Electromyographic Signal Dynamic Behavior in Neuropathies

Spectral Parameters Evaluation and Classiﬁcation

Maria Marta Santos

1,2

, Ana Lu

ısa Gomes

1,2

, Hugo Gamboa

1,2

, Mamede de Carvalho

, Susana Pinto

and Carla Quint

1,4

Departamento de Fsica, Faculdade de Ci

encias e Tecnologia, Universidade Nova de Lisboa, Lisboa, Portugal

PLUX - Wireless Biosignals, Lisboa, Portugal

Faculdade de Medicina, Instituto de Medicina Molecular, Universidade de Lisboa, Lisboa, Portugal

Instituto de Biof

ısica e Engenharia Biom

edica, Faculdade de Ci

encias, Universidade de Lisboa, Lisboa, Portugal

Keywords:

Amyotrophic Lateral Sclerosis (ALS), Coherence, Phase Locking Factor (PLF), Fractal Dimension (FD),

Lempel-Ziv (LZ), Detrended Fluctuation Analysis (DFA), Multiscale Entropy (MSE) , Surface Electromyog-

raphy (sEMG), Ipsilateral, Classiﬁcation.

Abstract:

Amyotrophic Lateral Sclerosis (ALS) is a neurodegenerative disease characterized by motor neurons degen-

eration, which reduces muscular force, being very difﬁcult to diagnose. Mathematical methods, such as

Coherence, Phase Locking Factor (PLF), Fractal Dimension (FD), Lempel-Ziv (LZ) techniques, Detrended

Fluctuation Analysis (DFA) and Multiscale Entropy (MSE) are used to analyze the surface electromiographic

signal’s chaotic behavior and evaluate different muscle groups’ synchronization. Surface electromiographic

signal acquisitions were performed in upper limb muscles, being the analysis executed for instants of contrac-

tion recorded from patients and control groups. Results from LZ, DFA and MSE analysis present capability

to distinguish between the patient and the control groups, whereas coherence, PLF and FD algorithms present

results very similar for both groups. LZ, DFA and MSE algorithms appear then to be a good measure of

corticospinal pathways integrity. A classiﬁcation algorithm was applied to the results in combination with

extracted features from the surface electromiographic signal, with an accuracy percentage higher than 70%

for 118 combinations for at least one classiﬁer. The classiﬁcation results demonstrate capability to distin-

guish both groups. These results can demonstrate a major importance in the disease diagnose, once surface

electromyography (sEMG) may be used as an auxiliary diagnose method.

1 INTRODUCTION

ALS is a fatal and very progressive disease, character-

ized by both upper and lower motor neurons degen-

eration, involving brainstem and also multiple spinal

cord innervation regions. This disorder is responsi-

ble for abnormal motor activity. ALS patients typ-

ically present fatigue, quickly progressive weakness

and reduced exercise capacity with loss of voluntary

movement, spasticity, fasciculations, dysphagia (dif-

ﬁculty in swallowing), dyspnea (difﬁculties in breath-

ing) and dysarthria (difﬁculties in speaking). After the

ﬁrst symptoms, death may occur within 3−5 years for

most of the patients. ALS is very difﬁcult to diagnose,

since there isn’t available a reliable biomarker of dis-

ease activity and progression (Kiernan et al., 2011;

Mitchell and Borasio, 2007).

Upper motor neuron integrity can be evaluated

through the investigation of oscillatory activity prop-

agation. The motor cortex activity can be recorded,

and both alpha (8 − 12Hz) and beta (15 − 30Hz) fre-

quency bands can be analyzed via coherence and PLF

(Farmer et al., 2007).

This work explores the analysis of ipsilateral ac-

quisitions, which was presented with promissory pre-

liminary results in (Camara, 2013), using different ap-

proaches.

Motor unit recruitment patterns complexity can be

quantiﬁed using FD. However, the strength of a mus-

cle’s contraction is better estimated based on Max-

imum Fractal Length (MFL), even for very small

muscle contraction strength, rather than FD (Poos-

apadi Arjunan and Kumar, 2012).

The LZ measure is a well suited feature regard-

227

Marta Santos M., Luisa Gomes A., Gamboa H., Carvalho M., Pinto S. and Quintão C..

