑ c p (t ) +r (t ) (1)
p =1
M

Where, M denotes the total number of IMFs. Hilbert transform is used to convert these

IMFs into analytic IMFs. An analytic IMF obtained from any IMF c(t) can be defined as:

jφ (t )
z (t ) = c(t ) + jcH (t ) = A ( t ) e (2)

Amplitude A(t), phase φ ( t ) and instantaneous frequency ω (t) of an analytic IMF can be

IET Review Copy Only

Page 7 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

defined as:

A(t) = c 2 (t ) + cH 2 (t ) (3)

 c H (t ) 
φ (t ) = arctan   (4)
 c (t ) 

dφ (t )
ω (t ) = (5)
dt

Instantaneous frequency is a measure of rate of rotation of analytic signal in complex plane.

In proposed method Hilbert transform has been applied on all the IMFs to get analytic IMFs.

Two parameters instantaneous frequency and instantaneous amplitude are extracted from

these analytic IMFs. Localization of instantaneous frequency of IMFs is good in time

frequency domain and useful to extract some essential features of signal. Each IMF can be

represented in analytic form as

jϕ p ( t )
z p ( t ) = Ap ( t ) e (6)

2.3 Area computation from analytic IMFs

Plot of an analytic IMF is a circular structure having unique center [28]. For fast

summarization of visual information in the graph, center tendency measurement (CTM) is a

useful method [29]. In this paper, CTM is used to get the radius (r) from circular plot of

analytic IMFs. Imaginary part of analytic signal is plotted against the real part of signal.

Analytic plot of IMFs fulfill these two conditions: Plot should have rotation direction, and a

fixed center. EMG signals are non-stationary and complicated in nature, and do not fulfill

above conditions. EMD algorithm is used to break any complicated EMG signal into numbers

of IMFs, which satisfies above conditions. Plot of Analytic IMFs can be used to calculate

area feature [29]. Area is calculated by using 95% radius of the circular plot. Variability of

the signal is quantified by CTM. It is computed by selecting all the points coming inside the

IET Review Copy Only

 IET Science, Measurement & Technology Page 8 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

radius r and dividing it by total number of points [30]. The CTM for analytic IMF can be

expressed as:

∑δ
n =1
d ( n)
CTM = (7)
N

where, N = Total number of points, and

1 If

δd ( n) =  ( R  z ( n) + I  z ( n) ) < r
2 2
(8)
0 otherwise

CTM represents a fraction of the total number of points lying inside the radius r.

2.4 Normalised instantaneous frequency computation from analytic IMFs

Normalised first derivative of instantaneous frequency can be represented as [31].

δf (9)
IFn =
f

where, δ f = diff ( f ) (modulus of first derivative of instantaneous frequency), IFn is

normalized by the mean value of instantaneous frequency; therefore, it is independent of

carrier frequency. Frequency of any single tone signal is constant, due to which after

normalization process, the mean value of instantaneous frequency difference ( δ f ) should be

small.

2.5 Bandwidth Feature:

Bandwidth of any signal is a measure of extent of frequency. This feature provides

information about the share of amplitude and frequency modulated bandwidth on whole

bandwidth of any signal. Aim of this feature is to provide the information, whether the spread

of frequency in EMG signals is due to the amplitude modulation (AM), or frequency

IET Review Copy Only

Page 9 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

modulation (FM), or due to both [32]. Share of frequency due to AM and FM in total

bandwidth of signal can be written as:

1   dA ( t ) 
2
 dϕ ( t ) 
2
 (10)
B2 = ∫   dt + ∫  − ω  A2 ( t ) dt 
E   dt   dt  


1 dϕ ( t ) 2
A ( t ) dt . Above Eq. 10 has two parts: first is dependent on amplitude and
E ∫ dt
Here ω =

second on the phase of a signal. Amplitude modulated bandwidth (BAM) and frequency

modulated bandwidth (BFM) can be represented as:

1  dA ( t )  (11)
2

E ∫  dt 
BAM =   dt

1  dϕ ( t )
2

− ω  A2 ( t ) dt
E ∫  dt
and BFM =  (12)


2.6 Spectral Momentum of power spectral density (SMPSD):

Power spectral density (PSD) is a useful parameter to find the power and dominant frequency

of any signal. PSD of any signal can be represented as:

