Computers in Biology and Medicine 138 (2021) 104922

Contents lists available at ScienceDirect

Computers in Biology and Medicine

journal homepage: www.elsevier.com/locate/compbiomed

A novel computer-aided diagnosis framework for EEG-based identification

of neural diseases
Muhammad Tariq Sadiq a, b, *, 1, Hesam Akbari c, 1, Siuly Siuly d, Adnan Yousaf e,
Ateeq Ur Rehman f
a
School of Automation, Northwestern Polytechnical University, Xi’an, 710072, China
b
Department of Electrical Engineering, The University of Lahore, Lahore, 54000, Pakistan
c
Department of Biomedical Engineering, Islamic Azad University, Tehran, 1411718541, Iran
d
Institute for Sustainable Industries & Liveable Cities, Victoria University, Melbourne, 14428, Australia
e
Department of Electrical Engineering, Superior University, Lahore, 54000, Pakistan
f
Department of Electrical Engineering, Government College University, Lahore, 54000, Pakistan

A R T I C L E I N F O A B S T R A C T

Keywords: Recent advances in electroencephalogram (EEG) signal classification have primarily focused on domain-specific
Electroencephalography approaches, which impede algorithm cross-discipline capability. This study introduces a new computer-aided
Neural diseases diagnosis (CAD) system for the classification of two distinct EEG domains under a unified sequential frame­
Two-dimensional modeling
work. The key motivation to consider two neural diseases by one framework is to develop a unified algorithm for
Geometrical features
Computer-aided diagnosis
EEG classification. The main contributions of this study are five-fold. First, EEG signals are decomposed into 10
intrinsic mode functions (IMFs) with the help of empirical wavelet transform. Second, a novel two-dimensional
(2D) modeling of IMFs is plotted to visualize the complexity of EEG signals. Third, several new geometrical
features are extracted to analyze the dynamic and chaotic essence. Fourth, significant features are selected by
binary particle swarm optimization algorithm (B–PSO). Fifth, selected features are fed to the k-nearest neighbor
classifier for EEG signal classification purposes. All the experiments are executed on one depression and two
epileptic EEG datasets in a leave one out cross-validation strategy. The proposed CAD system provides an average
classification accuracy of 93.35% in depression detection, 99.33% for regular against ictal, and 97.33% for
interictal versus ictal respectively. The overall empirical analysis authenticates that the proposed CAD out­
performs the existing domain-specific methods in terms of classification accuracies and multirole adaptability,
thus, can be endorsed as an effective automated neural rehabilitation system.

1. Introduction progression of the disease. On the contrary, epilepsy is factored as a

dreadful brain disorder originate by an abnormal dissent of the brain.
1.1. Background Detection of seizure can be useful for the treatment of epilepsy patients.
The seizure does not have explicit clinical signs in most epilepsy patients
Depression and epilepsy are two prevalent brain disorders in the [3–5].
world. According to the world health organization report, there are Electroencephalography (EEG) signals are commonly used for
around 264 and 50 million people of all ages suffering from the afore­ detecting brain activities and disorders. Depression EEG signals are
mentioned illness accordingly [1,2]. In recent years, depression disor­ more predictable (stationary behavior) with lesser morphological
ders have constantly been raised due to the mechanization and complexity than the normal or regular EEG signals, which are non-
modernization of life, and the decline in relations among society. predictable (non-stationary behavior) with more morphological
Depression is a treatable disease and its early diagnosis can avoid the complexity. However, epileptic seizures in the human brain display

* Corresponding author. School of Automation, Northwestern Polytechnical University, Xi’an, 710072, China.
E-mail addresses: [email protected] (M.T. Sadiq), [email protected] (H. Akbari), [email protected] (S. Siuly), adnan.
[email protected] (A. Yousaf), [email protected] (A.U. Rehman).
1
Co-first authors.

