Chaos, Solitons and Fractals 158 (2022) 112036

Contents lists available at ScienceDirect

Chaos, Solitons and Fractals

Nonlinear Science, and Nonequilibrium and Complex Phenomena
journal homepage: www.elsevier.com/locate/chaos

Alcoholic EEG signals recognition based on phase space dynamic and

geometrical features
Muhammad Tariq Sadiq a,⁎,1, Hesam Akbari b,1, Siuly Siuly c, Yan Li d, Peng Wen (Paul) e
a
Department of Electrical Engineering, The University of Lahore, Lahore 54000, Pakistan
b
Department of Biomedical Engineering, South Tehran Branch, Islamic Azad University, Tehran, Iran
c
Institute for Sustainable Industries & Liveable Cities, Victoria University, Melbourne 14428, Australia
d
School of Sciences, University of Southern Queensland, Toowoomba, Australia
e
School of Mechanical and Electrical Engineering, University of Southern Queensland, Toowoomba, Australia

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

Article history: Alcoholism is a severe disorder that leads to brain problems and associated cognitive, emotional and behavioral
Received 12 August 2021 impairments. This disorder is typically diagnosed by a questionnaire technique known as CAGE that measures the
Received in revised form 29 January 2022 severity of a drinking problem. This is a time-consuming, onerous, error-prone, and biased method. Hence, this
Accepted 23 March 2022
article aims to establish a novel framework for automatic detecting alcoholism using electroencephalogram
Available online 10 April 2022
(EEG) signals, which can mitigate these issues and help clinicians and researchers. In the proposed framework,
Keywords:
firstly, we explore the phase space dynamic of EEG signals for visualizing the chaotic nature and complexity of
Electroencephalography, EEG signals. Secondly, we discover thirty-four graphical features for decoding the chaotic behavior of normal
Phase space dynamic and alcoholic EEG signals. After that, we investigate thirteen feature selection in combination with eleven ma-
Graphical features chine learning and neural network classifiers to select the best combination for the development of an efficient
Feature selection framework. The experimental results reveal that the proposed method achieves the highest classification perfor-
Novel indexes mance involving 99.16% accuracy, 100% sensitivity and 98.36% specificity for the twenty-three features selected
Classification by Henry gas solubility optimization with feedforward neural network (FFNN). The proposed system provides a
new visual biomarker for alcoholic detection. In addition, we developed two new indexes using clinically relevant
features to distinguish normal and alcoholic classes with a single number. These indexes can help medical teams,
commercial users as well as product developers to develop a real-time system.
© 2022 Elsevier Ltd. All rights reserved.

1. Introduction analysis results can lead to misinformation about medical and psychiat-
ric conditions, since some patients may avoid providing honest informa-
According to World Health Organization (WHO), alcohol use is the tion because of embarrassment or fear of being stigmatized, and also
cause of almost 3.3 million deaths each year, which is around 5.9% of blood test is the invasive approach [4]. Clinically, electroencephalogra-
the world's overall death rate [1]. Approximately, 2 billion people are in- phy (EEG) is an effective approach for analyzing brain patterns due to
toxicated and 81.7 million of them are severe dependent upon alcohol its low cost, portability, being a non-invasive, and high temporal
with observable symptoms [2]. There can be numerous behavioral, resolution. EEG signals represent electrical picture of brain's working
emotional, and physiological effects in the individual due to excessive that reflects various physiological and pathological activities such as al-
consumption of alcohol. Uncontrolled alcohol intake is responsible for coholism.
liver cirrhosis, heart diseases, brain disorders, cancers, road accident Several computerized methods are proposed in the literature for the
deaths and injuries [3,4]. The early detection of alcoholism can prevent early detection of alcoholic EEG signals [6,7]. These methods are catego-
its irremediable consequences, such as disability and mortality [3,5]. To rized in various ways such as, linear and nonlinear features-based
measure alcohol effects on individuals, traditional methods are based on methods in time and frequency-domains; signal decomposition-based
blood tests, questionnaires, and physiological examinations. However, methods; as well as deep learning-based approaches. In previous re-
each of them has its limitations, for instance, the questionnaire-based search work, the linear and nonlinear features were extracted in time
and frequency domains from EEG using hidden Markov models
⁎ Corresponding author.
(HMMs) [8], largest Lyapunov exponent (LLE), approximate entropy
E-mail address: [email protected] (M.T. Sadiq). (ApEn), sample entropy (SamEn), higher-order spectra (HOS) [9,10],
1
Co-first author. granger causality [11] and synchronization likelihood [12] have been

https://doi.org/10.1016/j.chaos.2022.112036
0960-0779/© 2022 Elsevier Ltd. All rights reserved.
M.T. Sadiq, H. Akbari, S. Siuly et al. Chaos, Solitons and Fractals 158 (2022) 112036

extracted from EEG for detecting alcoholic EEG signals. These features The rest of paper is organized as following: Section 2 deals with the
are time-consuming and some of these features were unable to quantify materials and methods. Section 3 outlines the performance evaluation
the complex essence of EEG signals because time-domain features failed parameters. Section 4 describes results. Section 5 provides the discus-
to provide spectral information and frequency-domain approaches sion of our experimental outcomes and finally, Section 6 summarizes
were un-successful to deliver temporal indulgence. the study.
It is generally known that the frequency-related information at dif-
ferent time scales can aid to recognize the chaotic behavior of EEG sig-
nals. Taking this motivation, in the last two decades, various signal 2. Materials and methods
decomposition-based techniques such as wavelet packet transform
(WPT) [10], continues wavelet transform (CWT) [13], tuned-Q wavelet 2.1. EEG alcoholism dataset
transform (TQWT) [14], dual-tree complex wavelet transform (DT-
CWT) [15], three-band orthogonal wavelet filter bank (TBOWFB) [16], In this work, we utilized the open dataset provided by the University
flexible analytical wavelet transform (FAWT) [17], empirical mode de- of California, Irvine (UCI). The utilized dataset arises after a long percep-
composition (EMD) [18], Fourier-Bessel series expansion based empiri- tion on the hereditary inclination of alcoholic subjects and comprises of
cal wavelet transform (FBSE-EWT) [19], empirical wavelet transform two EEG classes alcoholic and normal. This dataset is accessible in three
(EWT) [20] and Fourier intrinsic band functions (FIBFs) [21] have been variations. First is a small dataset that contains two subjects, one for
suggested to obtain a time-frequency representation for the EEG signals. each class with all different stimuli. Two categories of stimuli were uti-
Recently, deep learning methods have been applied to many bio-signal lized during small dataset collection, such as single (s1) and double
processing applications including alcoholic EEG signals detection [22]. stimuli (s1 match s2 and s1 do not match s2). The second dataset is
In the aforementioned literature, some disadvantages or limitations large in which ten subjects of each class with separate training and test-
are found as listed: Firstly, the linear and nonlinear features cannot char- ing data set is used. The third dataset is collected from 122 subjects, each
acterize the system's momentary chaotic behavior, thus, unable to pres- with 120 trials for each subject. These datasets are recorded from each
ent a visual inspection approach for the medical team. Secondly, signal subject utilizing various stimuli having 90 pictures. In the current
decomposition methods may occupy inappropriate frequency order, work, experiments are performed utilizing a small dataset in light of
suffer from mode mixing, end effects, noise effects, narrowly spaced fre- the fact that the openly accessible full dataset is not complete. Few trials
quencies challenges along with selecting a mother wave or shift- are with empty files or labelled as err. The 64 electrodes are used to ob-
invariance complications. Thirdly, the EEG data are recorded from mul- tain EEG recordings. The position of each electrode is adjusted according
tiple channels which may result in a large number of features, however, to 10–20 international montage standard [20,23].
EEG recordings are restricted to small sessions. Due to the small number The dataset comprises EEG signals recorded for 32 s (about 16,400
of features and samples ratio, the curse of dimensionality occurs. samples) at 256 Hz sampling frequency and 12-bit resolution. The ex-
Fourthly, it is extremely difficult to track the variety of selected features periments in this research are carried out on a small dataset, and arti-
between the normal and alcoholic classes to make an accurate diagnosis facts such as eyes and muscular movements are minimized using a
so irrespective of traditional approaches, indexes are needed. baseline filter. Following filtering, only 30 EEG recordings from both
To address the above-mentioned problems, this study aims to intro- classes remain. The extensive 32 s EEG records are split into four equal
duce a new design that can automatically identify alcoholic EEG signals segments of 2048 samples with 8 s window size [24]. It is clear from
from normal EEG signals with higher performances. The proposed de- Fig. 1 that the amplitude of a normal subject is more when contrasted
sign consists of several modules such as (i) Visualize EEG signals with an alcoholic subject since the normal subject's brain is stirred
through Phase Space Dynamics (ii) Discover appropriate Graphical and energized while alcoholic subjects are oblivious. Summary of the
Features for alcoholism detection (iii) design a strategy for selecting op- used dataset is given in Table 1 [20].
timal feature set (iv) find out a sustainable Classification model for the
obtained feature set (v) introduce new Alcoholic Diagnostic indexes.
In this study, for the very first time, we propose the phase space dy-
namic and graphical features to understand the complexity and chaotic
behavior of normal and alcoholic EEG signals. In addition, we investigate
the ability of 13 feature selection (FS) algorithms with 11 classifiers to
decode the most appropriate combination for the development of effec-
tive computerize framework. Another key contribution of this research
is the formulations of two integrated indexes by employing the graphi-
cal features extracted from the phase space dynamics of the EEG signal.
The key contributions of this study are listed as follows:

1. A novel framework is designed and validated for automated identifi-

cation of alcoholic EEG signals.
2. By employing phase space dynamic, EEG signal characteristics are
converted into topological patterns of a geometrical entity embed-
ded in a space representing all possible system states, each state cor-
responds to the single attribute, and this rebuilt space has the same
topological properties as the original space.
3. Introduce several graphical features for the first time in classifying
normal and alcoholic EEG signals.
4. A comprehensive comparison among different combinations of fea-
ture selection (FS) and classifiers is made to decode the best combi-
nation for the development of the computerized system, which is
first of such kind of attempts in alcoholism research field.
5. In addition to the computerized system, two novel indexes are devel-
oped to assist medical team and commercial purposes. Fig. 1. Visual representation of normal and alcoholic EEG signals.

2
M.T. Sadiq, H. Akbari, S. Siuly et al. Chaos, Solitons and Fractals 158 (2022) 112036

Table 1
Description of dataset. An ¼ ½a1 , a2 , a3 , . . . , am−1  ð2Þ
Subjects Stimuli present No. of segments Sampled frequency No. of samples

Normal s1 obj 120 256 2048 Anþ1 ¼ ½a2 , a3 , a4 , . . . , am  ð3Þ

s2 match
s2 no match
Alcoholic s1 obj 120 256 2048 then the phase space dynamics of alcoholic and normal EEG signals are
s2 match obtained by plotting An versus An+1 as shown in Fig. 3.
s2 no match
2.2.2. Module 2: graphical feature extraction
2.2. Methods Non-linear features (e.g., Lyapunov exponent, mutual information
etc.) can measure the dynamics of phase space, but they need time-
The proposed computerized framework comprised of several mod- consuming calculations that are hindrances for practical uses. On the
ules as shown in Fig. 2 and describe in the following sub-sections. contrary, graphical features of EEG signals help calculate the complexity,
variation, scattering of the two dimensional (2D) phase space dynamic
2.2.1. Module 1: phase space dynamics shape, and can quantify the chaotic behavior of EEG signals in both reg-
Alcoholic EEG signals are chaotic in nature because these signals are ular and intoxicated classes with less time-cost. Thus, in this module, we
recorded from brain dynamical system. The visualization of chaotic na- propose to extract graphical features from EEG signals for identifying al-
ture of such EEG signals demand a higher domain representation. Irre- coholic and normal class. This study discovers 34 graphical features for
spective of the time and frequency domains, high dimensional space discrimination between the normal and alcoholic groups. These features
can help obtain useful information for dynamic and complex behaviors are detailed in Table 2 and illustrate in Fig. 4.
of EEG signals. Thus, in this study, we employed the phase space dy-
namic approach for plotting EEG signals in higher dimension space. 2.2.3. Module 3: feature selection
The phase space dynamic can be represented as follows [25], Generally, a feature selection algorithm intends to select a group of
high discriminative features from the original dataset. Among the fea-


phase space dynamic ¼ An , Anþτ , . . . , Anþðd−1Þτ ð1Þ ture selection algorithms, the wrapper methods are the most widely
used, mainly due to their excellent behaviors in solving this complex
where, A = [a1, a2, a3, …, am] represents the input EEG signals with m problem [30,31]. In this technique, the features are represented by the
samples, and n=1,2,…, m − (d − 1)τ. The estimation of time delay vector with S size, where S is the number of original features [32]. The
(τ) and embedded dimension (d) parameters put a major impact on wrapper methods first build a set of random solutions. In the process
computational complexity of phase space dynamic of EEG signals. In this of feature subset evaluation, the learning algorithm is implemented as
research, the values of τ and d are empirically chosen as 1 and 2 respec- an indicator to assess the quality of the solution. Iteratively, the wrapper
tively. After choosing the empirical values, the dynamic of EEG signals methods will evolve the achieved solutions until an end condition is sat-
are displayed in two-dimensional (2D) geometry with a one sample de- isfied [33]. In the end, the optimal feature subset (best solution) is per-
lay. Let An and An+1 are formulated as, ceived.

Fig. 2. Block diagram of the proposed computerized framework.

3
M.T. Sadiq, H. Akbari, S. Siuly et al. Chaos, Solitons and Fractals 158 (2022) 112036

different biomedical signals and to help the healthcare staff [47–49].

Keeping this motivation in mind, two alcoholic indexes for the
diagnosis of alcoholic EEG signals are derived from the graphical fea-
tures of the phase space dynamic of EEG signals after a series of exper-
iments.
We focused on developing an ADI1 by quantifying the area of phase
space dynamic shape by “SCCA”, “SCTA”, “SHCA”, “SDHC”, “SSVL”,
“AOCT”, “SDTC”, “TACR”, “SRCA”, “TDSD”, “ELPA”, and “CTMs” features.
After trial and error manner, the proposed index is formulated as:
rffiffiffiffiffiffiffiffiffiffiffiffi
ELPA
ADI1 ¼ ð4Þ
AOCT

In addition, according to Fig. 3, the points of the 2D shape of the

phase space dynamic for the normal group have more distance from
the 135-degree line and coordinate center compared to the alcoholic
group. For this reason, we propose an ADI2 using “SH45”, “SH135”,
“SDTC”, “TACR”, “SRCA”, and “CTMs” features. After conducting a series
of experiments, another index is developed as follows:

CTM90 þ CTM95
ADI2 ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð5Þ
SH135

Fig. 3. A sample phase space dynamics of (a) normal and (b) alcoholic EEG signals. 3. Performance evaluation parameters