 1 T 
2

sz ( f ) = lim  ∫ z (t ) e
− j 2π ft
dt  (13)
T →∞ 2T
 −T 

To find the greater order shape, spectral moment of PSD can be represented as [33]:
N
SM PSD = ∑ k ⋅ PSDk (14)
k =1

2.7 Mean of first derivative of instantaneous frequency (MFDIF):

To find the difference between two consecutive values of instantaneous frequency, first

derivative is calculated. Mean value of these differences is calculated to use it as feature. This

feature can be expressed as [31]:

IET Review Copy Only

 IET Science, Measurement & Technology Page 10 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

δ f = diff ( f i ) (15)

Mean value of first derivative of instantaneous frequency can be represented as:

N −1 δ f (l )
MFDIF = ∑ (16)
l =1 N − 1

where, N is the total number of samples present in first derivative of the instantaneous

frequency.

2.8 Least square support vector machine (LS-SVM)

These features used as input to least squares support vector machine (LS-SVM) classifier for

classification of ALS and normal EMG signals. The decision function for two-class problem

is defined as:

f ( x ) = sign ω T g ( x) + b  (16)

The optimization problem can be formulated as:

1 γ N
Minimize J (ω , b, e) = ω T ω + ∑ ei2 (17)
2 2 i =1

Subject to: yi ω T g ( xi ) + b  = 1 − ei , i = 1, 2,.... N (18)

Where, xi is the index of N inputs, yi is the index of outputs, which is 1 or -1 for input of

class 1 or 2 respectively. b and γ are the bias and regularization parameters respectively. The

α i is the Lagrangian multiplier. The solution of LS-SVM classifier is obtained as:

N 
f ( x) = sign  ∑ α i K ( x, xi ) + b  (19)
 i =1 

The polynomial kernel function for LS-SVM is expressed as:

K ( x, xi ) = g T ( x) g ( xi ) = (xT xi + 1)l (20)

Where, l is the order of polynomial kernel function. The detailed description of LS-SVM

classifier is available in [34-35].

IET Review Copy Only

Page 11 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

3. RESULTS

The non-stationary and non-linear EMG signals are decomposed into numbers of IMFs by

using EMD algorithm. Hilbert transform is used to extract the analytic signals from IMFs,

which enable to compute the analytic features namely area, normalized instantaneous

frequency (IFn), BAM, BFM, SMPSD, and MFDIF features. Fig. 2 (a) and Fig. 2 (b) depict ALS

EMG signal its IMFs and instantaneous frequency of ALS EMG signal. This figure clearly

shows that frequency decreases in higher IMFs. Kruskal-wallis test has been performed on

each feature separately, which provide the discrimination probability of the ALS and normal

EMG signals. Lower values of probability demonstrate better discrimination ability. The

proposed analytic features are used as input to LS-SVM classifier for classification of ALS

and normal EMG signals.

Table 1 shows probability values (p) of area and normalized instantaneous frequency

features with mean and standard deviation of ALS and normal EMG signals. It is clear from

the table that the small p values of first six IMFs of both features are statistically significant.

Area feature of ALS EMG signals is higher as compared to normal EMG signals due to

variation in the amplitude. Also IFn feature ALS EMG signals are higher as compared to

normal EMG signals due to variation in frequency. Fig. 3 and Fig. 4 show the discrimination

ability of area and normalized instantaneous frequency parameter respectively.

Table 2 shows the discrimination ability of BAM and BFM features. Small probability

values (p) of first four IMFs indicates that these features are significant (p <0.05). The BAM

and BFM feature of normal EMG signals are higher as compared to ALS EMG signals. The

higher BAM in normal EMG signals may be due to change of rate of amplitude and higher BFM

in normal EMG signals may be due to change of rate of frequency. Kruskal-wallis plots of

BAM and BFM features are shown in Fig. 5 and Fig. 6 respectively.

