Surface EMG Signal Segmentation and Classiﬁcation for Parkinson’s

Disease Based on HMM Modelling

Hichem Bengacemi

1,3 a

, Abdenour Hacine Gharbi

, Philippe Ravier

3 b

, Karim Abed-Meraim

and

Olivier Buttelli

Signal Processing Lab, École Militaire Polytechnique, Algiers, Algeria

LMSE laboratory, University Mohamed El Bachir El Ibrahimi of Bordj Bou Arreridj, El-Anasser, Bordj Bou Arréridj,

34030, Algeria

PRISME Lab, Université d’Orléans, INSA-CVL, 12 Rue de Blois, 45067, Orléans, France

Keywords:

Parkinson’s Disease Diagnostic, sEMG Signal Classiﬁcation, sEMG Signal Segmentation,

Wavelet Cepstral Coefﬁcient (WCC), HMM Models.

Abstract:

To increase the diagnostic accuracy, the techniques of artiﬁcial intelligence can be used as a medical support.

The Electromyography (EMG) signals are used in the neuromuscular dysfunction evaluation. This paper pro-

poses a new frame work for segmenting and classifying the surface EMG (sEMG)signals by segmenting the

EMG signal in regions of muscle activity (ACN) and non activity (NAN) for control group (healthy) and the

muscle activity (ACP) and non activity (NAP) for Parkinsonian group. This paper proposes an automatic sys-

tem of the neuromuscular dysfunction identiﬁcation for Parkinson disease diagnosis based on HMM modeling

by using on sEMG signals. Discrete Wavelet Transform (DWT), LP coefﬁcients and FLP coefﬁcients have

been used for feature extraction. The results have been evaluated on ECOTECH project database using the

signal classiﬁcation rate (CRS) and the Accuracy (Acc) criterion. The obtained results show highest perfor-

mance by using HMM models of 2 states associated with GMM of 6 Gaussians, combined with Log Wavelet

decomposition based Energy(LWE) descriptor based on Coiﬂet wavelet mother with decomposition level of

4. The proposed methodology leads to a classiﬁcation accuracy of leads to an Acc of 99.37 % and a CRS of

100 %.

1 INTRODUCTION

Parkinson’s disease (PD) is a neurodegenerative dis-

ease caused by dopaminergic degeneration. This dis-

ease is deﬁned clinically by movement disturbances,

motor disturbances, and loss of postural control. It

is characterized by several symptoms: freezing, pos-

tural instability and gait disturbance, resting tremor,

rigidity, Akinesia and Bradykinesia.

The diagnosis of PD is not always easy to make.

The latter is generally based on the symptoms de-

scribed by the patient and neurological examination

made by the doctor. Several research works are car-

ried out for the analysis, evaluation and identiﬁcation

of PD using several approaches such as: Hand writing

(Rosenblum et al., 2013), analysis and gait evaluation

by recording stride intervals (Wendling, 2008; Bhoi,

https://orcid.org/0000-0002-4141-5275

https://orcid.org/0000-0002-0925-6905

2017; Abdulhay et al., 2018; Kugler et al., 2013) ,

voice analysis (Manwatkar et al., ), medical imaging

(Xu and Zhang, 2019; Porter et al., 2020), the analysis

of electrovestibulography (EVestG) signals which are

in fact the vestibular response modulated by cortical

cerebral signals (Dastgheib et al., 2012) and ﬁnally

the analysis of gait using surface EMG signals which

are the subject of this work (Elamvazuthi et al., 2015;

Raut and Gurjar, 2015; Nazmi et al., 2016).

The problem of classiﬁcation and diagnosis of

neurodegenerative diseases including PD, is strongly

lied to the techniques of features extraction and those

of classiﬁcation (Sugavaneswaran et al., 2012; Car-

letti et al., 2006; Henmi et al., 2009; Hausdorff et al.,

1998; Okamoto et al., 2009; Surangsrirat et al., 2016;

Bhoi, 2017). In (Hausdorff et al., 1997), the au-

thors presented that the ﬂuctuations of stride inter-

val are increased in PD and correlate with the dis-

ease severity degree. In (Miller et al., 1996), the au-

thors showed that the variability of the EMG signal

930

Bengacemi, H., Gharbi, A., Ravier, P., Abed-Meraim, K. and Buttelli, O.

Surface EMG Signal Segmentation and Classiﬁcation for Parkinson’s Disease Based on HMM Modelling.

DOI: 10.5220/0012572900003654

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 930-938

ISBN: 978-989-758-684-2; ISSN: 2184-4313

recorded from the astrocnemius muscle is higher in

Parkinson’s patients. In (Oung et al., 2018), the au-

thors proposed a multi-class classiﬁcation, to indicate

the severity level of PD (mild, moderate, severe) us-

ing the empirical wavelet transformation (EWT) and

empirical packet-to-wavelet transformation (EWPT)

based on the motion signals and audio signals. In

(Putri et al., 2018), the authors proposed the clas-

siﬁcation of PD by combining voice recordings and

EMG signals using an Adaptive Neuro-Fuzzy Infer-

ence System (ANFIS) and artiﬁcial neural networks

(ANN). In (Yuvaraj et al., 2018), the authors used

the EEG signal by combining higher order spectra

(HOS) with some fuzzy k-nearest neighbor classiﬁca-

tion techniques (fuzzy K-nearest neighbor: FKNN), k

more near neighbors (K-nearest neighbor: KNN) and

the naive bayes approach (NB). In (Elamvazuthi et al.,

2015), ANN is combined with linear prediction co-

efﬁcients (LPCs) to classify neuromuscular disorders

(myopathic and neuropathic).In (Bengacemi et al.,

2021a), the authors have proposed the modelling of

muscle activity (AC) and non activity (NA) using

HMM models combined with wavelet analysis for

sEMG signal segmentation in in regions of AC and

non NA applied in PD. In the present work, we pro-

pose to extend the previous work using four HMM

models with adding a decision step for simultaneously

sEMG segmentation and PD diagnosis.