For the evaluation of the proposed framework, the 10-fold cross-

validation (CV) technique is used in the training and testing phase of
To exploiting the significant metaheuristic algorithm as a feature se-
the classifier. Four standard measurements are used for assessing the
lection method, in this study, we tested the ability of 13 algorithms in
performance of classifiers for normal vs. alcoholic EEG signal classifica-
alcoholic detection applications. These feature selection methods are
tion as described below:
named as: binary atom search optimization (ASO) [34], binary dragon-
Accuracy (ACC): the capability of a classifier for the correct classifica-
fly algorithm (BDA) [35], binary differential evolution (BDE) [36], binary
tion of normal and alcoholic EEG signals.
harris hawk optimization (HHO) [37], binary tree growth algorithm
Sensitivity (SEN): the ability of a classifier for the correct detection of
(TGA) [38], binary equilibrium optimizer (BEO) [39], binary genetic al-
alcoholic EEG signals.
gorithm (BGA) [40], binary grey wolf optimization (GWO) [41], binary
Specificity (SPE): the ability of a classifier for the correct detection of
particle swarm optimization (PSO) [42], henry gas solubility optimiza-
normal EEG signals.
tion (HGS) [43], sine cosine algorithm (SCA) [44], salp swarm algorithm
Area under the receiver operating curve (AUC): the ability to test the
(SSA) [45], and whale optimization algorithm (WOA) [46]. Table 3 de-
authenticity of a classifier, where a value close to 1 indicates the truth-
picts the parameter settings of utilized methods. To ensure a fair com-
fulness of the classifier, and vice versa.
parison, each method is using the similar settings of population size =
10 and maximum iterations = 30 in this work. The optimal feature vec-
4. Results
tor arrays obtained by different feature selection methods for best fit-
ness functions (FFNN and RNN) are noted in Table 7.
4.1. Experimental setup

2.2.4. Module 4: classification The experiments for the analysis were carried on a computer with an
After obtaining the optimal feature-subset, the selected features are Intel(R) Core i5 CPU and 6 GB of memory, using MATLAB 2014a. The
used as the input to the classification process. For the validation, a clas- phase space dynamics of normal and alcoholic EEG signals were plotted
sifier with the 10-fold cross-validation technique is adopted to measure in the first stage of the experimental setup, and 34 graphical features
the error rate. This study applies 11 common classifiers including neural were extracted afterward. As there were 120 trials for the normal
network (NN), multilayer neural network (MNN), recurrent neural net- group, and the same numbers for the alcoholic group, the features vec-
work (RNN), generalized regression neural network (GRNN), feed- tor consisted of a matrix of 120 × 34 dimensions for each group. As a re-
forward neural network (FFNN), cascade forward neural network sult, the feature vector for both normal and alcoholic classes has a size of
(CFNN), random forest (RF), linear discriminate analysis (LDA), k- 240 × 34. To reduce the size of the feature vector, the efficacy of 13 FS
nearest neighbor (KNN) with Euclidean distance, KNN with city block algorithms (ASO, BDA, BDE, HHO, TGA, BEO, BGA, GWO, HGS, PSO,
distance, and support vector machine (SVM). The detailed parameter SCA, SSA, and WOA) with 11 different classifiers as fitness function
settings of classifiers are tabulated in Table 4. (NN, MNN, CFNN, FFNN, GRNN, RNN, SVM, KNN with City block and
Euclidean distances, LDA and RF) have been experimented. For reducing
2.2.5. Alcoholic diagnostic indexes the bias effect, the results for classifications (i.e. ACC, SEN, and SPE)
Another key contribution of this study is to develop two alcoholic di- were computed in a 10-fold cross-validation strategy. In the 10-fold
agnostic indexes (ADI1 and ADI2). It is challenging to track the variety of cross-validation technique, the feature vector is divided randomly into
selected features between the normal and alcoholic cases to make an ten equal subsets. After that, the classifier is trained by nine subsets
accurate diagnosis. Even though classifiers can be used for diagnosis, it and tested by the remaining one subset. This procedure is repeated
is more convenient for the clinicians' use if we present them with a ten times, so each subset is set once as the testing data, and nine times
single integrated index that provides significant differences between as the training data. Finally, the performance of the classifier is obtained
the two classes. Many indexes have been developed for categorizing by averaging the results during ten times of training and testing the

4
M.T. Sadiq, H. Akbari, S. Siuly et al. Chaos, Solitons and Fractals 158 (2022) 112036

Table 2
Description of graphical features.

No. Features Meaning Illustration

1 “Summation of consecutive circles The SCCA quantifies the variation of phase space dynamic of EEG Fig. 4(a) shows the SCCA, which is formulated by SCCA =
area (SCCA)” [26] signals on 2D space. n
∑ C i , here i = 3, and Ci represents the area of ith circle.
i¼1
2 “Summation of consecutive The SCTA quantifies the variation of phase space dynamic of EEG Fig. 4(b) shows the SCTA, which is formulated by SCTA =
triangles area (SCTA)” [27] signals on 2D space with more flexibility in contrast to the SCCA. n
∑ T i , here i = 5, and Ti represents the area of ith triangle.
i¼1
3 “Summation of Heron's circulars The SHCA quantifies the self-similarity of phase space dynamic of Fig. 4(c) shows the SHCA, which is formulated by SHCA =
area (SHCA)” [26] EEG signals on 2D space. n
∑ πr 2i , here i = 5 and ri represents the radius of ith circle.
i¼1
4 “Summation of distances between The SDHC quantifies the self-similarity and amount of complexity Fig. 4(d) shows the SDHC, which is formulated by SDHC =
Heron's circulars (SDHC)” [26,27] of phase space dynamic of EEG signals. n
∑ Di , here i = 4, and Di is the distance between successive
i¼1
Heron's circulars.
5 “Summation of the angles between The SAHC quantifies the similarity of 2D shape to each other which Fig. 4(e) shows the SAHC, which is formulated by SAHC =
Heron's circular (SAHC)” [27] is same as correlation. n
∑ Ai , here i = 3, and Ai is angle between the successive
i¼1
Heron's circular centers.
6 “Summation of successive vector The SSVL quantifies the amplitude variation of EEG signals in the Fig. 4(f) shows the SSVL, which is formulated by SSVL =
lengths (SSVL)” [26] time domain. n
∑ Li , here i = 6, and Li is the length of each vector made by
i¼1
successive points.
7 “Shortest distance from each point The SH45 quantifies the data scatter rate in the second and fourth Fig. 4(g) shows the SH45, which is formulated by SH45 =
relative to the 45-degree line quarters which is define as the width of a 2D shape from the n
∑ SH i , here i = 7, and SHi is the distance from each point to
(SH45)” [26,27] 45-degree line. i¼1
45 degree line.
8 “Summation of shortest distance The SH135 quantifies the data scatter rate in the first and third Fig. 4(h) shows the SH135, which is formulated by SH135 =
from each point relative to the quarters, which is define as the width of a 2D shape from the n
∑ SH j , here j = 7, and SHj is the distance from each point to
135-degree line (SH135)” [26] 135-degree line. j¼1
135 degree line.
9 “Area of octagon (AOCT)” [26] AOCT measures the extent of data expansion of 2D shape on the Fig. 4(i) shows the AOCT
coordinate plane
10 “Summation of distances to a The SDTD quantifies the variation of 2D shape from the coordinate Fig. 4(j) shows the SDTC, which is formulated by SDTC =
coordinate center (SDTC)” [26] center. n
∑ Cr i , here i = 7, and Cri is the distance from each point to
i¼1
coordinate center.
11 “Summation of the angles between The SABP measures the variability of angles which quantifies the Fig. 4(k) shows the SABP, which is formulated as SABP =
three consecutive points (SABP)” amount of complexity of EEG signal in the time domain. n
∑ Aj , here i = 5, and Aj is the angle between the successive
[26,27] j¼1
vectors made by successive points.
12 “Summation of triangles area made The TACR combines the SSVL and SDTD, and quantifies the Fig. 4(l) shows the TACR, which is formulated by TACR =
successive points and coordinate variation and self-similarity of 2D shape simultaneously. n
∑ Tr i , here i = 6, and Tri is the area of each triangle made by
center (TACR)” [26] i¼1
two successive points and coordinate center.
13 “Summation of concessive The SCRA combines the SH45, SH135, and SDTC graphical features. Fig. 4(m) shows the SCRA, which is formulated by SCRA =
rectangular area (SCRA)” [26] Specifically, SCRA quantifies the scattering of data from the n
∑ Ri , here i = 4, and Ri is the area of each rectangular made
coordinate center, 45 and 135-degree lines simultaneously with i¼1
more sensitivity. by each points of 45 and 135 degree lines.
14 “Two-dimensional standard The TDSD quantifies the scattering of data points on the coordinate Fig. 4(n) shows the TDSD as a used graphical feature where
descriptors (TDSD)” [27] plane. STD1 and STD2 represents 45 and 135 degree lines.
15 “Elliptical area (ELPA)” [28] The shape of the phase space dynamic of EEG signals has an Fig. 4(o) shows the ELPA, where a and b are the two
elliptical pattern which represents the scattering of data. diameters of elliptical
16 “Central tendency measure (CTM)” The CTM quantifies the degree of variability of points on the Fig. 4(p) shows the CTMs as graphical features, where CTM90
to [26,29] coordinate plane by measuring the percentage of data scatter on represents the outer (larger) and CTM10 represents the inner
34 the 2D space. We employed the range of CTM to 5, 10, …, 90, 95, (smallest) circle respectively.
and computed the corresponding r. (i.e. 19 variants of CTM are
extracted such as CTM5, CTM10, …CTM90, CT95)