Electromyographic Signal Dynamic Behavior in Neuropathies - Spectral Parameters Evaluation and Classiﬁcation.

DOI: 10.5220/0005215602270234

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2015), pages 227-234

ISBN: 978-989-758-069-7

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

ing sEMG analysis, particularly during dynamic con-

tractions (highly non-stationary signal), since this

feature doesn’t make any assumption of stationarity

(Talebinejad et al., 2011).

The DFA method is proved to be efﬁcient to de-

scribe upper-limbs movements. DFA outperforms

other methods such as correlation dimension and

Higuchi methods, being a stable technique to quan-

tify fractality and to establish self-similarity, being a

robust method in the presence of nonstanionarity time

series and trends (Phinyomark et al., 2011).

Entropy is a feature which can detect and quantify

differences in the EMG signal amplitude distributions

due to neuromuscular conditions (pathology) (Kapla-

nis et al., 2010). This feature has been successfully

applied to physiological signals, in order to quantify

their degree of complexity. Multiscale entropy (MSE)

has been proved to be more effective than single-scale

entropy in this quantiﬁcation, since it considers mul-

tiple spatiotemporal scales (Zhang et al., 2013).

k-Nearest Neighbor is a relatively simple and fast

algorithm, important characteristics in the classiﬁca-

tion process (Kim et al., 2011). Decision tree meth-

ods have also been used for dealing with classiﬁcation

problems in various domains, such as pattern recog-

nition, data mining, web mining and signal process-

ing, among others. However, standard decision tree

algorithms can only handle discrete attributes (Wang

et al., 2006). The decision tree algorithm has usu-

ally good performance for large data sets in a short

time (Zhang et al., 2011). Random Forest algorithm

has been tested with real and simulated data sets. The

results have been proven to be very accurate. This al-

gorithm is fast, versatile and can be applied to very

large data sets. It has also been shown its robustness

against noise in the outcome compared with several

other methods (Roy and Larocque, 2012). Discrim-

ination between subject or patient group concerning

age, gender and injuries in athletes has been proven

to be effective using a classiﬁcation approach with

generic features and AdaBoost (Eskoﬁer et al., 2012).

ıve Bayes method has shown to be competitive

among much sophisticated induction algorithms con-

cerning experiments on real world data, despite the

assumption of conditional independence (Wang et al.,

2006).

2 METHODS

2.1 Coherence

Coherence function estimates values from 0 to 1, as-

suming the value 0 if there is no association between

two signals at a certain frequency, and the value 1 if

there is a perfectly linear association between them

(Farmer et al., 2007)

2.2 Phase Locking Factor (PLF)

PLF is a measure of synchronization between two

signals, in which the frequencies of interest are iso-

lated by the application of a narrow band-passed ﬁlter.

Then, the relationship between the phases of the two

signals, φ

(t) and φ

(t) are analyzed (Almeida et al.,

2011):

≡ |

∑

t=1

i[φ

(t)−φ

(t)]

| = |he

i[φ

(t)−φ

(t)]

i| (1)

where ρ

is the PLF and T is the number of discrete

samples.

PLF ranges from 0 to 1. While ρ

= 0 corre-

sponds to asynchronous signals (their phases are not

correlated), ρ

= 1 is attained if the two signals are

in perfect synchronization (Almeida et al., 2011).

2.3 Fractal Dimension (FD)

FD is one of the most used measurements for the eval-

uation of the dynamics of complex systems. A non-

integer FD usually indicates a chaotic behavior, and

the smallest integer bigger than FD is considered the

minimum number of independent variables capable to

describe this behavior (West, 1994).

For physiological signals, FD can be estimated us-

ing Higuchi algorithm, since it is suitable for non-

periodic and irregular time series.

This algorithm results on a plot of several curve

lengths, L(k), for a pre-determined k

max

, being k ∈

[1,k

max

] (Phinyomark et al., 2011).

2.4 Lempel-Ziv (LZ)

LZ is another tool used to analyze the deterministic

complexity of a highly non-linear chaotic setting. To

compute LZ, it is necessary to convert the sEMG sig-

nal into a symbolic sequence, conventionally a binary

sequence, by comparison to a threshold. Next, the

number of distinct patterns within a signal is obtained,

which is directly related to the complexity of the sys-

tem (Talebinejad et al., 2011).