https://doi.org/10.1016/j.compbiomed.2021.104922
Received 27 January 2021; Received in revised form 3 October 2021; Accepted 5 October 2021
Available online 12 October 2021
0010-4825/© 2021 Elsevier Ltd. All rights reserved.
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

increases in EEG signals due to increased synapse activity, that inevi­ and statistical features [25,26], nonlinear features (Lyapunov expo­
tably leads to ictal parts in EEG signals. The manual follow-up of lengthy nent, fractal demotions, and entropy-based) [27–34],
EEG records is a cumbersome and time taking activity for detecting correlation-based features [35–38], and graphical features (CTM and
neural disorders. Therefore, an automatic method would be required for elliptical) [22–24,39]. However, linear features are incapable to
the detection of depression and epilepsy illness [4,6]. detect non-stationary characteristics of depression and epilepsy EEG
signals; chaotic-based and entropy-based features require high pro­
1.2. Literature cessing time due to complex calculations; correlation-based features
are time-consuming and sensitive to noise; and CTM-based features
To build an automated EEG-based system various techniques are do not provide information about the instantaneous variations of the
employed which are categorized as, time-domain, frequency-domain, 2D shape.
time-frequency-based, and geometrical approaches. In literature, 3. Almost all the existing frameworks are restricted to a specific EEG
various time-domain features are extracted, including permutation en­ domain.
tropy [7], horizontal visibility graph [8], linear prediction error energy
[9], fractional linear prediction error [10], histogram [11], and statis­ 1.4. Motivation and contributions
tical parameters [12–14]. The EEG signals evaluated by those extracted
features are without considering the frequency components. To deal The key innovation of this work is to provide a new two-dimensional
with this problem, the Fourier transform (FT) is applied to the EEG (2D) illustration of IMFs and the introduction of new graphical features.
signals to extract features from the EEG signals spectrum, resulting in The motivation and novelties are listed as follows.
the generation of the frequency domain. Different features have been
extracted in the frequency-domain, consisting of entropies [11], 1. The motivation of our proposed 2D modeling is taken to resolve
mean-frequency, and root-mean-square bandwidth [15]. These spectral limitations in already available 2D approaches. To overcome the
features are important for stationary signals, but EEG recordings are limitations encountered by RPS and SODP methods in this study, we
extremely non-stationary. Besides, FT relies on sine and cosine functions developed a novel 2D method for illustration of chaotic behavior of
which may not have the ability to explain the EEG signals spectrum IMFs of EEG signals in depression and epilepsy problems. The pro­
evidently. posed method is fast in comparison with RPS and SODP because it
In recent studies, traditional wavelet transform-based methods have does not depend on any parameters, and is developed by combining
been proposed which are taken into account as multi-resolution analysis the number of angles concerning the IMFs of EEG signals. Further­
approaches, and provide time-frequency planes for the representation of more, the proposed 2D study analyzes the instantaneous variations of
the EEG signal components. These methods are better than the fre­ 2D structures rather than exhibiting fluctuations among different
quency domain methods, whome are based on a single resolution EEG signals.
analysis. The most popular wavelet-based transforms for the decompo­ 2. Another key novelty of this work is to introduce new graphical fea­
sition of EEG signals are comprised of time-scale decomposition (ITD) tures that are not only suitable for EEG depression detection but are
[16], discrete wavelet transform (DWT) [16], dual-tree complex wavelet equally effective for EEG signals taken from distinct epilepsy data­
transforms (DT-CWT) [17], tunable-Q wavelet transforms (TQWT) [18], sets. To address the aforementioned limitations found in existing
and flexible analytic wavelet transform (FAWT) [19]. The main defi­ features, we proposed three graphical features. These graphical
ciency of wavelet transform-based methods is that the filter bank is not features can solve the limitations of linear, chaotic, entropy, and
adapted to input signals [17,18]. An empirical mode decomposition correlation-based features by quantifying the chaotic and complex
(EMD) method has been proposed to address this problem, which is an behaviors of EEG signals in the nonlinear domain with less
adaptive algorithm that decomposes the nonlinear and non-stationary complexity and time. Besides, the value of these features has a direct
signals into respective modes or intrinsic mode functions (IMFs). But relationship with the instantaneous changes on 2D shape points.
the EMD deals with some shortcomings such as mode mixing, noise
sensitivity, heavy calculation, and inability to estimate the modes before In the present study, we developed a novel CAD system for the
decomposing. To overcome the issues of EMD, the empirical wavelet identification of depression and epileptic tasks in one place. Unlike the
transform (EWT) is proposed which designs an adaptive filter bank ac­ existing 2D approaches that represents only variations of EEG signal
cording to the existing main frequency components in the input signal shapes with heavy calculations, the proposed method considers the
spectrum and decomposes it into modes [20,21]. instantaneous variations of 2D shape and categorizes EEG signals with
Besides, the signal decomposition-based approaches, few geomet­ higher classification accuracy considering simple calculations. The rec­
rical approaches for EEG signals such as, second-order difference plot ommended framework comprises of a number of tasks such as: (1)
(SODP) [22] and reconstructed phase space (RPS) [23] are used to implement a signal decomposition approach to obtain intrinsic mode
provide a two-dimensional (2D) representation for the modes of EEG functions (IMF) or modes; (2) introduce a novel scheme for two-
signals. The RPS demands the two variables to be estimated, time delay dimensional modeling of modes; (3) extract several features; (4) select
and embedding factor. Using mutual information (MI) and false nearest suitable feature vectors; (5) examine appropriate classification model
neighbor (FNN), together, these two variables are determined [23,24]. for recognition of epileptic and depressed patients from normal subjects.
In this study, firstly we employed an empirical wavelet transform
1.3. Limitations (EWT) method to acquire modes from nonlinear and nonstationary EEG
signals. Secondly, we proposed a novel two-dimensional modeling
The gaps found in aforementioned discussed literature, is listed as scheme to visualize the chaotic behavior of depression and epileptic
follows. signals. Thirdly, we extract geometrical representative characteristics of
2D dynamics named as summation of distances from each point to co­
1. The large computational complexity and processing time due to MI ordinate center (SDC), summation of rectangles area (SRA) and sum­
and FNN embedded parameters is the major drawback of RPS mation of triangle area (STA). Fourthly, binary particle swarm
method [23]. On the other hand, the SODP illustrates the variability optimization (B–PSO) technique was implemented to reduce the feature
of the signal but does not clarify the complex nature of the signal for vectors. At last, we investigate different values of K in K-nearest
each instant [22]. neighbours (KNN) classifier with the relevant features for classification
2. In previous studies, various categories of features are extracted from of depression and epilepsy patients from normal subjects. In the litera­
depression and epilepsy EEG signals, which are categorized as, linear ture, several machine learning and neural network methods have been