The HMMs have been widely employed and in-

vestigated in the automatic speech recognition. Re-

cently, it has been successfully used for both medical

monitoring and diagnosis system applications such as

ECG classiﬁcation (Patil et al., 2017), EEG classiﬁ-

cation (Jiang et al., 2019), electrical appliances iden-

tiﬁcation (Nait-Meziane et al., 2016). Especially, this

method is also used for the PD classiﬁcation using

the raw gait data (Khorasani and Daliri, 2014). The

HMM has been also combined with support vector

machine (SVM) classiﬁer for natural gesture recog-

nition using EMG signals for upper limb prostheses

control (Rossi et al., 2015). In (Kwon et al., 2007),

the authors have combined the multilayer perceptrons

(MLP’s) and the HMM for classifying six motions

based on EMG signals. A HMM based classiﬁer is

used for speech recognition using myoelectric signals

from the muscles of vocal articulation (Chan et al.,

2002). In (Liu et al., 2015), the authors have used

the HMM on EMG signals to measure the EMG burst

presence probability (EBPP). The study was limited

to simulated signals and to one experimental signal

just for illustration purpose. In (Bengacemi et al.,

2021a), the authors have showed the effectiveness of

HMM for PD modelling where the authors have clas-

siﬁed the EMG activity (AC) versus the no EMG ac-

tivity (NAC). In this paper, we use HMM to classify

EMG activity versus no EMG activity, in which we

have deﬁned four classes ACN, NAN, ACP and NAP

(P: Parkinson and N: Normal). The proposed system

segments the signals on sequence of ACN and NAN

or on sequence of ACP and NAP, then makes a deci-

sion about the class of the EMG signal by verifying

the class type of the sequence P or N.

The proposed system uses the extracted features

from the EMGs signals recorded within the frame-

work of the ECOTECH project (Buttelli, 2012). The

extraction techniques used in this work are the linear

prediction (LP) coefﬁcients, the fractional prediction

coefﬁcients (FLP) and the Discrete Wavelet decompo-

sition based calculus Energy (DWE), Log Wavelet de-

composition based Energy (LWE) and Wavelet Cep-

stral Coefﬁcients (WCC)(Bengacemi et al., 2021a).

The main task consists of looking for optimal param-

eters of HMM and wavelets descriptors to achieve the

best surface EMG signals segmentation and classiﬁ-

cation. The proposed system is carried out in two

phases, namely: the learning phase and the evaluation

and test phase. The ﬁrst phase is reserved to model

the different classes, while the test phase is used to

evaluate the performance of the diagnostic system.

The main contributions of our work are: ﬁrstly,

this work exploits the principal advantage of wavelet

decomposition that is better adapted for extracting

the impulsive information of the action potentials

(AP) of the motor units (MU), especially in PD case.

Secondly, this work adapts the HMM to automatic

sEMG signal segmentation and classiﬁcation. Note

that HMM is one of the best tools to model signal

state transitions which is considered as the supervised

PD diagnostic task (Bengacemi et al., 2021a). Finally,

based on the real ECOTECH data base, we provide a

high performance analysis using two evaluation crite-

rion. This proposed approach is carried out in a learn-

ing and a test phases. The learning phase consists in

modeling the four classes ACN, NAN, ACP and NAP,

while the test phase aims to evaluate the performance

of the classiﬁcation systems using the HMM. These

two phases require a step of extracting discriminating

parameters from the four classes.

The rest of the paper is organized as follows: sec-

tion 2 describes the problem formulation and the pro-

posed methodology. Section 3 is dedicated to the per-

formance analysis and discussions while section 4 is

reserved for the concluding remarks.

Surface EMG Signal Segmentation and Classiﬁcation for Parkinson’s Disease Based on HMM Modelling

931

2 MATERIAL AND METHOD

In this section, we present in details the methodology

of our proposed work. More precisely, after deﬁn-

ing thoroughly the problem formulation, we introduce

gradually the proposed method starting by the perfor-

mance evaluation Tools. Then, the description of the

used database of sEMG signal has been presented.

Then, we present the used features extraction tech-

niques for EMG signal modeling.

2.1 Problem Formulation

In this paper, we present the PD diagnostic system

based on the HMM models. These have already been

used and tested as an sEMG signal segmentation tech-

nique, in (Bengacemi et al., 2021a). However, in this

work, we use this method for a task of diagnosing and

classifying PD. In this system, we used the HMMs for

the both segmentation and classiﬁcation of the sEMG

signal for the PD diagnostic task. The proposed ap-

proach involves transcribing the sEMG signal of a

control and parkinsonian subjects into a sequence of

activity zone followed by a non-activity zone. Four

classes are considered in this work such as ACP and

NAP for Parkinsonian subject and ACN and NAN for

a control subject.

(1) ACN: EMG activity class for a control subject;

(2) NAN: non-activity class (noise region) for a con-

trol subject;

(3) ACP: EMG activity class for a Parkinson’s pa-

tient;

(4) NAP: non-activity class (noise region) for a

Parkinson’s patient.