2.5 Detrended Fluctuation Analysis

(DFA)

The DFA method can be applied to the study of elec-

trophysiological signals, being a modiﬁed root mean

BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

228

square (RMS) analysis of a random walk (Phiny-

omark et al., 2011).

The outputs of the DFA algorithm are scaling ex-

ponents extracted from a log-log graph (for details

(Phinyomark et al., 2011)). These exponents assume

values between 0 and 2, according to the time series

behavior:

• If 0 < α <

the time series is anti-correlated

• If α

∼

the time series is uncorrelated, or indi-

cates White noise (the value at one instant cannot

be correlated with any previous value)

• If

< α < 1 the time series is correlated

• If α

∼

1 indicates Pink noise (

noise)

• If 1 < α <

indicates nonstationary or random

walk

• If α

∼

indicates Brownian noise (i. e. the inte-

gration of the White noise)

2.6 Multiscale Entropy (MSE)

Entropy has been often used to quantify complexity,

since traditional entropy deﬁnitions (e. g. Shannon-

entropy) are used to measure disorder and uncertainty,

as well as to characterize a systems’ gain of informa-

tion. Approximate Entropy and its modiﬁcation Sam-

ple Entropy are entropy-based complexity measures

with a single scale which are widely used in short

and noisy time series. Thus, in MSE one of this ap-

proaches is applied to different time scales, resulting

on a plot of the entropy value as a function of the fac-

tor scale, τ (Zhang et al., 2013).

2.7 Classiﬁcation

Classiﬁcation methods are used to identify the be-

longing of a novel observation in a set of categories

(sub-populations). These categories are obtained with

based on a training group of observations, therefore,

it is necessary to have previous knowledge of the cat-

egory membership of each observation of the train-

ing group. These observations are then analyzed

according to the extrated features. In this project,

the used features are: kurtosis, maximum frequency,

mean, median frequency, power band, spectral kur-

tosis, spectral skewness, spectral spread and correla-

tion. There will be also analyzed the results of the pre-

viously referred implemented algorithms (coherence,

PLF, FD, LZ, DFA and MSE). Classiﬁcation can be

implemented throughout a various number of algo-

rithms, the classiﬁers (k - Nearest Neighbor, Decision

Tree, Random Forest, AdaBoost and Na

ıve Bayes).

Leave-one-out cross validation iterator is used to split

data in train/test sets. Hence, all samples except one

are used as a train set, being this left out sample tested

after (Pedregosa et al., 2011).

3 ACQUISITIONS

3.1 Subjects

Measurements were performed in two different

groups of subjects: group of patients, with 21 mem-

bers presenting ALS, and group of control, with 26

members which do not evidence ALS disease. There-

fore, the patients group contains 21 subjects, varying

this number according to the analyzed channel (left or

right hand or forearm), since some patients presented

inability of self-controlled movement for one arm. All

participants from the patients group have been diag-

nosed within less than three years, except two mem-

bers which have been diagnosed previously. All par-

ticipants from both groups have ages between 23 and

77 years (mean of 59 years for the patient group and

45 years for the control group).

3.2 Acquisition Protocol

The acquisition protocol is identical to the one in (Ca-

mara, 2013). The performed task was repeated for 6

minutes or less according to maximum time borne by

the patients. Subjects sat down and placed both hands

and forearms on a desk in a parallel position, 10 cm

away from each other with hand palms facing one an-

other in 90 degrees ﬂexion with the elbow. While

listening to a programmed sound, which guided the

movement, subjects were asked to coordinately ele-

vate both index ﬁngers vertically with maximum ar-

ticular amplitude in an opposite direction from the

other ﬁngers position, hold that position for 3 sec-

onds while maintaining a certain force and return to

the original position, remaining in that position for

3 seconds while trying to relax the arms muscles as

much as possible.

3.3 Recordings

Contralateral and ipsilateral acquisitions were

recorded simultaneously, and 4 signals were acquired

from each subject using EMG sensors attached to

a bioPlux device. The sensors were placed on the

ﬁrst dorsal interosseus muscle for both left and right

hand, and on the extensor digitorum communis

muscle for both left and right forearm. Ground was

placed on ulna bone inferior extremity, where no

ElectromyographicSignalDynamicBehaviorinNeuropathies-SpectralParametersEvaluationandClassification

229

muscle activity is present. Figure 1 shows the surface

electrodes placements.