2
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

proposed for EEG signal classification. Among those methods, the KNN Table 1
algorithm has a special place that is because of its light calculations. The Information of volunteers with number of trials in normal and depression
KNN is a fast classifier with a simple theory that causes easy imple­ groups.
mentation in MATLAB. For this reason, in the current work, the KNN NG DG NA DA NT DT
algorithm is used as a classifier for detecting depression and seizure EEG M (S1) M (S23) 51 58 77 88
signals. According to the best of the author’s knowledge, this is the first M (S2) M (S24) 18 29 79 86
attempt towards the development of such a CAD system that provides M (S3) M (S25) 20 31 83 65
efficient identification of neural disorders of distinct EEG domains. M (S4) M (S26) 23 34 64 86
M (S5) M (S27) 33 44 86 77
M (S6) M (S28) 20 31 82 86
1.5. Organization M (S7) M (S29) 33 44 75 73
M (S8) M (S30) 19 28 76 88
M (S9) M (S31) 23 34 69 81
This paper has been arranged as following. Section 2 is focused on
M (S10) M (S32) 46 57 86 62
materials description. In section 3, the proposed methodology is pre­ M (S11) W (S33) 51 59 83 65
sented, which is comprised of EWT, 2D modeling of dynamics of EEG M (S12) W (S34) 58 56 88 78
signals, graphical features, PSO, and classifiers. Section 4 describe the M (S13) W (S35) 27 25 87 79
performance evaluation metrics. Sections 5 and 6 comprise the results M (S14) W (S36) 31 29 85 79
M (S15) W (S37) 35 33 68 86
and discussion respectively. At last, the conclusion is presented for the M (S16) W (S38) 51 49 63 80
proposed framework in section 7. W (S17) W (S39) 34 32 88 78
W (S18) W (S40) 25 23 77 73
2. Materials W (S19) W (S41) 48 46 67 68
W (S20) W (S42) 45 43 89 87
W (S21) W (S43) 44 42 68 66
In the present analysis, three databases are employed to determine W (S22) W (S44) 25 23 60 69
the effectiveness of the planned approach for detecting depression and
S# = Subject Number, M = Men, W=Women, NG=Normal Gender, DG =
epilepsy EEG recordings (one self-collected depression dataset and two
Depressed Gender, NA=Normal Age, DA = Depressed Age, NT=Normal Trials,
publicly accessible epilepsy data sets).
DT = Depressed Trials.