classifier. Let the number of selected features be γ, then during 10-fold means that the 2D shape of phase space dynamics of normal EEG signals
cross-validating, the size of the training matrix is 216 × γ and the size class scattered in more areas than the alcoholic group in the coordinate
of the test matrix is 24 × γ. plane (see Fig. 3). Besides, the std. values for all of the proposed
graphical features (expect of SHCA and TACR) in the normal group are
4.2. Statistical analysis higher than those in the alcoholic group, which indicates that the
alcoholic EEG signals are more similar among themselves compared to
To see the discrimination abilities of normal and alcoholic EEG fea- the normal EEG signals. Based on the observation above, it can be con-
tures, statistical analysis are performed. Table 5 shows the resulted cluded that alcoholic EEG signals have more similarities than differ-
mean and standard deviation (std) values for the extracted graphical ences, with less variations in contrast to the normal EEG signals.
features. It can be witnessed from Table 5 that the mean values for all Table 5 also highlights the fact that the p-value for all graphical features
of the proposed graphical features (except for SHCA, SAHC, and SABP) is below 0.001, which suggested the statistical significance of recom-
in the normal class are more in contrast to the alcoholic class, which mended features.

5
M.T. Sadiq, H. Akbari, S. Siuly et al. Chaos, Solitons and Fractals 158 (2022) 112036

Fig. 4. Illustration of extracted graphical features.

4.3. Significance of graphical features for alcoholic EEG signal detection

In this study graphical features are first time utilized for correct iden-
Table 3
Parameter settings utilized in the FS methods. tification of normal vs. alcoholic EEG signals classification. The signifi-
cance of graphical features in alcoholic EEG signal detection is detailed
Method Parameter Value
subsequently. In Table 5, we found that SHCA and SAHC features
All Population size 10 (which measure the similarities and correlations of EEG signals respec-
Maximum iterations 30 tively) have higher mean value in an alcoholic group than in the normal
K-fold cross-validation 10
ASO Depth weight (α) 50
group that indicates that EEG signals in the alcoholic group are more
Multiplier weight (β) 0.2 predictable than the normal group. A similar conclusion was drawn in
BDA Dmax 6 the study [2], wherein the mean value of centered correntropy was
BDE CR 0.9
HHO Number of hawks 10
TGA θ 0.8 Table 4
λ 0.5 Parameter settings for the classifiers used.
Size of the first group 3
Method Parameter Value
Size of the second group 5
Size of the fourth group 3 All K-fold cross-validation 10
BEO al 2 NN, MNN, FFNN, CFNN, and Hidden layer size 10
a2 1 RNN Maximum epochs allowed 50
GP 0.5 GRNN The spread of radial basis 1
BGA CR 0.8 functions
MR 0.01 RF Number of trees 50
GWO Number of wolves 10 LDA Function Linear
PSO c1 and c2 2 SVM Kernel function Radial basis
w [0.9,0.4] function
HGS Number of gas types 3 Sigma 1
SCA α 2 KNN Distance metric Euclidean, city--
SSA Number of salps 10 block
WOA Number of whales 10 k-value 5

6
M.T. Sadiq, H. Akbari, S. Siuly et al. Chaos, Solitons and Fractals 158 (2022) 112036

Table 5
Mean and standard deviation of extracted graphical features.

Feature Mean ± std of normal Mean ± std of alcoholic Feature Mean ± std of normal Mean ± std of alcoholic

SCCA 232,509 ± 136,459 168,605 ± 123,621 CTM15 30 ± 17 9±7

SCTA 17,544 ± 11,495 13,233 ± 9924 CTM20 48 ± 27 15 ± 11
SHCA 5715 ± 1625 8983 ± 2418 CTM25 71 ± 40 21 ± 15
SDHC 16,670 ± 4212 10,814 ± 3346 CTM30 101 ± 55 30 ± 21
SAHC 29,847 ± 7490 34,218 ± 7325 CTM35 136 ± 71 40 ± 27
SSVL 74,010 ± 43,436 53,669 ± 39,350 CTM40 177 ± 87 52 ± 35
SH45 2873 ± 722 1877 ± 569 CTM45 229 ± 107 67 ± 43
SH135 23,536 ± 3764 12,773 ± 3473 CTM50 290 ± 129 84 ± 53
AOCT 156,029 ± 44,106 53,383 ± 24,110 CTM55 362 ± 154 106 ± 65
SDTC 24,157 ± 3790 13,229 ± 3529 CTM60 450 ± 182 132 ± 80
SABP 31,939 ± 8122 37,033 ± 7462 CTM65 560 ± 217 165 ± 97
TACR 9797 ± 823 7238 ± 1016 CTM70 695 ± 255 206 ± 117
SRCA 42,830 ± 13,991 21,444 ± 10,651 CTM75 865 ± 304 260 ± 144
TDSD 124 ± 47 59 ± 30 CTM80 1094 ± 370 332 ± 181
ELPA 8.35E+08 ± 4.71E+08 1.09E+08 ± 9.16E+07 CTM85 1425 ± 467 440 ± 237
CTM05 7±4 2±2 CTM90 1949 ± 622 615 ± 330
CTM10 17 ± 9 5±4 CTM95 3022 ± 969 972 ± 551

found to be higher in the alcoholic group than in the normal group. Fur- p-value (as shown in Table 5). However, this is not suitable for real-
thermore, the mean values of the SSVL, SDTC, and CTMs (which mea- time computer-aided diagnosis framework. For this reason, several FS
sure the variability of EEG signals) are substantially higher in the methods (such as ASO, BDA, HHO, TGA, HGS, SSA, SCA, WOA, BEO,
normal EEG class compare to the alcoholic EEG class, indicating that BDE, BGA, GWO, and PSO as discussed in Section 2.2.3) has been
normal EEG class have more diversified and individualized morpholog- investigated to eradicate the redundant features. In this work, the ACC
ical characteristics than alcoholic EEG class. Similar findings were seen of classifiers in terms of the 10-fold CV strategy is used as the fitness
in a study [5], in which statistical features revealed that the alcoholic function in all FS methods. To identify the normal or alcoholic features
group had lower mean values than that from the normal group. correctly, several classification methods (including NN, MNN, CFNN,
The SH45, SH135, SRCA, and ELPA measures the scattering of points FFNN, GRNN, RNN, SVM, KNN with City block, and Euclidean distances,
for the phase space dynamic shape, and the value of frequency compo- LDA and RF as described in Section 2.2.4) are utilized for experimenta-
nents in the EEG signal have a direct relationship with the scattering of tion.
the phase space dynamic shape. In the current work, the mean values of The resulted ACC with various classifiers is shown in Table 6. For nor-
these aforementioned features in the normal EEG class are considerably mal vs. alcoholic signals, highest classification ACC of 99.16% is achieved
more than the alcoholic EEG class, which indicates the increasing num- by employing HGS with the FFNN. On contrary, lowest classification ACC
ber of frequency components of normal group EEG signals in contrast to of 91.16% is achieved by employing BDA and BDE FS methods with
the alcoholic group. The same consequence has been reported in [3], GRNN classifier, which is 7.5% lesser than the highest results. It is also
wherein area under power spectrum is utilized as a feature and normal understood from Table 6 that the performance of FS methods is in-
EEG signals showed more values compared to alcoholic EEG signals. creased with FFNN and RNN fitness functions, indicating these two clas-
sifiers are better than all others in the normal vs. alcoholic EEG signals
4.4. Results obtained with computerized framework classification task. Similar results have been found in studies [47,50,
51], where the FFNN classifier outperformed the other classifiers, indi-
This section reports the experimental results obtained with the pro- cating its dominance for correctly recognizing EEG signals in contrast
posed framework. The proposed framework is built after investigating to other classifiers.
different combinations of feature selection methods with classifiers Table 7 reports the feature vector size, chosen attributes, and respec-
used. A total of 143 combinations were examined from 13 FS methods tive classification accuracies obtained for FS approaches with FFNN and
and 11 classifiers. The key reason we employed different FS methods RNN fitness functions. It is clear from Table 7 that the SCA and PSO algo-
with classification models is to provide a comprehensive comparison rithms with FFNN and RNN fitness functions are resulted in 98.33% and
among several combinations, and to suggest an optimum combination 98.75% of classification accuracies by choosing only 10 and 11 feature
methodology. This is the first of such kind of study to investigate alco- vectors, respectively. Although these accuracies are slightly less than
holism. the 99.16% (i.e. highest classification accuracy obtained with 23 features
For classification of normal vs. alcoholic EEG signals, all 34 graphical obtained by the HGS FS method with FFNN fitness function), the num-
feature can be used as an input to the classifier because of their lowest ber of the selected features is significantly less. It infers that the SCA