The bioPlux device has eight analog channels with

12-bit resolution and also an external channel to be

used as a reference ground. The EMG sensors have

second order band pass ﬁlter with cutoff frequencies

of 25 and 450. EMG signals were recorded using a

sampling frequency of 1000 Hz, being the recorded

data transmitted via Bluetooth to a computer.

Figure 1: Simultaneous contralateral and ipsilateral exper-

imental setup: Bioplux research device, placement of four

EMG sensors and ground.

4 SIGNAL PROCESSING

All acquired signals were processed using python lan-

guage. First, the signals’ Direct Current (DC) compo-

nent was removed and a third order butter band pass

ﬁlter of 10 − 500 Hz was applied. From each pair

of ipsilateral signals, intervals of common contraction

were isolated from intervals of relaxation using the al-

gorithm referred in (Camara, 2013).

All the previously referenced algorithms were ap-

plied to moments of contraction. While coherence

and PLF were calculated twice for each subject, be-

ing calculated for a pair of signals, FD, LZ, DFA

and MSE were calculated four times, being applied

to each one of the four acquired signals individually.

Coherence and PLF were calculated for two com-

mon sections of data from an interval of one contrac-

tion. This was performed for all contractions, with

posterior averaging of all epochs. FD, LZ, DFA and

MSE calculus was applied to a concatenation of all

contractions.

4.1 Coherence Processing

All EMG signals were full-wave rectiﬁed and then,

for each contraction, coherence was calculated using

python libraries (matplotlib.mlab.cohere tool). Coher-

ence is then averaged for the moments of contrac-

tion. The used sampling frequency is 1000 Hz, the

Nonequispaced fast Fourier transform (NFFT) is 512

and the value that dictates the dependency between

FFT windows is NFFT/2. Coherence was averaged

for each frequency among all the subjects within the

same group.

4.2 PLF Processing

PLF algorithm was developed in (Camara, 2013).

All signals were full-wave rectiﬁed and each signal

was band pass ﬁltered in order to remove all the

other undesirable frequencies. The used ﬁlter was

[ f − 2, f + 2], being f the analyzed frequency. There-

fore, PLF calculus was performed as many times as

the number of frequencies to analyze. PLF was ob-

tained with resolution of 1Hz for all the frequencies

within the beta band (15−30Hz). PLF was computed

according to equation 1. Hence, a different PLF value

is obtained for each contraction of the analyzed sig-

nals and, in order to attain a ﬁnal value for each sub-

ject, PLF was averaged among all contractions within

the same acquisition. Finally, for each analyzed fre-

quency, PLF was averaged among all members within

each group, for the patient and the control groups.

4.3 FD Processing

In the Higuchi fractal dimension algorithm k

max

was

deﬁned as 128, as suggested in literature (Phinyomark

et al., 2011). The FD coefﬁcient was estimated and

averaged among all subjects within each group.

4.4 LZ Processing

The LZ coefﬁcient was calculated for a binary se-

quence obtained from the rectiﬁed ﬁltered used signal

with threshold deﬁned as 0.4.

Since the LZ coefﬁcient is greatly related to the

number of different patterns within each signal, and

the number of patterns is related to the length of the

signal, all the signals were cut in accordance with

the minimum reasonable signal length within both

groups. An average of the LZ coefﬁcient was cal-

culated for all members within the patients and the

control’s groups.

4.5 DFA Processing

The DFA was applied among all subjects, obtaining

two distinct scaling exponents, α

and α

. These scal-

ing exponents were averaged for all members within

each group.

BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

230

4.6 MSE Processing

For MSE, the Sample Entropy was computed for 20

scales. These values were averaged among all sub-

jects within each group, and plotted in a graph for

both populations.

4.7 Classiﬁcation

The feature extraction and the classiﬁcation algo-

rithms were developed in (Gomes, 2014).

The algorithms were adapted for the use of two

channels, and the features extracted were selected ac-

cording to this work objectives.

Features were extracted from a pair of signals

(right hand and forearm). The sampling frequency

was placed at 1000 Hz, being used a central window

with 2000 points. The results obtained from the pre-

viously described algorithms were joined to these fea-

tures.