2.1. Dataset 1: depression database

Table 2
In this study, we utilized the same depression dataset as we utilized Statistical analysis of gender, age and trial numbers of volunteers for normal and
in our previous studies [35,40,40], which is taken from AJA University depression groups in depression database.
of Medical Sciences. EEG recordings were taken from 22 healthy people Variable Normal Depression p- DOF Effect
and 22 depressed people to diagnose depression. Since previous research value size
[4,29,41–43] found that bipolar channels ”FP1-T3” and ”FP2-T4” from Gender M = 16, W = M = 10, W = 0.07 42 0.08
the left and right halves of the brain, respectively, performed well in 6 12
diagnosing depression, the EEG recordings in this study were also ac­ Age (mean ± 34.54 ± 38.63 ± 11.72 0.43 42 0.02
std) 12.60
quired by bipolar channels at a sampling frequency of 256 HZ for 10 min Trial numbers 1700 1700 0.98 42 1.27E-5
with a bipolar montage. The position of bipolar channels on the left and
right halves of brain is depicted in Fig. 1. Table 1 represents gender, age, M = Men, W=Women, DOF = Degree of Freedom.
and the number of trials for each subject in normal and depression
groups, whereas, Table 2 demonstrates the demographics information respectively as follows,
for normal and depression groups. To find the number of values in a DF = N − 1 (1)
calculation that has the freedom to vary, the degree of freedom is uti­
lized. On the other hand, to measure the strength of the relationship where DF represents the degree of freedom and N denotes the sample
among different genders, the effect size is also estimated. The mathe­ size.
matical formulation for the degree of freedom and effect size are given
SSeffect ​
η2p = (2)
SSeffect ​ + SSerror ​

where η2p represents the effect size of choice, SS denotes effect and error
sums of squares. For the partial eta square effect size calculation
method, 0.01, 0.06, and 0.14 represent small, medium, and large effects
accordingly.
In this database, the performance of the proposed method is tested
for classifying normal and depressed EEG signals. This experiment was
approved by the Research Ethics Committee of AJA University of Med­
ical Sciences (Approval ID: IR.AJAUMS.REC.1399.049), Tehran, Iran
[35,40,40]. The eye-blinking artifact of recorded data was discarded
visually by a signal recording expert and approved by two psychiatrists
and a neurologist. Also, a Butterworth low-pass filter with a cut-off
frequency of 80 Hz was applied to EEG signals as preprocessing. Fig. 2
shows a visual representation of the normal and depressed EEG signals
in the left and right half of the brain, respectively. In the current study,
the normal against depressed classification task is performed for the EEG
signals. All recorded EEG signal trials have 2000 samples and in total, we
Fig. 1. Location of bipolar channels on left and right halves for depres­ have 1700 trials for normal and 1700 trials for the depression group.
sion recording.

3
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

Fig. 2. A sample of normal and depression EEG signals for the left and right half of the brain.