We consider K observations (samples)

{

x[n]

}

n=1:K

of sEMG signal. Given a chosen analysis frame

length, these measurements are divided into overlap-

ping

frames. For each signal’s frame, we are inter-

ested in determining whether it contains a signal s[n]

embedded in a random background noise w[n] (EMG

activity) or, on the contrary, it is just the confusing

manifestation of the noise (no EMG activity). Hence,

we have a decision problem expressed as:

Γ :











: no EMG activity for control subject

: EMG activity for control subject

: no EMG activity for Parkinsonian subject

: EMG activity for Parkinsonian subject

(1)

This task is known as sEMG signal segmentation

and classiﬁcation based on the HMM approach which

In this work, we used 50% overlapping windows.

allows us to classify a signal either in class P (parkin-

sonian) or in class N (control) according to the type of

the sequence of the areas ACP, NAP or ACN, NAN.

Thus, this system makes it possible to segment the

signal into activity zones and non-activity zones, and

also to perform the task of diagnosing Parkinson’s dis-

ease. In the following, we will present the methodol-

ogy of diagnosis and classiﬁcation of PD adopted in

this present system.

Figure 1: Segmentation and classiﬁcation of EMG signal

based on HMM models.

In the learning phase, each class (ACN, NAN or

ACP, NAP) of region is modeled by an HMM model

of N

states

states, each state being represented by a

GMM model of N

GMM

Gaussian with a diagonal co-

variance matrix. The parameters of the HMM and

GMM models are estimated using the HEREST com-

mand of the HTK tool, applied to the sequence of fea-

ture vectors extracted from the EMG signals from the

database of the ECOTECH project (Buttelli, 2012). In

addition, this estimation requires the transcription of

the reference text which contains the class sequence

of each signal. The feature vector sequences are ex-

tracted using the LPC, FLP, DWE, LWE and WCC

descriptors, applied to each surface EMG signal.

In the segmentation and classiﬁcation phase, the

HVITE command of the HTK tool uses the trained

HMM models and the constraint model (language

model) to transcribe each input sequence of features

vectors into a sequence of classes (ACN, NAN) or a

sequence of classes (ACP,NAP) and detect the bound-

aries of their segments (Young et al., 2006; Chibelushi

et al., 2002). The constraint problem is to accept only

the sequence of classes in which each ACN tag (class)

is followed by the NAN tag for the control subject’s

surface EMG signal and each ACP tag (class) is fol-

lowed by the NAP tag for the surface EMG signal for

the Parkinson’s patient. Then, a decision is made for

SEMG signal classiﬁcation by verifying the class of

obtained sequence, either P (Parkinsonian subjects) or

N (Control subjects).

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

932

2.2 Performance Evaluation Tools

The performance evaluation can be performed using

the HRESULTS command of the HTK tool which

compares each test transcript of an EMG signal to

its corresponding reference transcript (Young et al.,

2006). The result of identifying and classifying seg-

ments is evaluated using the precision Acc deﬁned in

the equation (2).

The precision Acc is used to evaluate the number

of bursts of EMG activity correctly detected, taking

into account the constraints already described previ-

ously.

Acc =

N − D − S−I

(2)

N represents the total number of segment labels in

the reference transcriptions of EMG signals, D is the

number of removed labels, S is the number of substi-

tuted labels and I is the number of the inserted labels.

The classiﬁcation system is composed of the follow-

ing three steps.

For the performance test and evaluation stage, we

adopt the following performance indicators: the clas-

siﬁcation precision (Acc) and the signal classiﬁcation

rate (C RS), deﬁned in the equation (2) and the equa-

tion (3) respectively.

The precision allows us to have an overall mea-

surement calculated from the classiﬁcation of all the

zones of the test base (ACN-NAN or ACP-NAP) inde-

pendently of the signals. However, the diagnostic task

consists in classifying the signals, this diagnostic task

uses the segmentation results which makes it possible

to classify the signal in Parkinson’s class (P) or in con-

trol class (N) according to the type of the sequence of

zones (ACN-NAN or ACP-NAP). The classiﬁcation

rate is calculated using the equation 3 which allows

us to get the perfomance analysis of the proposed di-

agnostic system.

CRS = NsC/NsT (3)

where CRS is signal classiﬁcation rate, NsC is the

classiﬁed signals correctly and NsT is the total num-

ber of signals.

2.3 Surface EMG Signal Database

For this study, nine healthy subjects and eight Parkin-

sonian patients were recruited in the frame of the

French national research project ECOTECH (But-

telli, 2012). This work was approved by the local

ethics committee and subjects provided written con-

sent prior to commencement.

A speciﬁc lower limb muscles of gait activity have

been measured. Patients were prepared for electrodes

placement by shaving the skin and cleaning it with al-

cohol wipes. EMG sensors were placed on the muscle

belly parallel to the main direction of muscle ﬁbres in

accordance with study on the innervation zone (Bar-

bero et al., 2012). Data were collected using an on

board system of wearable sensors (20-450 Hz band-

width, 16 bits per sample, 1926 Hz sampling rate).

Data collection provides several burst activities from

each right soleus muscle corresponding to several gait

cycles. The data base description is reported in tables

1 and 2.

Table 1: Description of sEMG signals for healthy subjects.

Subjects Number of EMG bursts Signal duration (second)

Data base for training phase

Control

22 26.0685

Control

10 11.2128

Control

11 14.3998

Control

11 14.7441

Data base for testing phase

Control

11 11.1635

Control

6 7.7121

Control

6 6.5298

Control

12 14.3458

Control

26 28.5702

Table 2: Description of sEMG signals for Parkinsonian sub-

jects.

Subjects Number of EMG bursts Signal duration (second)

Data base for training phase

Park

10 28.5702

Park

10 9.1317

Park

5 4.8657

Park

37 39.6152

Data base for testing phase

Park

10 11.2876

Park

9 8.9152

Park

5 4.4742

Park

5 4.6487

2.4 Features Extraction and EMG

Signal Modeling

The feature extraction plays a critical role to get a ro-

bust diagnosis system. This process transforms the

raw sEMG signal into a feature vector. Generally, the

used features in EMG signals analysis can be divided

into three categories: time domain, frequency do-

main and time-frequency domain features (Hogan and

Mann, 1980; Tsai et al., 2014; Englehart et al., 1999).