Table 6
The ACC (%) from the selected features by classifiers and FS methods.

Classifier ASO BDA BDE HHO TGA BEO BGA GWO HGS PSO SCA SSA WOA

NN 97.5 97.5 97.5 97.5 97.5 97.5 97.5 97.08 97.08 97.91 97.5 97.08 97.08
MNN 97.5 97.5 98.33 97.91 97.08 97.08 97.08 97.08 97.08 94.16 93.75 95 93.33
CFNN 97.5 98.33 97.91 97.91 97.91 98.75 97.91 97.91 98.33 98.33 98.33 97.91 98.33
FFNN 98.75 98.75 98.75 98.33 98.33 98.33 98.33 98.75 99.16 98.75 98.33 98.33 98.33
GRNN 95 91.66 91.66 92.5 92.5 93.33 95.41 92.5 92.08 92.08 95.41 93.33 92.5
RNN 98.75 98.75 98.33 98.33 98.33 98.33 98.33 98.75 98.75 98.75 98.33 98.33 98.75
SVM 95.41 95.83 94.58 95 95.41 94.58 95.41 95.41 95.83 96.25 95.83 95.41 95
KNN (City block) 96.25 96.25 94.58 95 94.16 96.25 95 94.58 95.83 96.25 97.08 95.41 95.41
KNN (Euclidean) 96.25 95.41 94.16 96.25 96.25 96.25 94.16 95 95 95.83 95 95 95.41
LDA 95.83 94.58 94.58 93.75 93.33 94.58 95 94.16 94.58 94.16 93.75 95 94.16
RF 95.83 96.25 96.25 96.66 95.83 96.66 95.83 95.83 95.83 96.66 96.66 96.25 96.25

7
M.T. Sadiq, H. Akbari, S. Siuly et al. Chaos, Solitons and Fractals 158 (2022) 112036

Table 7
Feature vector lengths and selected features by FS algorithms.

Method Selected features by FFNN fitness function No of ACC

arrays (%)

ASO “SCCA, AOCT, SRCA, CTM25, CTM50, CTM60, CTM75, CTM80, CTM85, CTM90, CTM95” 11 98.75
BDA “SCCA, SCTA, SHCA, SDHC, SAHC, SH45, SDTC, SABP, TACR, SRCA, TDSD, ELPA, CTM20, CTM35, CTM40, CTM45” 23 98.75
BDE “SCCA, SDHC, SH45, SH135, AOCT, SDTC, SABP, TACR, SRCA, TDSD, ELPA, CTM5, CTM10, CTM20, CTM25, CTM45” 23 98.75
HHO “SDHC, SH45, SDTC, SABP, TDSD, ELPA, CTM10, CTM15, CTM30, CTM40, CTM45, CTM55, CTM60, CTM65, CTM80, CTM85, CTM90” 17 98.33
TGA “SCCA, SCTA, SHCA, SDHC, SAHC, SSVL, SH135, AOCT, TACR, SRCA, TDSD, ELPA, CTM15, CTM45, CTM50, CTM55, CTM65, CTM70, CTM75, CTM80, 23 98.33
CTM85, CTM90, CTM95”
BEO “SCCA, SCTA, SHCA, SDHC, AOCT, TACR, SRCA, ELPA, CTM5, CTM10, CTM20, CTM25, CTM30, CTM50, CTM55, CTM60, CTM75, CTM80, CTM85, CTM95” 20 98.33
BGA “SCCA, SCTA, SDHC, SAHC, SSVL, SH45, SDTC, SABP, CTM15, CTM25, CTM30, CTM35, CTM40, CTM45, CTM55, CTM65, CTM70, CTM75, CTM85, CTM90, 21 98.33
CTM95”
GWO “SCCA, SCTA, SHCA, SDHC, SAHC, SSVL, SH45, SH135, AOCT, SABP, TACR, SRCA, TDSD, ELPA, CTM20, CTM35, CTM40, CTM45, CTM50, CTM70, CTM85, 23 98.75
CTM90, CTM95”
HGS “SHCA, SAHC, SSVL, SH45, SH135, SDTC, SABP, TACR, SRCA, ELPA, CTM5, CTM10, CTM15, CTM25, CTM35, CTM40, CTM45, CTM60, CTM65, CTM75, 23 99.16
CTM85, CTM90, CTM95”
PSO “SDHC, SSVL, SH45, SH135, TACR, TDSD, ELPA, CTM5, CTM10, CTM20, CTM40, CTM55, CTM60, CTM80, CTM90” 15 98.75
SCA “SSVL, SH135, AOCT, TACR, ELPA, CTM5, CTM10, CTM45, CTM50, CTM75” 10 98.33
SSA “SCTA, SHCA, SAHC, SSVL, TDSD, ELPA, CTM20, CTM35, CTM50, CTM70, CTM90” 12 98.33
WOA “SCCA, SCTA, SHCA, SDHC, SSVL, AOCT, SDTC, TACR, SRCA, ELPA, CTM10, CTM15, CTM25, CTM30, CTM35, CTM40, CTM50, CTM60, CTM75, CTM80, 23 98.33
CTM85, CTM90, CTM95”

Method Selected features by RNN fitness function No of ACC

arrays (%)