Posteriorly several combinations of the extracted

features, the developed algorithm results and of both

of them were organized and classiﬁed.

5 RESULTS AND DISCUSSION

In this section, we show results obtained from right

arm signals, being the most relevant conclusions also

achieved for left arm analysis.

5.1 Coherence Results

Coherence mean values for the group of patients and

the group of control are presented in Figure 2.

Figure 2: Mean coherence dependency on frequency for pa-

tients by the straight line and for controls by the dashed line

with NFFT placed as 512. The ﬁrst grey box delimitates

the frequencies corresponding to the alpha band (8 − 12

Hz). The second grey box delimitates the frequencies cor-

responding to the beta band (15 − 30 Hz). These results are

from the right arm.

By the observation of the graphics presented in

Figure 2, coherence mean values are very similar

for both the patients and the control groups. Con-

trol group mean value for coherence is slightly higher

within the alpha band. Coherence pooled value for

patients is 0,23 ± 0,21 for the right arm. Coherence

pooled value for the control group is 0, 22 ± 0,21 for

the right arm. These values are different from those

found in literature (Fisher et al., 2012), since slightly

higher values of coherence were expected for the con-

trol group within the beta band than for the patients

group. These differences may be explained by the

sampling frequency used, differences in the acquisi-

tion protocol, differences in the used algorithm or pa-

rameters and the tested subjects themselves (age, gen-

der, lifestyle, etc.). The obtained results are also dif-

ferent from the expected for ipsilateral acquisitions in

(Camara, 2013), since only preliminary results were

obtained previously.

5.2 PLF Results

PLF mean values depending on frequency for both

groups are presented in Figure 3.

Figure 3: Mean PLF values dependency on frequency for

patients by the straight line and for controls by the dashed

line; The ﬁrst grey box delimitates the frequencies corre-

sponding to the alpha band (8 − 12 Hz). The second grey

box delimitates the frequencies corresponding to the beta

band (15 − 30 Hz). These results are from the right arm.

Observing the graphics presented in Figure 3, PLF

values appear to be very similar for both groups. Al-

pha band appears to demonstrate a higher difference

between both groups PLF values. However, since

this work proceeds the investigation of PLF within

the beta band (Camara, 2013), PLF was only com-

puted for 10 Hz within the alpha band. PLF pooled

value for the patients group is 0, 237 ± 0,006 for the

right arm. PLF pooled value for the control group is

0,242 ±0,009 for the right arm. These results are dif-

ferent from the expected for ipsilateral acquisitions in

(Camara, 2013), since only preliminary results were

obtained previously. Differences in results may also

be explained by the length of the used signal and the

sampling frequency used, and characteristics such as

age, gender and lifestyle may inﬂuence this analysis.

ElectromyographicSignalDynamicBehaviorinNeuropathies-SpectralParametersEvaluationandClassification

231

5.3 FD Results

FD coefﬁcient mean and standard deviation values are

presented in Table 1, for the group of patients and the

group of control. Table 2 represents the MFL (the

point of lowest scale) values for each group.

Table 1: Mean and standard deviation values of FD coefﬁ-

cient for patients and control group for the right arm.

Patients group Control group

Right hand −1.985 ± 0.008 −1.983 ± 0.008

Right forearm −1.983 ± 0.013 −1.984 ± 0.006

Table 2: MFL values for patients and control group.

Patients group Control group

Right hand 3.72 ± 0.39 3.99 ± 0.20

Right forearm 3.72 ± 0.47 3.80 ± 0.22

From the observation of Table 1 FD coefﬁcient is

almost identical for both groups. Therefore, FD co-

efﬁcient is not a good measure of distinction between

patients and control group. However, the obtained FD

values are very similar to the results obtained from

different upper limb activities in (Phinyomark et al.,

2011). Observing Table 2, MFL is always slightly

higher for the control group.

5.4 LZ Results

Table 3 shows the LZ coefﬁcient mean value and stan-

dard deviation for a binary sequence.

Table 3: LZ coefﬁcient for a binary sequence obtained from

the rectiﬁed ﬁltered used signal with threshold deﬁned as

0.4.