As a particular class within the time-frequency meth-

ods, the time-scale methods have gained high interest

because the scale parameter provides a natural analy-

sis of biological phenomena, that is to say a high time

precision for rapid events (low scales) and conversely

a poor time precision with high frequency precision

for slow events (high scales). Moreover, they show a

high tuning ﬂexibility in their design useful for per-

formance seeking. In our work, we are interested in

the use of Discrete Wavelet transform, particularly the

Wavelet Cepstral Coefﬁcient (WCC) coefﬁcients. In

this study, we have also analysed the Discrete Wavelet

Energy (DWE) normalized on total energy of win-

Surface EMG Signal Segmentation and Classiﬁcation for Parkinson’s Disease Based on HMM Modelling

933

dow analysis, the logarithm of wavelet energy (LWE)

and the Wavelet Cepstral Coefﬁcient (WCC) com-

puted from the discrete cosine transform (DCT) of

LWE (Hacine-Gharbi and Ravier, 2018) (see Fig.2).

All the features were calculated using the discrete

wavelet transform (DWT) which mother wavelet is

characterized by two digital low-pass and high-pass

ﬁlters. The DWT provides coefﬁcients by an iter-

ative down sampling-ﬁltering procedure achieved at

successive scales beginning on the N-length original

signal x[n] up to a desired decomposition level L

decomp

(that should be lower than the maximum decomposi-

tion level L

max

= log

[n] provided N is a power of 2

or rounded to its nearest high value). The iterative

procedure extracts the set of wavelet coefﬁcients d

[n]

at each scale i from 1 up to L

decomp

value plus the

decomp

[n] at the last scale. Then the DWE features

E(d

) and E(a

decomp

) are composed of the energy val-

ues calculated as the squared absolute magnitude sum

of the wavelet coefﬁcients at each scale i. The LWE

are the log of the DWE coefﬁcients previously nor-

malized by the number of samples per scale. Finally,

the WCC coefﬁcients are obtained by applying the

inverse DCT on the previous LWE coefﬁcients for

decorrelation. In this work, we investigated the im-

pact of these descriptors on the performance results of

PD classiﬁcation. Hence different experiences have

been carried out to search for the optimal conﬁgura-

tion. These features are widely used in: speech recog-

nition (Lei and Kun, 2016; Adam et al., 2013), Elec-

trical Appliances Identiﬁcation (Hacine-Gharbi and

Ravier, 2018; Bengacemi et al., 2021a; Bengacemi

et al., 2021b) and Recognition of Heart Sound (Xiong

et al., 2019).

In this work, we have also integrated the dynamic

features which are the ﬁrst order differential coefﬁ-

cients (also called delta ∆) and second order (called

delta-delta ∆∆), resulting from the initially calcu-

lated coefﬁcients DWE, LWE and WCC, considered

as static features. The static’s features contain only

information on a given frame. In order to improve the

representation of the frame’s information, it is often

proposed to introduce new features in the vector of

features. (Furui, 1981; Furui, 1986) proposed the use

of dynamic features which present the spectral transi-

tion information in the signal. The dynamic features

are calculed using HCopy command of the HTK tools

library (Hidden Markov Model Toolkit).

Let C

(t) is the extracted feature k of frame t,

then the corresponding differential coefﬁcient ∆C

calculated on 2η

∆

analysis frames by estimating the

slope of the linear regression of the coefﬁcient C

time t (Young et al., 2006):

∆C

(t) =

i=+η

∆

∑

i=−η

∆

i.C

(t + i)

i=+η

∆

∑

i=−η

∆

(4)

The second order differential coefﬁcients ∆∆ (delta-

delta or acceleration) are calculated in the same way

from the ﬁrst order coefﬁcients.

The proposed system can be seen as pattern recog-

nition system which requires a training and recogni-

tion phases. The ﬁrst one is used for modelling the

temporal pattern classes and the second one is used

for Parkinson’s diseases classiﬁcation. Hence, both

phases require feature extraction step to convert each

signal in sequence of features vectors obtained by di-

viding the signal into overlapping windows and com-

puting from each window a set of features that consti-

tutes the feature vector (see Fig.2). This sequence of

vectors can be considered as input sequence of obser-

vations in modelling or classiﬁcation steps.

Figure 2: Features extraction steps (Bengacemi et al.,

2021b).

3 RESULTS AND DISCUSSION

In this paragraph, we evaluate the performance of the

proposed HMM approach with the different feature

extraction techniques. Like any classiﬁcation system,

we have two phases, the ﬁrst is the learning phase and

the second is the testing phase. The ﬁrst phase is per-

formed with a set of nine real EMG signals, made up

of four signals from Parkinson’s subjects and ﬁve sig-

nals from control subjects. In the second phase, we

take four signals from Parkinson’s subjects and four

signals from control subjects, as shown in the tables 1

and 2.

Several experiments are carried out to ﬁnd the op-

timal conﬁguration which gives the best performance

by studying: (i) The parameters of the HMM mod-

els (number of states and the number of Gaussians),

(ii) The study of different types of descriptors for fea-

ture extraction (iii) The optimal combination of the

mother wavelet with the level of decomposition. The

three experiments are processed in this order:

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

934

Table 4: Comparison of performances in Acc % and CRS % for the LWE descriptor using ’Coi f 5’ for different analysis

frame’s duration.