2.2. Dataset 2: University of Bonn epilepsy database states. Subgroups C and D were documented from five patients whom
have recovered fully from seizure control at epileptic areas following
The second database is a benchmark dataset, which is freely down­ operation. Subgroup E is comprised of EEG signals being used track the
loadable from the university of Bonn website [44]. This database activity of epileptic seizures. In this research, the signals in the A and B
comprised of five subgroups, namely A, B, C, D, and E. Every subset subgroups are assumed to be regular EEG signals, while the C and D
contains 100 EEG signals, which were acquired at a sampling frequency subgroups are believed to be interictal EEG signals, while the signals in
of 173.61 Hz. The duration of every EEG signal is 23.6 s, so the number the E subclass are known to be ictal EEG messages.
of samples are 4097. Subgroups A and B were obtained from five normal
persons, respectively, taking into account eye-open and eye-closed

Fig. 3. A sample of ictal, interictal and normal EEG signals.

4
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

2.3. Dataset 3: Indian epilepsy database are determined as B = {B1, B2, …, B9}. It is done by assuming that the
first midpoint is between 0 and L1. Finally the EWT boundaries are
The third database is another epilepsy database, which is collected defined as:
by “Neurology and Sleep Centre, Hauz Khas, New Delhi” [45–47]. We {
ω = [ 0 ​ B ​ 1 ]1 , [ ​ B ​ 1 ​ B2 ​ ]2 , [ ​ B ​ 2 ​ B3} ​ ]3 ,
named this as an Indian database, which consists of three subsets, spe­ (3)
………, [ ​ B ​ 8 ​ B ​ 9 ]9 , [ ​ B ​ 9 ​ fs ​ /2 ]10
cifically normal, interictal, and ictal. Every subset contains 50 instances
and each instance consists of 1024 samples. Indian database was ac­ The EWT filter bank is created based on the most recent Lit­
quired at a sampling frequency of 400 Hz, and the duration of each tlewood–Paley, and Meyer wavelets after differentiation of a range of
signal is 2.56 s. In the current study, the tasks are described as normal or frequencies. In the Fourier domain, scaling function and wavelet func­
regular against ictal, and interictal versus ictal EEG signal classification tion are defined for the EWT filter bank as follows [49]:
tasks for evaluating the proposed method. Fig. 3 depicts a plot for ictal, ⎧
interictal, and normal EEG signals. ⎪
⎪ 1, if |ωf | ≤ (1 − λ)ω1
⎪ (
⎪ )
⎨
For both Bonn and Indian databases, the EEG signals were recorded π β(λ, ω1 )
φ(ωf ) = cos , if (1 − λ)ω1 ≤ |ωf | ≤ (1 + λ)ω1 (4)
by the gold pated electrodes using international standard 10–20 system. ⎪
⎪ 2
⎪
⎪
⎩
0, otherwise
3. Methods
⎧
⎪
⎪ 1, if (1 + λ)ωi
In this study, the EWT is used to decompose the EEG signals into 10 ⎪
⎪
⎪
⎪
⎪
modes. Then, 2D modeling of modes is plotted, and significant graphical ⎪
⎪
⎪
⎪
features are extracted based on their shapes and incorporated with the ⎪
⎪
⎪
⎪ ≤ |ωf | ≤ (1 − λ)ωi+1
⎪
PSO algorithm. Thus, the feature vector arrays are reduced and are fed ⎪
⎪
⎪
(
π β(λ, ωi+1 )
)
⎪
into the KNN classifier in the ten-fold cross-validation technique. Fig. 4 ⎪
⎨ cos
⎪ , if (1 − λ)ωi+1 ≤ |ωf |
2
exhibits a step by step procedure for the proposed framework. The de­ ψ i=2,…,m (ωf ) = (5)
⎪ ≤ (1 + λ)ωi+1
tails of the proposed system are explained in several steps that are ⎪
⎪
⎪
⎪
⎪ ( )
subsequently presented. ⎪
⎪
⎪ πβ(λ, ωi )
⎪
⎪ sin , if (1 − λ)ωi ≤ |ωf |
⎪
⎪ 2
⎪
⎪
⎪
⎪
⎪ ≤ (1 + λ)ωi
3.1. Step 1: empirical wavelet transform ⎪
⎪
⎪
⎪
⎩ 0, otherwise
For the analysis of nonlinear and non-stationary EEG signals,
( )
empirical wavelet transform (EWT) is utilized in the present study. The
where β(λ, ωi ) = β , ωf is the bandwidth of the EWT filter bank
|ωf |− (1− λ)
EWT overcome the mode-mixing problem encountered in the EMD and 2λωf