Patients group Control group

Right hand 0.17 ± 0.13 0.27 ± 0.16

Right forearm 0.27 ± 0.29 0.15 ± 0.13

From the observation of Table 3, LZ coefﬁcient

presents higher values for the control group for the

right hand, and higher values for the patients group for

the right forearm. Tables 3 presents a good distinction

between patient and control groups. Therefore, for a

binary sequence obtained from the ﬁltered used signal

and for a binary sequence obtained from the rectiﬁed

ﬁltered used signal, LZ algorithm appears to distin-

guish subjects between both groups regarding pooled

LZ coefﬁcient values, however, with some variance

among subjects within each group.

Table 4: DFA α

coefﬁcient mean and standard deviation

values for both groups for the right arm.

Patients group Control group

Right hand 1.23 ± 0.17 1.35 ± 0.14

Right forearm 1.39 ± 0.09 1.44 ± 0.11

Table 5: DFA α

coefﬁcient mean and standard deviation

values for both groups for the right arm.

Patients group Control group

Right hand 0.50 ± 0.12 0.52 ± 0.09

Right forearm 0.62 ± 0.16 0.59 ± 0.10

5.5 DFA Results

Tables 4 and 5 present the DFA coefﬁcients, α

and

mean values for both groups, respectively.

As observed in Tables 4 and 5, DFA coefﬁcients

generally present higher values for the control group.

Therefore, DFA algorithm appears to distinguish both

groups.

5.6 MSE Results

Figure 4 represents the Sample Entropy mean value

for each scale for both patient and control groups.

Observing Figure 4, patients group exhibit higher

entropies for the right hand, whereas control group

exhibit higher entropies for the right forearm. There-

fore, despite statistically the differences between pa-

tients and control’s groups are very small, the MSE

tendency appears to be distinct between both groups.

MSE pooled values for the patients and control

groups are presented in Table 6.

Observing the results presented in Figure 4 and

Table 6, MSE algorithm appears to be capable of dis-

tinguish both groups.

(a)

(b)

Figure 4: Sample Entropy mean value for each scale. The

straight line represents the patient group, and the dashed

line represents the control group. The errorbars represent

the standard deviation. (a) Results for the right hand. (b)

Results for the right forearm.

BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

232

Table 6: MSE mean and standard deviation values, for both

patients and control groups.

Patients group Control group

Right hand 3.15 ± 0.46 2.77 ± 0.48

Right forearm 3.08 ± 0.49 3.50 ± 0.52

5.7 Classiﬁcation Results

Table 7 presents the 14 combinations and the 3 classi-

ﬁers with better results for the right arm for a Leave-

one-out cross validation.

Table 7: Classiﬁcation results for the right arm for Deci-

sion Tree, Random Forest and AdaBoost Classiﬁers. 1) sum

power band + MSE + DFA (α

+ α

); 2) mean + MSE +

DFA (α

+ α

); 3) mean + MSE + DFA (α

+ α

) + LZ; 4)

spectral skewness + MSE + DFA α

+ α

) + LZ; 5) spectral

spread + MSE + DFA (α

+ α

) + LZ; 6) spectral kurtosis

+ MSE + DFA (α

+ α

) + LZ; 7) sum power band + MSE

+ coherence + PLF + DFA (α

+ α

); 8) kurtosis + MSE

+ DFA (α

+ α

); 9) kurtosis + MSE + DFA (α

+ α

) +

LZ; 10) kurtosis + MSE + coherence + PLF + DFA (α

); 11) maximum frequency + MSE + coherence + PLF

+ DFA (α

+ α

); 12) mean +MSE + DFA (α

+ α

); 13)

spectral skewness + MSE + coherence + DFA (α

+ α

);

14) spectral spread +MSE + coherence + PLF + DFA (α

Combinations Decision Tree Random Forest AdaBoost

Right Hand

1) 0,806 0,694 0,722

2) 0,806 0,778 0,722

3) 0,833 0,611 0,750

4) 0,806 0,667 0,667

5) 0,806 0,639 0,722

Right Forearm

6) 0,667 0,750 0,806

Right Arm

7) 0,778 0,750 0,806

8) 0,556 0,750 0,806

9) 0,500 0,806 0,778

10) 0,694 0,833 0,694

11) 0,667 0,611 0,806

12) 0,556 0,806 0,639

13) 0,667 0,861 0,750

14) 0,694 0,833 0,694

Classiﬁcation results demonstrate the algorithms

distinction capability, since the presented results show

the percentage of cases that the algorithm classiﬁed

correctly (accuracy percentage).