The analysis frame duration (ms) 16.61 20 30 33.22 40 50 60 66.45 70 80 90 100 110 120 130 132.91 140 150

Acc % 61,01 53,46 51,57 67,92 72,96 79,25 69,18 44,03 68,55 99,37 68,55 83,02 72,33 84,91 46,54 47,80 69,18 32,08

CRS % 75.00 62.50 75.00 75.00 75.00 87.50 75.00 62.50 75.00 100 75.00 87.50 75.00 87.50 62.50 62.50 75.00 50.00

(1) Comparison of the performance of the LPC, FLP,

DWE, DWE and WCC descriptors;

(2) Finds the optimal duration of the analysis frame;

(3) Finding the best combination between the mother

wavelet and the level of decomposition.

(4) Analyse the performance of the obtained results.

For the ﬁrst experiment, we have varied the num-

ber of states HMM N

states

, the number of Gaussians

GMM

and the level of wavelet decomposition L

decomp

for wavelet analysis

, The order P for the LPC de-

scriptor and the order L for the FLP descriptor. The

best descriptor with the optimal parameters found are

used for the second experiment to study the duration

of the analysis frame. Next, we look for the optimal

conﬁguration for the mother wavelet.

3.1 Performances Comparison of

Different Descriptors

This part presents the performance evaluation results

of ﬁve descriptors namely LPC, FLP, DWE, LWE and

WCC for the diagnosis of PD. In this experiment, we

are looking for the optimal conﬁguration that gives

the best performance in terms of Acc and CRS. For

each descriptor, we vary the number of states for

each experiment N

states

= (2,3,4, 5,6, 7,8, 9,10), the

number of components of Gaussians in GMM mod-

eling N

GMM

= (1,2,3,6, 12,24, 48), the order P =

(2,3,4, 5,6, 7,8, 9,10) of the LPC descriptor and the

order L of the FLP descriptor.

The obtained results are presented in the table 3

which show the optimal conﬁgurations in terms of

number of Gaussians N

GMM

and number of states

states

for each wavelet analysis descriptor DWE,

LWE and WCC with the optimal order P for the de-

scriptor LPC and the optimal order L for the FLP de-

scriptor. These results demonstrate the gain in perfor-

mance of the wavelet analysis descriptors which show

an Acc greater than 88 % and a CRS greater than 87

% compared to the LPC and FLP descriptors. In par-

ticular, the descriptor LW E with Acc = 98.11% and

CRS = 100% for N

GMM

= 6 and N

states

= 2.

The mother wavelet ’Coi f 5’ was chosen with an anal-

ysis frame duration equal to 66.45 ms and a wavelet de-

composition level L

decomp

= 4, found as optimal parameters

in the segmentation by HMM modelling (Bengacemi et al.,

2021a).

Table 3: Comparison of performance in Acc % and CRS %

for DWE, LPC, LWE and WCC descriptors using ’Coi f 5’

and L

decomp

= 4 with a the analysis frame equal to 66.45

ms.

Descriptors LPC FLP DWE LW E WCC

Optimal parameters N

GMM

= 24 N

GMM

= 6 N

GMM

= 3 N

GMM

= 6 N

GMM

= 12

states

= 3 and P = 3 N

states

= 3 and L = 5 N

states

= 3 N

states

= 2 N

states

= 3

Acc % 66.97 83.65 88.05 98.11 91.82

CRS % 75 75 87.5 100 87.5

3.2 Inﬂuence of the Duration of the

Analysis Window

After having chosen the descriptor LW E, we study in

this paragraph the duration of the appropriate analy-

sis frame taking into account the advantages of the

wavelet analysis, which is appropriate for the non-

stationarity of EMG signals. We vary the analysis

frame duration for the mother wavelet ’C oi f 5’, the

number of GMMs N

GMM

= 6, the number of states

states

= 2 and the level of decomposition equal to

decomp

= 4 for the descriptor LWE. The table 4

shows the values of Acc for each value of the duration

of the analysis frame. The best performance is ob-

tained for an analysis frame duration equal to 80 ms,

which corresponds to the values of Acc equal to 99.37

% and CRS equal to 100 %. This analysis frame du-

ration is used in the performance analysis for the op-

timal choice of the mother wavelet.

3.3 Choice of the Mother Wavelet and

Decomposition Level

Several studies on surface EMG analysis have con-

cluded that the Daubechies wavelet family (Db) is the

most suitable wavelet for the analysis of the EMGs

signal (Hussain et al., 2009; Mahaphonchaikul et al.,

2010; Phinyomark et al., 2009). In (Too et al., 2018),

the authors concluded that the ’Sym4’ is the most suit-

able for EMG pattern recognition. In (Bengacemi

et al., 2021b), we found that Coi f 5 is the most suit-

able for the segmentation of the surface EMG signal.

This part of the study aims to select the optimal order

of the mother wavelets within its family for a dura-

tion of the analysis frame equal to 80 ms, a number of

GMM N

GMM

= 6 , a number of states N

states

= 2 with

a decomposition level varying between 1 to log2(N)

(N is the number of samples in the analysis window

(max level = 7)). In this study, we consider the fol-

lowing wavelet families:

Surface EMG Signal Segmentation and Classiﬁcation for Parkinson’s Disease Based on HMM Modelling

935

• The Daubechies family with orders 1 to 8: Db1,

Db2, ... , Db10;

• The Symlets family with orders 1 to 8: Sym1,

Sym2, ..., Sym8;

• The Coiﬂets family with orders 1 to 5: Coif1,

Coif2,..., Coif5.

The obtained results of Acc and L

decomp

are reported

in the tables 5, 6 and 7 for each of the three wavelet

family, respectively from which we can note that the

best results:

- For the wavelet family Daubechies (see the table

5), we notice that the average of the classiﬁcation

precision Acc is greater than 87 % and the average

of the classiﬁcation rate CRS is greater than 87%.