unlike discrete wavelet transform (DWT), the bandwidth of filters in [49]. Also, β(y) is introduced as follows:
EWT changes according to the input signal spectrum to extract the main ⎧
frequency components. In the EWT algorithm, the operator must ⎨ 0 ​ if ​ y ≤ 0
β(y) = β(y) + β(1 − y) = 1 ∀y ∈ [0, 1] (6)
determine the number of IMF’s. To extract 10 empirical IMF’s, the fre­ ⎩
1 ​ if ​ y ≥ 1
quency spectrum of the input signal is obtained by fast Fourier transform
( )
(FFT) algorithm with 09 maximums, which are considered as main
frequency components, and detected from the frequency spectrum as Moreover, λ < min ωωi+1 − ωi
i− 1 +ωi
maintains that the EWT coefficients are
[48]: L = {L1, L2, …, L9}. Then, the midpoints between local maximums

Fig. 4. Proposed system for identification of depression and epilepsy EEG signals.

5
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

in L2 (R) space. The factor of λ tights the filter bank structure and makes ( x )( ) ( )
Sn+1 − Sxn Sxn+2 − Sxn+1 + Syn+1 − Syn (Syn+2 − Syn+1 )
the lowest overlapping of bandwidths between the lower and upper B(n) = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̅ √ ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( x )2 ( )2 ( x )2
frequencies. The factor of λ creates ignorable stop-band ripples for EWT Sn+1 − Sxn + Syn+1 − Syn + Sn+2 − Sxn+1 + (Syn+2 − Syn+1 )
2

filter bank with the ability to solve mode-mixing problems. The inner
(8)
product of the input signal with scaling function and wavelet function
give IMF1, IMF2, …., IMF10, respectively. Fig. 5 and Fig. 6 show the We calculated the differences of S2(n) to quantify the variability of
IMF’s of depression and epilepsy EEG signals accordingly. the input pattern as follow:
( ) [ ]
z(n) = diff s2 (n) = s2n+1 − s2n (9)
3.2. Step 2: Proposed 2D modeling of dynamics of EEG signals
Finally, we can plot input pattern s(n) in 2D space by plotting the X
In the second step, 2D modeling of IMFs is proposed. Irrespective of (n) versus Y(n) as follow:
RPS and SODP approaches, the proposed 2D modeling is developed
2Z(n) = [X(n), Y(n)] (10)
considering angles of IMFs of EEG signals. The key advantages of the
proposed strategy are its simplicity and quickness as it does not require where X(n) and Y(n) are evaluated by angels B(n) and distances of Z(n)
any embedded parameters, and it provides instantaneous changes in 2D as follow:
structures. To define the proposed 2D modeling, it is assumed that s(n) is
a signal with n samples, which are characterized by s(n) = {s1 , s2 , …, sn }. X(n) = Z(n) × cos(B(n)) (11)
{ }
The second power of the s(n) is calculated as s2 (n) = s21 , s22 , …, s2n .
2 Y(n) = Z(n) × sin(B(n)) (12)
Then, the mean value is calculated for s (n). Now we can illustrate a 2D
model for s2(n) by assuming that S(n) is [50]: Fig. 7 and Fig. 8 show the 2D modeling of IMF’s of depression and
[ ] [ ] epilepsy EEG signals accordingly.
S(n) = s2 (n) − mean, (− mean)n = Sxn , Syn (7)

where Sxn and Syn represent the x and y dimensions of the nth place of S(n) 3.3. Step 3: Proposed graphical features
accordingly. Afterward, a sequence (B(n)) can be defined for relevant
angles among the three consecutive places in S(n) as follow [50]: To identify the chaotic behavior of anomalies in EEG signals, several
graphical features [51] are extracted from 2D modeling of IMF’s, which
are outlined in the following sub-sections.

Fig. 5. IMFs of (a) left depression (b) right depression (c) left normal (d) right normal EEG signals of the brain halves respectively obtained with the EWT.