Classiﬁcation results were obtained for the right

arm for a leave-one-out cross validation tested with

k - Nearest Neighbor, Decision Tree, Random For-

est, AdaBoost and Na

ıve Bayes classiﬁers. In spite

of 201 combinations have been arranged, 118 combi-

nations present values higher than 70.0% for at least

one classiﬁer, 41 combinations present results higher

than 77,8% for at least one classiﬁer, and 14 combina-

tions present results higher than 80.6% for one classi-

ﬁer (Decision Tree, Random Forest or AdaBoost).

The best results include MSE, DFA, LZ, coher-

ence and PLF algorithm results, and also the extracted

features mean, maximum frequency, spectral kurto-

sis, spectral skewness, spectral spread, sum power

band and kurtosis, in several different combinations.

The top 14 best combinations include the results of

both DFA and MSE algorithms. The best obtained

combination is spectral skewness + MSE + coher-

ence + DFA (α

and α

), with an accuracy percent-

age of 86,1% for a Random Forest Classiﬁer. There-

fore, it is proved that in spite of some algorithms

may present slender differences between both control

and patients groups, the combination of these algo-

rithms with other measures can improve the distinc-

tion capability between members of patients and con-

trol groups.

6 CONCLUSIONS

In this work, methodology was developed to evalu-

ate the complexity of sEMG signal acquired from dif-

ferent muscle groups of healthy subjects and patients

with ALS. FD, LZ, DFA and MSE algorithms were

implemented and all of these algorithms as well as

coherence and PLF algorithms were applied to the ac-

quired ﬁltered signals.

Contrary to the results presented in literature, co-

herence analysis does not present signiﬁcant differ-

ences between the group of patients and the group

of control. PLF analysis also does not present any

signiﬁcant differences between both groups. Results

from both algorithms appear to be slightly higher for

the control group for the alpha band frequencies. For

further work is suggested to compute PLF for all the

frequencies within the alpha band with resolution of

1 Hz.

FD analysis results in FD coefﬁcients very sim-

ilar for all the signals for both patients and control

groups. Therefore, this algorithm does not seem good

to obtain a distinction between both groups. How-

ever, MFL presents slightly higher values regarding

the control group, being this a better measure of dis-

tinction between both groups than the FD coefﬁcient.

LZ analysis presents better results, being LZ coef-

ﬁcient higher for the control group for the right hand,

and higher for the patients group for the right fore-

arm. These results are for a binary sequence obtained

from the rectiﬁed ﬁltered used signal. Therefore, LZ

coefﬁcient appears to be a good reﬂection of neural

degeneration. Since the used threshold to obtain the

binary sequence was deﬁned as 0.4, it is not adapted

ElectromyographicSignalDynamicBehaviorinNeuropathies-SpectralParametersEvaluationandClassification

233

to each individual signal. Therefore, it is suggested

for further work a threshold obtained as a percentage

of the standard deviation of each signal.

DFA analysis presents higher values for both α

and α

DFA coefﬁcients for the control group. Then,

this algorithm seems to be a good measure to reﬂect

neural degeneration.

MSE analysis presents higher values for the right

hand for the patients group, and higher values for the

right forearm for the control group. This algorithm

appears to be a good indicator of neural degeneration.

LZ, DFA and MSE analysis have then potential as

a quantitative test for upper and lower neural integrity

concerning ALS disease.

Classiﬁcation results demonstrate to provide a

good distinction of both groups, being the combina-

tion of various algorithms with features proved to be

advantageous to improve both groups distinction ca-

pability.

REFERENCES

Almeida, M., Vig

ario, R., and Bioucas-Dias, J. (2011).

Phase locked matrix factorization. In Proc. of the EU-

SIPCO Conference, pages 1728–1732.

Camara, M. (2013). Coherence and phase locking dis-

ruption in electromyograms of patients with amy-

otrophic lateral sclerosis. Master’s thesis, Faculdade

de Ci

encias e Tecnologia da Universidade Nova de

Lisboa.