We also notice that for Db7, Db8 and Db9 with

decomp

= 4, we have a Acc = 99.37 and a CRS =

100%.

- For the wavelet family Symlets (see the table 6),

we notice that the average of the classiﬁcation

rates Acc is greater than 87 % and the average

of the classiﬁcation rate CRS is greater than 89%.

We also notice that for Sym4 and L

decomp

= 2 we

have a Acc = 97.48 and a CRS = 100%.

- For the wavelet family Coiﬂets (see the table 7),

we notice that the average of the classiﬁcation

rates Acc is greater than 96 % and the average

of the classiﬁcation rate CRS is greater than 97%.

We also notice that for Coi f 5 and L

decomp

= 4

leads to an Acc = 99.37 and a CRS = 100%.

The obtained results demonstrate the robustness of the

performance of the proposed approach in terms of the

value of Acc and of CRS where we notice that all the

mean values of Acc and CRS are greater than 87 % . In

particular, we notice that the wavelet family ’Coiﬂets’

gives an average Acc greater than 96 % and an aver-

age CRS 97 %. The mother wavelets Db7, Db8, Db9

and Coi f 5 with Ldecomp = 4 give the same values of

Acc = 99.37 and of CRS = 100%. This improvement

of the results is obtained through the various exper-

iments described previously without any exhaustive

empirical calculation.

Table 5: The performance in terms of Acc %, CRS % and

decomp

optimal for LWE using the Daubechies wavelet

family.

Daubechies Db1 Db2 Db3 Db4 Db5 Db6 Db7 Db8 Db9 Db10 moyenne

80 ms

decomp

2 5 3 2 5 2 4 4 4 2 //

Acc % 80.50 86.16 82.39 74.21 79.87 91.82 99.37 99.37 99.37 83.02 87.60

CRS % 87.5 75 87.5 75 75 87.5 100 100 100 87.5 87.5

Table 6: The performance in terms of Acc %, CRS % and

decomp

optimal for LWE using the Symlets wavelet family.

Symlets Sym1 Sym2 Sym3 Sym4 Sym5 Sym6 Sym7 Sym8 moyenne

80 ms

decomp

2 5 3 2 4 4 4 2 //

Acc % 80.50 86.16 82.39 97.48 86.16 83.65 83.65 96.23 87.02

CRS % 87.5 75 87.5 100 87.5 87.5 87.5 100 89.06

Table 7: The performance in terms of Acc %, CRS % and

decomp

optimal for LWE using the Coiﬂets wavelet family.

Coiﬂets Coi f 1 Coi f 2 Coi f 3 Coi f 4 Coi f 5 moyenne

80 ms

decomp

1 4 2 2 4 //

Acc % 91.19 97.48 97.48 97.48 99.37 96.60

CRS % 87.5 100 100 100 100 97.50

4 CONCLUSION

The classiﬁcation and diagnosis of diseases has im-

portant clinical application. This paper describes the

new diagnostic systems for dealing with the classiﬁ-

cation problem of Parkinson’s disease. The proposed

system is based on HMM modeling and wavelet anal-

ysis which is most suitable for non-stationary signals,

especially surface EMG signals. The results show

the HMM system achieved good classiﬁcation accu-

racy of ACN,NAN,ACP,NAP (segmentation) and sig-

nal classiﬁcation rate (diagnostic) suitable for clini-

cal applications. The evaluation of diagnostic per-

formance is performed by performing various experi-

ments on surface EMG signals. The proposed system

gives better results which lead to an Acc of 99.37 %

and a CRS of 100 %. Consequently, this proposed

approach represents an appropriate solution for the

analysis of EMGs signals and its use for both EMG

signal segmentation and diagnostic purposes, in par-

ticular Parkinson’s disease. We have also seen the ef-

fectiveness of the HMM method for PD.

ACKNOWLEDGMENTS

The present paper used collected data from the French

national project ECOTECH supported by the French

National Agency for research under the contract No.

ANR-12-TECS-0020.

REFERENCES

Abdulhay, E., Arunkumar, N., Narasimhan, K., Vellaiap-

pan, E., and Venkatraman, V. (2018). Gait and tremor

investigation using machine learning techniques for

the diagnosis of parkinson disease. Future Genera-

tion Computer Systems, 83:366–373.

Adam, T., Salam, M., and Gunawan, T. S. (2013). Wavelet

based cepstral coefﬁcients for neural network speech

recognition. In 2013 IEEE International Conference

on Signal and Image Processing Applications, pages

447–451. IEEE.

Barbero, M., Merletti, R., and Rainoldi, A. (2012). Atlas

of muscle innervation zones: understanding surface

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

936

electromyography and its applications. Springer Sci-

ence & Business Media.

Bengacemi, H., Gharbi, A. H., Ravier, P., Abed-Meraim,

K., and Buttelli, O. (2021a). Surface emg signal seg-

mentation based on hmm modelling: Application on

parkinson’s disease. ENP Engineering Science Jour-

nal, 1(1):63–74.

Bengacemi, H., Hacine-Gharbi, A., Ravier, P., Abed-

Meraim, K., and Buttelli, O. (2021b). Surface emg

signal classiﬁcation for parkinson’s disease using wcc

descriptor and ann classiﬁer. In ICPRAM, pages 287–

294.

Bhoi, A. K. (2017). Classiﬁcation and clustering of parkin-

son’s and healthy control gait dynamics using LDA

and K-means. International Journal Bioautomation,

21(1).

Buttelli, O. (2012). Agence nationale de la recherche.

http://www.agence-nationale-recherche.fr/Projet-

ANR-12-TECS-0020.

Carletti, T., Fanelli, D., and Guarino, A. (2006). A new

route to non invasive diagnosis in neurodegenerative

diseases? Neuroscience letters, 394(3):252–255.