6
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

3.3.1. Summation of distances from each point to coordinate center

Distances of points from coordinate can be used as a graphical
feature to quantify the scattering of data in 2D space. In this regard, the
summation of distances is calculated from each point to the coordinate
center (SDC) which is defined as follows:
∑n √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
SDC = X(i)2 + Y(i)2 (13)
i=1

where n represents the overall amount of the 2D plot of IMF’s. Also, X(i)
and Y(i) specify the point coordinates on the 2D plane.

3.3.2. Summation of rectangles area

To quantify the width of the 2D plot of IMFs in EEG signal, the
shortest distance is computed from each point to the 45 and 135-degree
lines as follow:
|X(i) − Y(i)|
D45 (i) = √̅̅̅ (14)
2

|X(i) + Y(i)|
D135 (i) = √̅̅̅ (15)
2

where D45(i) and D135(i) are the shortest distances from the ith point of
the 2D plot of IMFs in EEG signals to the 45 and 135-degree lines,
respectively. Also, X (i) and Y (i) specify the point coordinates on the 2D
plane. If it is assumed that D45(i) and D135(i) are two sides of a rectangle,
we can calculate the rectangle area for each point on a 2D plane. In other
words, the summation of rectangles area (SRA) created by 45 and 135-
degree lines is calculated as a feature. The SRA can be defined as follow:
∑
n
D45 (i) × D135 (i)
SRA = (16)
i=1
2

3.3.3. Summation of triangle area

Every three successive points make two vectors in 2D projection
which results in a triangle. Therefore, a 2D projection is dealt with
successive triangles of IMF in EEG signals. In this study, the summation
is computed for triangle areas (STA) created by three points as a
graphical feature. The STA measures the self-similarity between 2D
Space EEG data with high versatility unlike cross-correlation [10]. STA
is defined as follows:
⃒ ⎡ ⎤⃒
n− 2 ⃒ X(i) X(i + 1) X(i + 2) ⃒⃒
1∑ ⃒
STA = ⃒det ⎣ Y(i) Y(i + 1) Y(i + 2) ⎦ ⃒ (17)
2 i=1 ⃒⃒ ⃒
⃒
1 1 1

where n corresponds the overall amount of points on the 2D space, i, i +

1, and i + 2 of X(.) and Y(.) indicate the coordinates for the first, second,
and third successive points on the 2D projection plot.
Fig. 9 depicts the graphical features.

3.4. Step 4: feature selection by PSO algorithm

For the selection of significant features, we first normalize all the

features and then a binary edition of particle swarm optimization (PSO)
is utilized in the proposed system. The binary PSO is comprised of pbest
and gbest solutions, which update the velocity and position respectively.
Similar, to the PSO, pbest and gbest refer to the best personal and global
solutions, respectively. The velocity is updated for each solution as
follows [52]:

vdi (t + 1) = wvdi (t) + c1 r1 (pbestdi (t) − xdi (t))+

(18)
c2 r2 (gbestd (t) − xdi (t))

Fig. 6. IMFs of (a) ictal, (b) interictal and (c) normal EEG signals respectively where v represents the velocity, x expresses the position of the particle
obtained with the EWT. (solution), w denotes the inertia weight, c1 and c2 describe the acceler­
ation factors, r1 and r2 characterize two independent random numbers

7
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

Fig. 7. 2D modeling of IMFs of depression and normal EEG signals for left (a and b) and right (c and d) half of the brain, respectively.

8
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

Fig. 8. The 2D modeling of IMFs of ictal (a), interictal (b) and normal (c) EEG signals respectively.