ASO “SCCA, SCTA, SHCA, SAHC, SH45, SH135, SABP, TACR, ELPA, CTM5, CTM10, CTM15, CTM20, CTM30, CTM35, CTM50, CTM55, CTM60, CTM65, 23 98.75
CTM75, CTM80, CTM90, CTM95”
BDA “SCCA, SCTA, SDHC, SH45, AOCT, SDTC, SABP, TACR, SRCA, TDSD, ELPA, CTM10, CTM20, CTM30, CTM50, CTM60, CTM65, CTM70, CTM80, CTM90” 20 98.75
BDE “SCCA, SCTA, SHCA, SAHC, SSVL, AOCT, SABP, TACR, SRCA, ELPA, CTM5, CTM10, CTM15, CTM20, CTM25, CTM30, CTM35, CTM40, CTM45, CTM50, 24 98.33
CTM55, CTM60, CTM85, CTM95”
HHO “SHCA, SAHC, SH45, SH135, ELPA, CTM10, CTM20, CTM30, CTM35, CTM50, CTM55, CTM65, CTM75, CTM80, CTM90, CTM95” 16 98.33
TGA “SHCA, SAHC, SH45, SH135, ELPA, CTM10, CTM20, CTM30, CTM35, CTM50, CTM55, CTM65, CTM75, CTM80, CTM90, CTM95” 16 98.33
BEO “SHCA, SAHC, SSVL, SH45, SH135, SDTC, ELPA, CTM10, CTM15, CTM30, CTM45, CTM55, CTM65, CTM75, CTM80, CTM95” 16 98.33
BGA “SHCA, SAHC, SH45, SH135, ELPA, CTM10, CTM20, CTM30, CTM35, CTM50, CTM55, CTM65, CTM75, CTM80, CTM90, CTM95” 16 98.33
GWO “SDHC, SAHC, SH45, SH135, AOCT, SDTC, SABP, ELPA, CTM10, CTM20, CTM30, CTM35, CTM60, CTM75, CTM80, CTM95” 16 98.75
HGS “SCCA, SCTA, SHCA, SDHC, SAHC, SSVL, SH45, SH135, SDTC, TACR, TDSD, ELPA, CTM5, CTM10, CTM15, CTM20, CTM35, CTM45, CTM50, CTM55, 26 98.75
CTM60, CTM65, CTM75, CTM80”
PSO “SAHC, SH135, SABP, TDSD, ELPA, CTM10, CTM20, CTM40, CTM55, CTM75, CTM95” 11 98.75
SCA “SHCA, SAHC, SH45, SH135, ELPA, CTM10, CTM20, CTM30, CTM35, CTM50, CTM55, CTM65, CTM75, CTM80, CTM90, CTM95” 16 98.33
SSA “SHCA, SAHC, SH45, SH135, ELPA, CTM10, CTM20, CTM30, CTM35, CTM50, CTM55, CTM65, CTM75, CTM80, CTM90, CTM95” 16 98.33
WOA “SCCA, SH135, SDTC, CTM5, CTM15, CTM20, CTM25, CTM45, CTM50, CTM65, CTM70, CTM80, CTM85, CTM95” 14 98.75

Fig. 5. Comparing the sensitivity (%) for the selected features by various FS and classification methods.

8
M.T. Sadiq, H. Akbari, S. Siuly et al. Chaos, Solitons and Fractals 158 (2022) 112036

Fig. 6. Comparing the specificity (%) for the selected features by various FS and classification methods.

and PSO FS approaches are better than others in reduction of feature framework is shown in Fig. 8. We have provided phase space dynamic,
vector sizes while providing high ACC values. feature extraction, and classification test time. As seen in Fig. 8, the
The SEN and SPE results of different combinations are shown in phase space dynamic and graphical feature extraction time is only
Figs. 5 and 6 respectively. The HGS with FFNN provides 100% and 0.295 s, which is way less than signal decomposition or non-linear fea-
98.36% SEN and SPE results, which indicate that there is just a 1.64% pos- ture extraction techniques. Besides, the computer-aided diagnosis
sibility that a signal with a normal label incorrectly predicted as alco- framework requires only test time of a classifier i.e. 0.03 s and 0.034 s
holic class. Furthermore, it is noted that neural networks provide for FFNN and RNN, respectively, by utilizing Matlab software, signifying
relatively higher results for both SEN and SPE in comparison with tradi- the aptness of the proposed framework for real-world systems.
tional machine learning classifiers. Specifically, there are up to 10.84%
and 8.1% improvements in SEN and SPE achieved by HGS + FFNN or
WOA + RNN and HHO + FFNN respectively in comparison with all
other combinations.
Fig. 7 shows the AUC for all the FS methods by employing the best
fitness functions (FFNN and RNN). It is clear from the bar graphs, for
all the FS methods both FFNN and RNN provide 0.98 or more value, in-
dicating the authenticity of both fitness functions in the identification of
normal and alcoholic EEG signals. The time complexity of the proposed

Fig. 7. Area under the receiver operating curve for computerized framework. Fig. 8. Time complexity for computerized framework.

9
M.T. Sadiq, H. Akbari, S. Siuly et al. Chaos, Solitons and Fractals 158 (2022) 112036

Fig. 9. Scatter plot of most discriminative features.

4.5. Results obtained with alcoholic diagnostic indexes alcoholic subjects by using only a single number and assist medical
professionals.
Fig. 9 shows the scatter plots of features employed for the develop-
ment of ADI1 and ADI2. It is clear from Fig. 9 there is a significant 5. Discussion
difference among the two-class features. Table 8 represents the ADI1
and ADI2 values (mean ± std) for normal and alcoholic classes. It is 1. The complexity and chaotic behavior of EEG signals were analyzed by
obvious from Table 8 that the values of ADI indexes for the two classes the phase space dynamic in the proposed frame for normal and alco-
are distinctly different from each other. As seen in Fig. 10 the range of holic classes in this paper. In comparison to the alcoholic group, the
ADI1 and ADI2 for the two classes provide a distinct separation 2D shape of EEG signals in the normal group covers bigger and
between the two classes. So, it is evidence that the proposed two wider area, with more scattering patterns from the coordinate center
indexes can provide clear discrimination among the normal and and bisector of trigonometric regions, indicating its suitability as a vi-
sual biomarker of alcoholic diagnosis for physicians, and help neurol-
Table 8 ogists to investigate the effects of alcohol on the brain. The smaller
The mean ± std and p-value for the ADIs. mean values of alcoholic EEG signal features noted in Table 5 indicate
that the alcoholic EEG signal has more common morphological char-
Indexes Normal Alcoholic p-Value
acteristics compared to the normal group.
ADI1 69.71 ± 9.84 40.07 ± 9.50 4.18E − 39 2. In exploring the effective FS with classification combination in the
ADI2 32.10 ± 8.25 13.47 ± 6.02 1.11E − 36
recommended framework, the proficiency of the 13 FS methods

Fig. 10. Ranges of ADI indexes for the normal and alcoholic classes: (a) ADI1 and (b) ADI2.

10
M.T. Sadiq, H. Akbari, S. Siuly et al. Chaos, Solitons and Fractals 158 (2022) 112036

Table 9
Comparing the proposed framework with the exiting works.

Ref. Author, year Method, feature FS CV Classifier ACC (%) SEN (%) SPE (%)

[8] Zhong and Ghosh, 2002 HMMs and coupled HMMs Not used 10-fold NN 82.98 – –
[9] Acharya, Sree et al., 2012 LLE, ApEn, SamEn and HOS features p-Value 3-fold SVM 91.7 90 93.33
[10] Faust, Yu et al., 2013 WPT, HOS p-Value 10-fold KNN 95.8 95.8 95.8
[13] Upadhyay, Padhy, et al., 2014 CWT, statistical features Not used 10-fold SVM 94.29 – –
[14] Patidar, Pachori et al., 2017 TQWT, center correntropy PCA 10-fold LS-SVM 97.02 96.53 97.5
[11] Bae, Yoo et al., 2017 Granger causality Not used 5-fold SVM 90 95.3 82.4
[15] Sharma, Sharma et al., 2018 DTCWT, two entropies p-Value 10-fold SVM 97.91 – –
[16] Sharma, Deb et al., 2018 TBOWFB, log-energy entropy p-Value 10-fold LS-SVM 97.08 97.08 97.08
[12] Mumtaz, Kamel et al., 2018 Synchronization likelihood ROC 10-fold SVM 98 99.9 95
[18] Thilagaraj and Rajasekaran, 2019 EMD, power band and fractal dimension ICA 10-fold KNN 98.91 99.02 99.24
[19] Anuragi, Sisodia et al., 2020 FBSE-EWT, line-length, log energy entropy and norm entropy Not used Leave one out LS-SVM 98.8 98.3 99.1
[20] Anuragi and Sisodia, 2020 EWT, statistical features p-Value Leave one out LS-SVM 98.75 98.35 99.16
[22] Farsi, Siuly et al., 2020 Deep learning Not used 10-fold LSTM 93 95 92
Our work Sadiq et al., 2021 Phase space dynamic, graphical features HGS 10-fold FFNN 99.16 100 98.36