Chan, A. D., Englehart, K., Hudgins, B., and Lovely, D. F.

(2002). Hidden markov model classiﬁcation of my-

oelectric signals in speech. IEEE Engineering in

Medicine and Biology Magazine, 21(5):143–146.

Chibelushi, C. C., Deravi, F., and Mason, J. S. (2002). A

review of speech-based bimodal recognition. IEEE

transactions on multimedia, 4(1):23–37.

Dastgheib, Z. A., Lithgow, B., and Moussavi, Z. (2012).

Diagnosis of parkinson’s disease using electrovestibu-

lography. Medical & biological engineering & com-

puting, 50(5):483–491.

Elamvazuthi, I., Duy, N., Ali, Z., Su, S., Khan, M. A., and

Parasuraman, S. (2015). Electromyography (EMG)

based classiﬁcation of neuromuscular disorders using

multi-layer perceptron. Procedia Computer Science,

76:223–228.

Englehart, K., Hudgins, B., Parker, P. A., and Stevenson,

M. (1999). Classiﬁcation of the myoelectric signal

using time-frequency based representations. Medical

Engineering and Physics, 21(6):431–438.

Furui, S. (1981). Cepstral analysis technique for automatic

speaker veriﬁcation. IEEE Transactions on Acoustics,

Speech, and Signal Processing, 29(2):254–272.

Furui, S. (1986). Speaker-independent isolated word recog-

nition using dynamic features of speech spectrum.

IEEE Transactions on Acoustics, Speech, and Signal

Processing, 34(1):52–59.

Hacine-Gharbi, A. and Ravier, P. (2018). Wavelet cepstral

coefﬁcients for electrical appliances identiﬁcation us-

ing hidden markov models. In ICPRAM, pages 541–

549.

Hausdorff, J. M., Cudkowicz, M. E., Firtion, R., Wei, J. Y.,

and Goldberger, A. L. (1998). Gait variability and

basal ganglia disorders: Stride-to-stride variations of

gait cycle timing in parkinson’s disease and hunting-

ton’s disease. Movement disorders, 13(3):428–437.

Hausdorff, J. M., Mitchell, S. L., Firtion, R., Peng, C.-

K., Cudkowicz, M. E., Wei, J. Y., and Goldberger,

A. L. (1997). Altered fractal dynamics of gait: re-

duced stride-interval correlations with aging and hunt-

ington’s disease. Journal of applied physiology,

82(1):262–269.

Henmi, O., Shiba, Y., Saito, T., Tsuruta, H., Takeuchi, A.,

Shirataka, M., Obuchi, S., Kojima, M., and Ikeda, N.

(2009). Spectral analysis of gait variability of stride

interval time series: comparison of young, elderly

and parkinson’s disease patients. Journal of Physical

Therapy Science, 21(2):105–111.

Hogan, N. and Mann, R. W. (1980). Myoelectric

signal processing: Optimal estimation applied to

electromyography-part ii: Experimental demonstra-

tion of optimal myoprocessor performance. IEEE

Transactions on Biomedical Engineering, (7):396–

410.

Hussain, M., Reaz, M. B. I., Mohd-Yasin, F., and Ibrahimy,

M. I. (2009). Electromyography signal analysis using

wavelet transform and higher order statistics to deter-

mine muscle contraction. Expert Systems, 26(1):35–

48.

Jiang, D., Lu, Y.-n., Yu, M., and Yuanyuan, W. (2019). Ro-

bust sleep stage classiﬁcation with single-channel eeg

signals using multimodal decomposition and hmm-

based reﬁnement. Expert Systems with Applications,

121:188–203.

Khorasani, A. and Daliri, M. R. (2014). Hmm for classi-

ﬁcation of parkinson’s disease based on the raw gait

data. Journal of medical systems, 38(12):147.

Kugler, P., Jaremenko, C., Schlachetzki, J., Winkler, J.,

Klucken, J., and Eskoﬁer, B. (2013). Automatic

recognition of parkinson’s disease using surface elec-

tromyography during standardized gait tests. In 2013

35th Annual International Conference of the IEEE En-

gineering in Medicine and Biology Society (EMBC),

pages 5781–5784. IEEE.

Kwon, J.-W., Kim, J.-H., and Choi, H.-H. (2007). Classiﬁ-

cation of the emg signal using cascaded classiﬁer. In

World Congress on Medical Physics and Biomedical

Engineering 2006, pages 1222–1225. Springer.

Lei, L. and Kun, S. (2016). Speaker recognition using

wavelet cepstral coefﬁcient, i-vector, and cosine dis-

tance scoring and its application for forensics. Journal

of Electrical and Computer Engineering, 2016.

Liu, J., Ying, D., and Rymer, W. Z. (2015). EMG burst pres-

ence probability: A joint time–frequency representa-

tion of muscle activity and its application to onset de-

tection. Journal of Biomechanics, 48(6):1193–1197.

Mahaphonchaikul, K., Sueaseenak, D., Pintavirooj, C.,

Sangworasil, M., and Tungjitkusolmun, S. (2010).

EMG signal feature extraction based on wavelet trans-

form. In ECTI-CON2010: The 2010 ECTI Interna-

tional Confernce on Electrical Engineering/Electron-

ics, Computer, Telecommunications and Information

Technology, pages 327–331. IEEE.

Manwatkar, A. P., Salwe, S., Bagade, A., and Raut, R. A

review on detection of parkinson’s disease.

Miller, R. A., Thaut, M. H., McIntosh, G. C., and Rice,

R. R. (1996). Components of EMG symmetry and

variability in parkinsonian and healthy elderly gait.