IET Review Copy Only

 IET Science, Measurement & Technology Page 12 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

Table 3 shows the performance of mean of first derivative of instantaneous frequency

(MFDIF) and Spectral Momentum of power spectral density (SMPSD) features. From the

Table, it is clear that all the IMFs of SMPSD feature are significant due to p < 0.05. Whereas

IMF2 to IMF7 of MFDIF feature are significant. From the Table 3 and Fig. 7- Fig. 8, it can be

notice that the MFDIF and SMPSD features of ALS EMG signals are higher as compared to

normal EMG signals. Larger values of MFDIF parameter might be due to higher rate of

change of instantaneous frequency whereas larger value of SMPSD parameter is due to greater

power intensity of ALS EMG signals compared to normal EMG signals.

The small p value of these features motivates us to use these features for automatic

classification of ALS and normal EMG signals. These six features are simultaneously

employed to LS-SVM classifier to for classification of ALS and normal EMG signals. The

classification performance evaluated by each IMF to validate the proposed features for

classification of ALS and normal EMG signals. The performance evaluation fidelity

parameter sensitivity, specificity, and accuracy of each IMF are shown in Table 4. The

classification accuracy for classification of ALS and normal EMG signals is 95% for second

IMFs with polynomial kernel function.

4. DISCUSSION

For the first time, EMD method with six analytic features and LS-SVM classifier have been

proposed in this paper to classify the ALS and normal EMG signals. EMD decomposes any

non-stationary signal into symmetric and band limited signals known as IMF. IMF

components are converted into analytic signals which facilitate the computation of

instantaneous amplitude and instantaneous frequency (IF). The proposed features are

extracted using instantaneous amplitude and IF which is possible due to EMD. Kruskal-wallis

test is used on all the features separately to find the discrimination ability of each feature.

IET Review Copy Only

Page 13 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

Kruskal-wallis plot in Figs. 3-8 show that the proposed features are significant for

discrimination of ALS and normal EMG signals.

The LS-SVM classifier with polynomial kernel has been used for the classification ALS and

normal EMG signals with ten-fold cross-validation. LS-SVM classifier has nice properties

that it has moderate complexity, algorithm implements in nonlinear decision regions,

adaptive implementation, and converges to minimum mean square error solutions. The

capabilities of LS-SVM classifier are controlled by the choice of kernel function. This

enables the proper solution of complex regression and pattern recognition problems. The

main objective of this classifier to choose a suitable hyper plane that maximizes the

separation margin between it and nearest data points of each class. Table 5 shows the

comparison of proposed method with another existing classification method in the same

database. From the result of Table 5, it is clear that classification performance of proposed

method is better than another shown method [10].

5. CONCLUSION

The analytic features such as area, normalized instantaneous frequency, BAM, BFM, SMPSD and

MFDIF features with LS-SVM classifier have been explored in this paper for classification of

ALS and normal EMG signals. EMD is a promising method to decompose EMG signals into

some set of IMFs. Analytic representation of IMFs has capability to compute these features.

Discrimination performance of all the features has been evaluated using Kruskal-wallis test

and found significant for classification purpose. Initial four IMFs of all features provide

satisfactory and significant discrimination between the ALS and normal EMG signals. The

six features together with LS-SVM classifier have shown the 95% classification accuracy in

second IMFs with polynomial kernel function. Proposed method can be helpful for the

clinicians to take the decision about ALS disease. Future work includes the use of these

IET Review Copy Only

 IET Science, Measurement & Technology Page 14 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

features to classify other neural diseases (Essential tremor, Tourette's syndrome, Multiple

sclerosis etc.).

Signal 5000 2000

IMF1
0 0
−5000 −2000
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
1000 1000
2

IMF3
IMF

0 0
−1000 −1000
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
500 2000
IMF4

IMF5
0 0
−500 −2000
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
1000 2000

IMF7
6
IMF

0 0
−1000 −2000
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000

Residue
1000 50
8
IMF

0 0
−1000 −50
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
Sample Number Sample Number

(a)

4 4
x 10 x 10
10 10
2
1

5 5
IF
IF

0 0
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
4 4
x 10 x 10
10 10
IF3

5 5
IF

0 0
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
4 4
x 10 x 10
10 10
5

5
IF

5
IF

0 0
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
4 4
x 10 x 10
10 10
8
7

5
IF

5
IF

0 0
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
Sample Number Sample Number

(b)

Fig. 2. (a) Empirical mode decomposition of ALS EMG signal, (b) Instantaneous frequency

plot for eight IMFs of ALS EMG signal

IET Review Copy Only

Page 15 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

8
10 IMF1 IMF2 IMF3 IMF4

7
10

6
log(area)

10

5
10

4
10

ALS Normal ALS Normal (a) ALS Normal ALS Normal

IMF5 IMF6 IMF7 IMF8

7
10

6
10
log(area)

5
10

4
10

ALS Normal ALS Normal (b) ALS Normal ALS Normal

Fig. 3. Comparison of area parameter for ALS and normal EMG signals (a) IMF1-IMF4 and

(b) IMF5- IMF8.