Eskoﬁer, B. M., Kraus, M., Worobets, J. T., Stefanyshyn,

D. J., and Nigg, B. M. (2012). Pattern classiﬁca-

tion of kinematic and kinetic running data to distin-

guish gender, shod/barefoot and injury groups with

feature ranking. Computer methods in biomechanics

and biomedical engineering, 15(5):467–474.

Farmer, S. F., Gibbs, J., Halliday, D. M., Harrison, L. M.,

James, L. M., Mayston, M. J., and Stephens, J. A.

(2007). Changes in emg coherence between long and

short thumb abductor muscles during human develop-

ment. The Journal of physiology, 579(2):389–402.

Fisher, K. M., Zaaimi, B., Williams, T. L., Baker, S. N.,

and Baker, M. R. (2012). Beta-band intermuscular

coherence: a novel biomarker of upper motor neu-

ron dysfunction in motor neuron disease. Brain,

135(9):2849–2864.

Gomes, A. L. (2014). Human activity recognition with ac-

celerometry: Novel time and frequency features. Mas-

ter’s thesis, Faculdade de Ci

encias e Tecnologia da

Universidade Nova de Lisboa.

Kaplanis, P. A., Pattichis, C. S., Zazula, D., et al.

(2010). Multiscale entropy-based approach to au-

tomated surface emg classiﬁcation of neuromuscular

disorders. Medical & biological engineering & com-

puting, 48(8):773–781.

Kiernan, M. C., Vucic, S., Cheah, B. C., Turner, M. R.,

Eisen, A., Hardiman, O., Burrell, J. R., and Zoing,

M. C. (2011). Amyotrophic lateral sclerosis. The

Lancet, 377(9769):942–955.

Kim, K. S., Choi, H. H., Moon, C. S., and Mun, C. W.

(2011). Comparison of¡ i¿ k¡/i¿-nearest neighbor,

quadratic discriminant and linear discriminant analy-

sis in classiﬁcation of electromyogram signals based

on the wrist-motion directions. Current Applied

Physics, 11(3):740–745.

Mitchell, J. D. and Borasio, G. D. (2007). Amyotrophic

lateral sclerosis. The lancet, 369(9578):2031–2041.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer,

P., Weiss, R., Dubourg, V., Vanderplas, J., Passos,

A., Cournapeau, D., Brucher, M., Perrot, M., and

Duchesnay, E. (2011). Scikit-learn: Machine learning

in Python. Journal of Machine Learning Research,

12:2825–2830.

Phinyomark, A., Phukpattaranont, P., Limsakul, C., and

Phothisonothai, M. (2011). Electromyography (emg)

signal classiﬁcation based on detrended ﬂuctuation

analysis. Fluctuation and Noise Letters, 10(03):281–

301.

Poosapadi Arjunan, S. and Kumar, D. K. (2012). Compu-

tation of fractal features based on the fractal analysis

of surface electromyogram to estimate force of con-

traction of different muscles. Computer Methods in

Biomechanics and Biomedical Engineering, (ahead-

of-print):1–7.

Roy, M.-H. and Larocque, D. (2012). Robustness of ran-

dom forests for regression. Journal of Nonparametric

Statistics, 24(4):993–1006.

Talebinejad, M., Chan, A. D., and Miri, A. (2011). A

lempel–ziv complexity measure for muscle fatigue es-

timation. Journal of Electromyography and Kinesiol-

ogy, 21(2):236–241.

Wang, L.-M., Li, X.-L., Cao, C.-H., and Yuan, S.-M.

(2006). Combining decision tree and naive bayes for

classiﬁcation. Knowledge-Based Systems, 19(7):511–

515.

West, B. J. (1994). Fractal physiology, volume 2. Oxford

University Press.

Zhang, X., Chen, X., Barkhaus, P. E., and Zhou, P. (2013).

Multiscale entropy analysis of different spontaneous

motor unit discharge patterns. Journal of Biomedical

and Health Informatics, 17(2).

Zhang, X., Chen, X., Li, Y., Lantz, V., Wang, K., and Yang,

J. (2011). A framework for hand gesture recognition

based on accelerometer and emg sensors. Systems,

Man and Cybernetics, Part A: Systems and Humans,

IEEE Transactions on, 41(6):1064–1076.

BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

234