with 11 classifiers was checked and HGS + FFNN achieved the structures of normal and alcoholic EEG signals in a more simpler man-
highest classification ACC of 99.16%, SEN of 100%, and SPE of 98.36% ner; (ii) the investigation of various FS with different classification
by selecting 23 features. On the other hand, the SCA + FFNN and methods for alcoholic EEG signal detection, wherein, the HGS + FFNN
PSO + RNN, with the least number of feature vector arrays (i.e. 11 selects the best feature vectors, the RNN and FFNN are the best classi-
and 10) achieved classification ACC of 98.75% and 98.3% accordingly. fiers and the SCA + FFNN and PSO + RNN are the best combinations
3. In Table 9, the proposed framework is compared with state-of-the- that significantly reduce the feature vector length while keeping a
art works that were experimented with the same database. The com- high accuracy; (iii) providing new biomarkers and novel diagnostic in-
parison is made in terms of ACC, SEN, and SPE, wherein, the available dexes which can act as clinical tools for neurologists.
literature feature extraction methods, feature selection strategies, Consequentially, the key advantages of this study to help psychia-
CV, and classifiers are also reported in Table 9 for better understand- trists for alcoholic subject detection in the following three ways, Firstly,
ing. It is clear from Table 9 that almost all the available systems are visual-based diagnostic (see Fig. 3). Secondly, index-based diagnosis (as
made with the help of signal decomposition with nonlinear features, shown in Fig. 10 and Table 8). Thirdly, computerized framework. The
which inherent several limitations such as unsuitable frequency proposed framework can be used as a computer application for clinics
order, mode mixing, narrowly spaced frequencies mixing as well as and hospitals as a fast and inexpensive technique for alcoholic subject
heavy calculations. detection.

The proposed framework achieved the best classification ACC, SEN, 6. Conclusion
and SPE values of 99.16%, 100%, and 98.36% respectively. In comparison
to all other techniques discussed in Table 9, phase space dynamics with Alcoholism has become one of the common problems in the world. If
graphical features provide up to 16.18%, 10%, and 15.96% improvements left untreated, it can lead to cardiovascular diseases, depression, self-
in ACC, SEN, and SPE with straightforward and simplest manner. harm, or suicide. Although qualified psychiatrists can diagnose alcohol-
4. In the proposed framework, the computational algorithm time for ism by medical imaging and counseling, yet, computerized methods are
each EEG with 2048 samples for extraction of 34 graphical feature needed to avoid human errors. In the present study, a new computer-
vectors form phase space dynamic using i5-M480 CPU (2.67 GHz), ized framework has been developed based on the phase space dynamic
6GB RAM, and MATLAB 2014a is around 0.5 s, which demonstrates of EEG signals for the detection of alcoholism subjects. Specifically, the
its effectivity and simplicity of the proposed method. Also, our pro- significant graphical features are extracted from the phase space dy-
posed method required less than 0.1 s for the classification of an namic of EEG signals, then selected and fed to several classifiers. The
input test signal, required for real-time applications. proposed framework achieved a classification ACC of 99.16%, SEN of
5. In Section 2.2.5, we suggested two new indexes ADI1 and ADI2 for 100%, and SPE of 98.36% in the 10-fold CV strategy. Besides, two new in-
diagnostic indicators for alcoholic disorders with a single number. dexes have been proposed for medical professional to detect the alco-
As seen in Fig. 10 and Table 8, both the proposed indexes can be holic subject with a single number indication. Another significance of
used as an alcoholism diagnostic tool and employed as a reference the proposed study is to introduce a visual biomarker for psychiatrists
for the neurologist to measure the amount of brain damage in an and neurologists to diagnose alcoholics and analyze the intrinsic behav-
alcoholism subject. In the existing literature, we found only one ior of an alcoholic affected brain. The proposed framework has very fast
alcoholic diagnostic index [14], which is based on TQWT, center computations, which made it a suitable choice for real-time applica-
correntropy, and principal component analysis (PCA), wherein, the tions. In the future, the proposed framework will be applied to investi-
TQWT decomposed EEG signals into 4 levels, and the center gate other brain degenerative disorders, such as schizophrenia,
correntropy with seven lags was computed from each sub-band as autism, or depression, etc.
the features. At last, the PCA was applied to the extracted features
and its coefficients were used for the development of an index. The CRediT authorship contribution statement
centered correntropy needed intensive calculations and for seven
lags, the computations became even more complicated. However, Muhammad Tariq Sadiq: Conceptualization, Methodology, Soft-
our proposed indexes do not need any time-frequency decomposi- ware, Validation, Formal analysis, Investigation, Writing - original
tion method before feature extraction, and the features were ex- draft, Writing - review & editing, Visualization, Project administration.
tracted directly from the phase space dynamic of EEG signals Hesam Akbari: Conceptualization, Methodology, Software, Validation,
which represent the straightforward and fast calculations. Formal analysis, Investigation, Writing - original draft, Writing - review
& editing, Visualization. Siuly Siuly: Writing - review & editing, Visual-
The highlights of this study comprised of (i) utilization of the phase ization. Yan Li: Writing - review & editing, Visualization. Peng (Paul)
space dynamics and graphical features for discrimination of nonlinear Wen: Writing - review & editing, Visualization.