IET Review Copy Only

 IET Science, Measurement & Technology Page 16 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

IMF1 IMF2 IMF3 IMF4

log(IFn)

−1
10

ALS Normal ALS Normal (a) ALS Normal ALS Normal

IMF5 IMF6 IMF7 IMF8

−1
10
log(IF )
n

−2
10

ALS Normal ALS Normal (b) ALS Normal ALS Normal

Fig. 4. Comparison of Normalised instantaneous frequency parameter for ALS and normal

EMG signals (a) IMF1-IMF4 and (b) IMF5- IMF8.

IET Review Copy Only

Page 17 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

TABLE 1. Probability value of the area and Normalised instantaneous frequency features of

the IMFs used in test

Features Area IFn

IMF No. Signal type Mean ± standard deviation Probability Mean ± standard deviation Probability
ALS 1.68×106±5.42×106 0.2427±0.0971
IMF1 0.03 10-5
Normal 6
1.53×10 ±4.65×10 6 0.196±0.0875
ALS 2.99×106±8.98×106 0.2211±0.0809
IMF2 10 -9 10-12
Normal 6
2×10 ±6.23×10 6 0.15±0.0672
ALS 3.89×106±1.73×107 0.1490±0.0599
IMF3 10-7 10-11
5 6 0.1021±0.0391
Normal 8.44×10 ±1.32×10
ALS 2.8×106±3.39×106 0.1179±0.0381
IMF4 10-9 10-15
Normal 6
1.07×10 ±1.44×10 6 0.0805±0.0310
ALS 3.64×106±5.42×106 0.0968±0.0543
IMF5 10-6 10-10
6 6 0.0652±0.0296
Normal 1.6×10 ±2.34×10
ALS 4.28×106±6.68×106 0.0802±0.0444
IMF6 10-8 10-11
6 6 0.0487±0.0257
Normal 1.43×10 ±2.34×10
ALS 2.72×106±3.52×106 0.076±0.0454
IMF7 6 6 10-9 0.0001
Normal 1.19×10 ±1.54×10 0.0527±0.0322
ALS 2.34×106±3.56×106 0.0678±0.0466
IMF8 10-9 0.89
Normal 1.05×106±1.81×106 0.0772±0.0635

IET Review Copy Only

 IET Science, Measurement & Technology Page 18 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

IMF1 IMF2 IMF3 IMF4

0.035

0.03

0.025
log(BAM)

0.02

0.015

0.01

0.005

0
ALS Normal ALS Normal (a) ALS Normal ALS Normal

IMF5 IMF6 IMF7 IMF8

−2
10
log(BAM)

−3
10

ALS Normal ALS Normal (b) ALS Normal ALS Normal

Fig. 5. Comparison of BAM feature for ALS and normal EMG signals (a) IMF1-IMF4 and (b)

IMF5- IMF8.

IET Review Copy Only

Page 19 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

IMF1 IMF2 IMF3 IMF4

3
) 10
FM
log(B

2
10
ALS Normal ALS Normal (a) ALS Normal ALS Normal

IMF5 IMF6 IMF7 IMF8

log(BFM)

2
10

1
10
ALS Normal ALS Normal (b) ALS Normal ALS Normal

Fig. 6. Comparison of BFM feature for ALS and normal EMG signals (a) IMF1-IMF4 and (b)

IMF5- IMF8.