Surface EMG Signal Segmentation and Classiﬁcation for Parkinson’s Disease Based on HMM Modelling

937

Electroencephalography and Clinical Neurophysiolo-

gy/Electromyography and Motor Control, 101(1):1–7.

Nait-Meziane, M., Hacine-Gharbi, A., Ravier, P., Lamar-

que, G., Le Bunetel, J.-C., and Raingeaud, Y. (2016).

Hmm-based transient and steady-state current sig-

nals modeling for electrical appliances identiﬁcation.

In Proceedings of the 5th International Conference

on Pattern Recognition Applications and Methods,

pages 670–677. SCITEPRESS-Science and Technol-

ogy Publications, Lda.

Nazmi, N., Abdul Rahman, M. A., Yamamoto, S.-I., Ah-

mad, S. A., Zamzuri, H., and Mazlan, S. A. (2016).

A review of classiﬁcation techniques of EMG signals

during isotonic and isometric contractions. Sensors,

16(8):1304.

Okamoto, M., Matsubara, Y., Shima, K., and Tsuji, T.

(2009). EMG pattern classiﬁcation using hierarchi-

cal network based on boosting approach. Interna-

tional Journal of innovative computing, information

and control, 5(12):4935–4943.

Oung, Q. W., Muthusamy, H., Basah, S. N., Lee, H., and

Vijean, V. (2018). Empirical wavelet transform based

features for classiﬁcation of parkinson’s disease sever-

ity. Journal of medical systems, 42(2):29.

Patil, D. D., Singh, R., Thakare, V. M., and Gulve, A. K.

(2017). Automatic classiﬁcation of ecg arrhythmia us-

ing morphological parameters with hmm and svm. In-

ternational Journal of Applied Engineering Research,

12(20):10376–10384.

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

(2009). An optimal wavelet function based on wavelet

denoising for multifunction myoelectric control. In

2009 6th International Conference on Electrical Engi-

neering/Electronics, Computer, Telecommunications

and Information Technology, volume 2, pages 1098–

1101. IEEE.

Porter, E., Roussakis, A.-A., Lao-Kaim, N. P., and Pic-

cini, P. (2020). Multimodal dopamine transporter

(dat) imaging and magnetic resonance imaging (mri)

to characterise early parkinson’s disease. Parkinson-

ism & Related Disorders.

Putri, F. T., Ariyanto, M., Caesarendra, W., Ismail, R., Pam-

budi, K. A., and Pasmanasari, E. D. (2018). Low cost

parkinson’s disease early detection and classiﬁcation

based on voice and electromyography signal. In Com-

putational Intelligence for Pattern Recognition, pages

397–426. Springer.

Raut, R. K. and Gurjar, A. A. (2015). Bio-medical (EMG)

signal feature extraction using wavelet transform for

design of prosthetic leg. International Journal of Elec-

tronics, Communication and Soft Computing Science

& Engineering (IJECSCSE), 4:81.

Rosenblum, S., Samuel, M., Zlotnik, S., Erikh, I., and

Schlesinger, I. (2013). Handwriting as an objective

tool for parkinson’s disease diagnosis. Journal of neu-

rology, 260(9):2357–2361.

Rossi, M., Benatti, S., Farella, E., and Benini, L. (2015).

Hybrid emg classiﬁer based on hmm and svm for hand

gesture recognition in prosthetics. In 2015 IEEE Inter-

national Conference on Industrial Technology (ICIT),

pages 1700–1705. IEEE.

Sugavaneswaran, L., Umapathy, K., and Krishnan, S.

(2012). Ambiguity domain-based identiﬁcation of al-

tered gait pattern in ALS disorder. Journal of neural

engineering, 9(4):046004.

Surangsrirat, D., Thanawattano, C., Pongthornseri, R.,

Dumnin, S., Anan, C., and Bhidayasiri, R. (2016).

Support vector machine classiﬁcation of parkinson’s

disease and essential tremor subjects based on tempo-

ral ﬂuctuation. In Engineering in Medicine and Biol-

ogy Society (EMBC), 2016 IEEE 38th Annual Inter-

national Conference of the, pages 6389–6392. IEEE.

Too, J., Abdullah, A., Saad, N. M., Ali, N. M., and Musa,

H. (2018). A detail study of wavelet families for EMG

pattern recognition. International Journal of Electri-

cal and Computer Engineering (IJECE), 8(6):4221–

4229.

Tsai, A.-C., Hsieh, T.-H., Luh, J.-J., and Lin, T.-T. (2014).

A comparison of upper-limb motion pattern recogni-

tion using EMG signals during dynamic and isometric

muscle contractions. Biomedical Signal Processing

and Control, 11:17–26.

Wendling, B. (2008). Gait analysis of left and right stride

intervals in neurodegenerative diseases.

Xiong, L., Zhao, Z., Pan, J., Yang, H., and Wang, W. (2019).

Recognition of heart sound based on wavelet cepstrum

coefﬁcient and probabilistic neural network.

Xu, J. and Zhang, M. (2019). Use of magnetic resonance

imaging and artiﬁcial intelligence in studies of diag-

nosis of parkinson’s disease. ACS chemical neuro-

science, 10(6):2658–2667.

Young, S., Evermann, G., Gales, M., Hain, T., Kershaw, D.,

Moore, G., Odell, J., Ollason, D., Povey, D., Valtchev,

V., et al. (2006). The htk book (for htk version. 3.3),

cambridge university engineering department, 2005.

URL http://htk. eng. cam. ac. uk/docs/docs. shtml.

Yuvaraj, R., Acharya, U. R., and Hagiwara, Y. (2018).

A novel parkinson’s disease diagnosis index using

higher-order spectra features in eeg signals. Neural

Computing and Applications, 30(4):1225–1235.

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

938