11
M.T. Sadiq, H. Akbari, S. Siuly et al. Chaos, Solitons and Fractals 158 (2022) 112036

Declaration of competing interest [26] Akbari H, Sadiq MT, Rehman AU, Ghazvini M, Naqvi RA, Payan M, Bagheri H, Bagheri
H. Depression recognition based on the reconstruction of phase space of eeg signals
and geometrical features. Appl Acoust. 2021;179:108078.
The authors declare that they have no conflict of interest. [27] Moridani MK, Setarehdan SK, Nasrabadi AM, Hajinasrollah E. A novel approach to
mortality prediction of icu cardiovascular patient based on fuzzy logic method.
Biomed Signal Process Control. 2018;45:160–73.
References
[28] Akbari H, Sadiq MT, Payan M, Esmaili SS, Baghri H, Bagheri H. Depression detection
based on geometrical features extracted from sodp shape of eeg signals and binary
[1] Organization WH. Global status report on alcohol and health 2018. World Health
pso. Traitement du Signal. 2021.;38(1).
Organization; 2018.
[2] Prince J. Substance use disorder and suicide attempt among people who report com- [29] Bhattacharyya A, Sharma M, Pachori RB, Sircar P, Acharya UR. A novel approach for
promised health. Subst Use Misuse. 2018;53(1):9–15. automated detection of focal eeg signals using empirical wavelet transform. Neural
[3] Arunkumar N, Kumar KR, Venkataraman V. Entropy features for focal eeg and non Comput Appl. 2018;29(8):47–57.
focal eeg. J Comput Sci. 2018;27:440–4. [30] Faris H, Heidari AA, Ala’M A-Z, Mafarja M, Aljarah I, Eshtay M, Mirjalili S. Time-
[4] Siuly S, Li Y, Zhang Y. Eeg signal analysis and classification. Health information sci- varying hierarchical chains of salps with random weight networks for feature selec-
ence. Springer; 2016 978-3-319-47653-7. tion. Expert Systems with Applications. 2020;140:112898.
[5] Khan DM, Yahya N, Kamel N, Faye I. Effective connectivity in default mode network [31] Hammouri AI, Mafarja M, Al-Betar MA, Awadallah MA, Abu-Doush I. An improved
for alcoholism diagnosis. IEEE Trans Neural Syst Rehabil Eng. 2021;29:796–808. dragonfly algorithm for feature selection. Knowledge-Based Syst. 2020;203:106131.
[6] Zhu G, Li Y, Wen PP, Wang S. Analysis of alcoholic eeg signals based on horizontal [32] de Souza RCT, de Macedo CA, dos Santos Coelho L, Pierezan J, Mariani VC. Binary coy-
visibility graph entropy. Brain Inform. 2014;1(1–4):19–25. ote optimization algorithm for feature selection. Pattern Recogn. 2020;107:107470.
[7] Zhu G, Li Y, Wen P. Evaluating functional connectivity in alcoholics based on maxi- [33] Nguyen BH, Xue B, Zhang M. A survey on swarm intelligence approaches to feature
mal weight matching. J Adv Comput Intell Intell Inform. 2011;15(9):1221–7. selection in data mining. Swarm and Evolutionary Computation, 54. ; 2020.
[8] Zhong S, Ghosh J. Hmms and coupled hmms for multi-channel eeg classification. [34] Too J, Rahim Abdullah A. Binary atom search optimisation approaches for feature se-
Proceedings of the 2002 International Joint Conference on Neural Networks. lection. Connect Sci. 2020;32(4):406–30.
IJCNN’02 (Cat. No. 02CH37290). IEEE; 2002. p. 1154–9. [35] Mirjalili S. Dragonfly algorithm: a new meta-heuristic optimization technique for
[9] Acharya UR, Sree SV, Chattopadhyay S, Suri JS. Automated diagnosis of normal and solving single-objective, discrete, and multi-objective problems. Neural Comput
alcoholic eeg signals. Int J Neural Syst. 2012;22(03):1250011. Applic. 2016;27(4):1053–73.
[10] Faust O, Yu W, Kadri NA. Computer-based identification of normal and alcoholic eeg [36] Zorarpac E, Özel SA. A hybrid approach of differential evolution and artificial bee col-
signals using wavelet packets and energy measures. J Mech Med Biol. 2013;13(03): ony for feature selection. Expert Systems with Applications, 62. ; 2016. p. 91–103.
1350033. [37] Too J, Abdullah AR, Mohd Saad N. A new quadratic binary Harris hawk optimization
[11] Bae Y, Yoo BW, Lee JC, Kim HC. Automated network analysis to measure brain effec- for feature selection. Electronics. 2019;8(10):1130.
tive connectivity estimated from eeg data of patients with alcoholism. Physiol Meas. [38] Too J, Abdullah AR, Mohd Saad N, Mohd Ali N. Feature selection based on binary tree
2017;38(5):759. growth algorithm for the classification of myoelectric signals. Machines. 2018;6(4):
[12] Mumtaz W, Kamel N, Ali SSA, Malik AS, et al. An eeg-based functional connectivity 65.
measure for automatic detection of alcohol use disorder. Artif Intell Med. 2018;84: [39] Faramarzi A, Heidarinejad M, Stephens B, Mirjalili S. Equilibrium optimizer: a novel
79–89. optimization algorithm. Knowledge-Based Syst. 2020;191:105190.
[13] Upadhyay R, Padhy P, Kankar P. Alcoholism diagnosis from eeg signals using contin- [40] Holland JH. Genetic algorithms. Sci Am. 1992;267(1):66–73.
uous wavelet transform. 2014 annual IEEE India conference (INDICON). IEEE; 2014. [41] Emary E, Zawbaa HM, Hassanien AE. Binary grey wolf optimization approaches for
p. 1–5. feature selection. Neurocomputing. 2016;172:371–81.
[14] Patidar S, Pachori RB, Upadhyay A, Acharya UR. An integrated alcoholic index using [42] Kennedy J, Eberhart RC. A discrete binary version of the particle swarm algorithm.
tunable-q wavelet transform based features extracted from eeg signals for diagnosis 1997 IEEE International conference on systems, man, and cybernetics. Computa-
of alcoholism. Appl Soft Comput. 2017;50:71–8.
tional cybernetics and simulation. IEEE; 1997. p. 4104–8.
[15] Sharma M, Sharma P, Pachori RB, Acharya UR. Dual-tree complex wavelet
[43] Hashim FA, Houssein EH, Mabrouk MS, Al-Atabany W, Mirjalili S. Henry gas solubil-
transform-based features for automated alcoholism identification. Int J Fuzzy Syst.
ity optimization: a novel physics-based algorithm. Futur Gener Comput Syst. 2019;
2018;20(4):1297–308.
101:646–67.
[16] Sharma M, Deb D, Acharya UR. A novel three-band orthogonal wavelet filter bank
[44] Mirjalili S. Sca: a sine cosine algorithm for solving optimization problems.
method for an automated identification of alcoholic eeg signals. Appl Intell. 2018;
Knowledge-based Syst. 2016;96:120–33.
48(5):1368–78.
[45] Dhiman G, Kumar V. Spotted hyena optimizer: a novel bio-inspired based
[17] Anuragi A, Sisodia DS. Alcohol use disorder detection using eeg signal features and
metaheuristic technique for engineering applications. Adv Eng Softw. 2017;114:
flexible analytical wavelet transform. Biomed Signal Process Control. 2019;52:
48–70.
384–93.
[18] Thilagaraj M, Rajasekaran MP. An empirical mode decomposition (emd)-based [46] Mirjalili S, Lewis A. The whale optimization algorithm. Adv Eng Softw. 2016;95:
scheme for alcoholism identification. Pattern Recogn Lett. 2019;125:133–9. 51–67.
[19] Anuragi A, Sisodia DS, Pachori RB. Automated alcoholism detection using fourier- [47] Sadiq MT, Yu X, Yuan Z, Aziz MZ. Identification of motor and mental imagery eeg in
bessel series expansion based empirical wavelet transform. IEEE Sens J. 2020;20 two and multiclass subject-dependent tasks using successive decomposition index.
(9):4914–24. Sensors. 2020;20(18):5283.
[20] Anuragi A, Sisodia DS. Empirical wavelet transform based automated alcoholism de- [48] Raghavendra U, Koh JEW, Gudigar A, Chan WY, Hamid MTR, Rahmat K, Fadzli F, Ng
tecting using eeg signal features. Biomed Signal Process Control. 2020;57:101777. KH, Ooi CP, Ciaccio EJ, et al. Development of breast papillary index for differentiation
[21] Mehla VK, Singhal A, Singh P. A novel approach for automated alcoholism detection of benign and malignant lesions using ultrasound images. J Ambient Intell Humaniz
using fourier decomposition method. J Neurosci Methods. 2020;346:108945. Comput. 2021;12(2):2121–9.
[22] Farsi L, Siuly S, Kabir E, Wang H. Classification of alcoholic eeg signals using a deep [49] Acharya UR, Ng EY-K, Tan J-H, Sree SV, Ng K-H. An integrated index for the identifi-
learning method. IEEE Sens J. 2020;21(3):3552–60. cation of diabetic retinopathy stages using texture parameters. J Med Syst. 2012;36
[23] Acharya JN, Hani AJ, Cheek J, Thirumala P, Tsuchida TN. American clinical neurophys- (3):2011–20.
iology society guideline 2: guidelines for standard electrode position nomenclature. [50] Sadiq MT, Yu X, Yuan Z, Aziz MZ. Motor imagery bci classification based on novel
Neurodiagn J. 2016;56(4):245–52. two-dimensional modelling in empirical wavelet transform. Electr Lett. 2020;56
[24] Snodgrass JG, Vanderwart M. A standardized set of 260 pictures: norms for name (25):1367–9.
agreement, image agreement, familiarity, and visual complexity. J Exp [51] Sadiq MT, Yu X, Yuan Z, Aziz MZ, Siuly S, Ding W. A matrix determinant feature ex-
PsycholHuman Learn Memory. 1980;6(2):174. traction approach for decoding motor and mental imagery eeg in subject specific
[25] Zabihi M, Kiranyaz S, Rad AB, Katsaggelos AK, Gabbouj M, Ince T. Analysis of high- tasks. IEEE Trans Cogn Dev Syst. 2020. https://doi.org/10.1109/TCDS.2020.
dimensional phase space via poincaré section for patient-specific seizure detection. 3040438. In press.
IEEE Trans Neural Syst Rehabil Eng. 2016;24(3):386–98.

12