IET Review Copy Only

 IET Science, Measurement & Technology Page 20 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

TABLE 2. Probability value of the BAM and BFM features of the IMFs used in test

Features BAM BFM

IMF No. Signal type Mean ± std. Probability Mean ± std. Probability
3
ALS 0.057±0.0077 1.33×10 ±358.36
IMF1 4.26×10-7 5.65×10-6
Normal 0.0189±0.0077 1.52×108±349.71
ALS 0.012±0.0075 646.87±213.77
IMF2 6.72×10-7 3.962×10-3
Normal 0.0144±0.0073 705.85±175.85
ALS 0.0114±0.0106 421.35±204.83
IMF3 3.7×10-4 1.36×10-3
Normal 0.0112±0.0069 438.57±125.50
ALS 0.0103±0.0095 285.83±151.95
IMF4 2.134×10-3 0.018045
Normal 0.0106±0.0081 305.55±96.56
ALS 0.0102±0.009 226.08±121.92
IMF5 0.013758 0.218188
Normal 0.0091±0.0067 215.66±92.13
ALS 0.0086±0.0069 167.54±95.31
IMF6 0.504716 0.762604
Normal 0.0087±0.0074 158.77±85.01
ALS 0.0101±0.0079 136.49±76.29
IMF7 0.1444821 0.752903
Normal 0.0084±0.0073 118.61±69.42
ALS 0.009±0.0077 103.39±62.45
IMF8 0.9181117 0.1257617
Normal 0.0109±0.0073 120.03±62.21

10
10 IMF1 IMF2 IMF3 IMF4

8
)
PSD

10
log(SM

6
10

ALS Normal ALS Normal (a) ALS Normal ALS Normal

IMF5 IMF6 IMF7 IMF8

log(SMPSD)

10
10

ALS Normal ALS Normal (b) ALS Normal ALS Normal

Fig. 7. Comparison of SMPSD feature for ALS and normal EMG signals (a) IMF1-IMF4 and
(b) IMF5- IMF8

IET Review Copy Only

Page 21 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

IMF1 IMF2 IMF3 IMF4

3
log(MFD ) 10
IF

2
10

ALS Normal ALS Normal (a) ALS Normal ALS Normal

IMF5 IMF6 IMF7 IMF8

3
10
log(MFDIF)

2
10

ALS Normal ALS Normal (b) ALS Normal ALS Normal

Fig. 8. Comparison of MFDIF feature for ALS and normal EMG signals (a) IMF1-IMF4 and
(b) IMF5- IMF8

TABLE 3. Probability value of the SMPSD and MFDIF features of the IMFs used in test

Features SMPSD MFDIF

IMF No. Signal type Mean ± std. Probability Mean ± std. Probability
ALS 6.71×108±1.88×109 812.09±328.52
IMF1 8 8 8.73×10
-7 0.4776
Normal 1.14×10 ±4.38×10 806.59±274.54
ALS 5.99×108±1.24×109 390.65±164.12
IMF2 8 8 1.15×10
-5 0.027433
Normal 2.81×10 ±8.19×10 346.59±135.83
ALS 2.22×108±4.78×108 176.43±105.58
IMF3 8 8
4×10-4 5.20×10-6
Normal 1.43×10 ±2.98×10 130.33±77.89
ALS 4.24×108±8.88×108 86.95±68.08
IMF4 8 8
3.18×10-5 1.00×10-4
Normal 2.08×10 ±4.71×10 59.79±41.44
ALS 3.71×108±6.23×108 48.01±38.27
IMF5 0.0132 8.60×10-3
Normal 2.42×108±4.84×108 33.05±20.89
ALS 8.18×108±2.08×109 30.54±19.54
IMF6 8 9
3.49×10-6 9.60×10-3
Normal 5.6×10 ±2.81×10 23.51±14.61
ALS 2.53×109±8.27×109 206.99±258.38
IMF7 3.89×10-06 6.64×10-05
Normal 5.68×109±5.69×1010 136.4±142.38
ALS 5.06×1011±4.65×1012 96.25±83.36
IMF8 0.002 0.4085285
Normal 9.07×109±6.42×1010 83.5±89.09

IET Review Copy Only

 IET Science, Measurement & Technology Page 22 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

Table 4 Sensitivity, specificity, and accuracy of IMFs with Polynomial kernels of the LS-
SVM classifier for classification between ALS and Normal EMG signals

IMF No. Sensitivity (%) Specificity (%) Accuracy (%)

IMF1 93.43 91.72 94.75
IMF2 93 92.54 95
IMF3 93.71 92.54 94.5
IMF4 93.4 91.93 94.85
IMF5 92.10 89.12 92
IMF6 95.5 85.53 90.5
IMF7 94.36 86.16 89.12
IMF8 91.38 89.12 88

Table 5 A comparison of performance of the classification between ALS and Normal EMG
signals method studied on the same dataset.

Method Sensitivity (%) Specificity (%) Accuracy (%)

Doulah et al. [10] 76 98 92.5
Proposed method 93 92.54 95

References

[1] Yousefi J., and Hamilton-Wright A. Characterizing EMG data using machine-

learning tools. Computers in Biology and Medicine 2014; 51: 1-13.

[2] Tsai AC, Luh JJ, and Lin TT. A novel STFT-ranking feature of multi-channel EMG

for motion pattern recognition. Expert Systems with Applications 2015; 42 (7):3327-

3341.

[3] Rosenfeld J, and Swash M. Lumping or splitting ALS, PLS, PMA, and the other

motor neuron diseases. Neurology 2006; 66 (5): 624-625.

[4] Cristini J. Misdiagnosis and missed diagnoses in patients with ALS. Journal of the

American Academy of Physician Assistants 2006; 19 (7): 29-35.

[5] Ko KD, El-Ghazawi T, Kim D, and Morizono H. Predicting the severity of motor

neuron disease progression using electronic health record data with a cloud computing

Big Data approach. IEEE conference on Computational Intelligence in

Bioinformatics and Computational Biology, pp. 1-6. May 2014.

IET Review Copy Only

Page 23 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

[6] Kasi PK, Krivickas LS, Meister M, Chew E, Schmid M, Kamen G, Clancy EA, and

Bonato P. Motor unit firing characteristics in patients with amyotrophic lateral

sclerosis. Annual Northeast IEEE Conference on Bioengineering pp. 1-2, June 2009.

[7] Fattah SA, Doulah ABMSU, Jumana MA, and Iqbal MdA. Evaluation of different

time and frequency domain features of motor neuron and musculoskeletal diseases.

International Journal of Computer Applications 2012; 43 (23): 34-40.

[8] Kasi PK, Krivickas LS, Melvin M, Chew E, Bonato P, Schmid M, Kamen G, Liu P,

and Clancy EA. Characterization of Motor Unit Behavior in Patients with

Amyotrophic Lateral Sclerosis. 4th International IEEE/EMBS Conference on Neural

Engineering, pp. 10-13, April 2009.

[9] Pal P, Mohanty N, Kushwaha A, and Singh B. Feature extraction for evaluation of

Muscular Atrophy. IEEE International Conference on Computational Intelligence and

Computing Research, pp. 1-4, December 2010.

[10] Doulah ABMSU, and Fattah SA. Neuromuscular disease classification based

on mel frequency cepstrum of motor unit action potential, International Conference on

Electrical Engineering and Information & Communication Technology (ICEEICT),

pp. 1-4, April 2014.

[11] Fattah SA, Doulah ABMSU, Iqbal MA, Shahnaz C, Zhu WP, and Ahmad MO.

Identification of motor neuron disease using wavelet domain features extracted from

EMG signal, IEEE International Symposium on Circuits and Systems (ISCAS), pp.

1308-1311, May 2013.

[12] Fattah SA, Iqbal MdA, Jumana MA, and Doulah ABMS, Identifying the

motor neuron disease in EMG signal using time and frequency domain features with

comparison. An International Journal Signal & Image Processing 2012; 3 (2): 99-114.

IET Review Copy Only

 IET Science, Measurement & Technology Page 24 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

[13] Doulah ABMS, and Jumana MA, ALS disease detection in EMG using time-

frequency method. International Conference on Informatics, Electronics & Vision

(ICIEV), pp. 648-651, May 2012.

[14] Pattichis CS, and Elia AG. Autoregressive cepstral analyses of motor unit

action potentials. Medical Engineering and Physics 1999; 21 (6-7): 405-419.

[15] Subasi A. Classification of EMG signals using combined features and soft

computing techniques. Application of Soft Computing 2012; 12 (8): 2188–2198.

[16] Merlo A, Farina D. A Fast and Reliable Technique for Muscle Activity

Detection from Surface EMG Signals. IEEE transactions on biomedical engineering,

2003; 50(3):316-323.

[17] Katsis CD, Exarchos TP, Papaloukas C, Goletsis Y, Fotiadis DI, and Sarmas I.

A two-stage method for MUAP classification based on EMG decomposition.

Computers in Biology and Medicine 2007; 37 (9): 1232-1240.

[18] Wang G, Wang Z, Chen W, Zhuang J. Classification of surface EMG signals

using optimal wavelet packet method based on Davies-Bouldin criterion, Medical &

Biological Engineering Computing 2006; 44(10):865-72.

[19] Maitrot A, Lucas MF, Doncarli C, Farina D. Signal-dependent wavelets for

electromyogram classification, Medical & Biological Engineering & Computing,

2005; 43(4):487.

[20] Istenic R, Kaplanis PA, Pattichis CS, Zazula D. Multiscale entropy-based

approach to automated surface EMG classification of neuromuscular disorders,

Medical & Biological Engineering & Computing 2010; 48(8):773-81.

[21] Gokgoz E, Subasi A. Effect of multiscale PCA de-noising on EMG signal

classification for diagnosis of neuromuscular disorders, Journal of Medical Systems,

2014;38(4):1-10.

IET Review Copy Only

Page 25 of 26 IET Science, Measurement & Technology
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

[22] Wang G, Ren O. Classification of surface electromyographic signals by means

of multifractal singularity spectrum, Medical & Biological Engineering and

Computing 2013; 51:277–284.

[23] Fusfeld R. Classification of the electromyogram by a pattern-recognition

method Medical and Biological Engineering and Computing 1982; 20, 496-500.

[24] M. Nikolic, Detailed Analysis of Clinical Electromyography Signals EMG

Decomposition, Findings and Firing Pattern Analysis in Controls and Patients with

Myopathy and Amytrophic Lateral Sclerosis. PhD Thesis, Faculty of Health Science,

University of Copenhagen, August 2001.

[25] Huang NE, Shen Z, Long SR, Wu MC, Shih EH, Zheng Q, Yen NC, Tung

CC, and Liu HH. The empirical mode decomposition method and the Hilbert

spectrum for non-stationary time series analysis. Proceedings of the Royal Society of

London A 1998; 454: 903-995.

[26] Flandrin P, Rilling G, and Goncalv´es P. Empirical mode decomposition as a

filter bank. IEEE Signal Processing Letter 2004; 11 (2): 112–114.

[27] Bajaj V, and Pachori RB, Classification of seizure and nonseizure EEG

Signals using empirical mode decomposition, IEEE Transaction on Information

Technology in Biomedicine 2012;16 (6): 1135-1142.

[28] Pachori RB, Bajaj V. Analysis of normal and epileptic seizure EEG signals

using empirical mode decomposition. Computer Methods and Programs in

Biomedicine 2011; 104 (3): 373-381.

[29] Cohen ME, Hudson DL, and Deedwania P. Applying continuous chaotic

modeling to cardic signal analysis. IEEE Engineering in Medicine and Biology

Magazine 1996; 15: 97–102.

IET Review Copy Only

 IET Science, Measurement & Technology Page 26 of 26
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

[30] Pachori RB, Hewson D, Snoussi H., and Duchêne J. Postural time-series

analysis using empirical mode decomposition and second-order difference plots, IEEE

International Conference on Acoustics, Speech, and Signal Processing, pp. 537–540,

April 2009.

[31] Ohsawa Y, Tsumoto S. Chance discoveries in real world decision making:

data- based interaction of human intelligence and artificial intelligence. Springer

Science & Business Media 2006:109-110.

[32] Cohen L. and Lee C., Instantaneous bandwidth for signals and spectrogram,

1990, 2450–2454.

[33] Zhou SM, Gan JQ, and Sepulveda F. Classifying mental tasks based on

features of higher-order statistics from EEG signals in brain–computer interface,

Information Sciences 2008;178(6):1629–1640.

[34] Vapnik V. The nature of statistical learning theory. NewYork: Springer-

Verlag, 1995.

[35] Sukens JAK, J. Vandewalle J. Least squares support vector machine

classifiers. Neural Process. Lett. 1999; 9(3): 293–300.

IET Review Copy Only