considering an interval range of 0.1. Here, i represents the particle where x characterizes the solution, pbest and gbest represent the best
sequence in the population, d is the search area space, and t indicates the personal and global solutions over the whole populace, F(.) denotes the
amount of iterations. It should be noted that vmax and vmin refer to the function of fitness, and t describes the iterations amount.
maximum and minimum velocity, respectively [52,53]. Furthermore,
the conversion of velocity into probability value and updating in particle 3.5. Step 5: classification
position are formulated as:
( ) 1 The K-nearest neighbor (KNN) classifier is utilized to classify the
S vdi (t + 1) = (19) different domains extracted graphical features. The KNN is a supervised
1 + exp( − vdi (t + 1))
classifier with a easy theory and implementation [54]. The distance
{ ( )
1, ​ if ​ rand ​ < S vdi (t + 1) computation and the number of K are taken into account as two pa­
xdi (t + 1) = (20) rameters for the KNN classifier. In this study, Euclidean distance with
0, ​ otherwise ​
various numbers of K from 2 to 9 are the chosen parameters of the KNN
classifier.
where rand denotes a random number, which follows a uniform distri­
The K-nearest neighbor (KNN) classifier is utilized to classify the
bution function with a range of 0 and 1. In the present study, the
different extracted graphical features. The KNN is a supervised classifier
minimization function was applied, and ”pbest” and ”gbest” show a vital
with an easy theory and implementation [54]. The distance computation
performance in moving the particle into the optimal global. These are
and the number of K are taken into account as two parameters in KNN
updated by following expressions:
⎧ classifier. In this study, K is varied from 2 to 9 to choose the best value
⎨ xi (t + 1), for decision making. There are different metrics for computation of the
pbesti (t + 1) = ​ if ​ F(xi (t + 1)) < F(p ​ best ​ i (t)) (21) distance between n vectors such as Euclidean distance, city block, and
⎩
​ pbest ​ ti (t), ​ othervise ​ Minkowski metrics. The computation is performed through the
⎧ assumption that xi and yi are two vectors. Euclidean distance is calcu­
⎨ pbesti (t + 1), lated as [55]:
gbest ​ (t + 1) = ​ if ​ F(pbesti (t + 1)) < F(gbest ​ (t)) (22)
⎩
gbest ​ (t), ​ otherwise ​

9
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

right half of the brain EEG signals, it is assumed that the depressed and
normal groups are considered as abnormal and normal groups. In
normal against ictal, and interictal versus ictal classification tasks, it is
assumed that the normal and interictal groups are normal, and the ictal
group is an abnormal group.

5. Results

5.1. Statistical analysis

In this section we first explain the statistical significance of features,

subsequently, classification parameters and computational complexity
of the proposed framework are discussed. Table 3 and Table 4 present
the mean and standard deviation for the extracted graphical features.
According to these tables, the mean values of STA and SRA features for
IMF’s of normal EEG signals are greater than the IMF’s of depression
EEG signals, except for IMF9 and IMF10 in the left half of the brain, and
IMF10 in the right half of the brain. In both epilepsy databases, the mean
scores of the same attributes in ictal EEG recordings are considerably
higher than the IMF’s of regular and interictal EEG signals. It is also
noted in Tables 3 and 4, the other attribute is the standard deviation of
the features. In the depression group, the standard deviations of the
features are smaller than regular in both halves of the brain. In other
words, we can assume that the EEG signals for depression are analogous
to each other. Also, the standard deviations of the features in the ictal
Fig. 9. Illustration of Graphical Features (a) SDC (b) SRA, and (c) STA group are higher than interictal and normal groups. It indicates the
respectively. variability of ictal EEG signals in comparison with interictal and normal
EEG signals.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ n The probability (p) value shows the capability of the extracted fea­
d(x, y) = |xi − yi |2 (23) tures in a binary classification task. In the depression database, the
i=1 normal against depression EEG signals classification task is tested for the
City block metric is defined as [55]: recorded EEG signals from the left and right half of the brain. Two binary
classification tasks consist of regular against ictal, and interictal versus
∑ ictal EEG recordings, which are tested for both Bonn and Indian data­
n
d(x, y) = |xi − yi | (24)
i=1 bases. Indeed, the features with the p-value less than 0.05 are nominated
as a significant features in classification. In this study, the Kruskal-Wallis
Minkowski metric is estimated by Ref. [55]:
(KW) statistical analysis is employed to calculate the p value for classi­
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ n fication tasks. Table 5 indicates the achieved p values for these tasks.
|xi − yi |p (25) The p values are less than 0.05 for all features in three databases
p
d(x, y) =
i=1
excluding the STA and SDC in IMF6, and the SRA in IMF10 of depression
database. Table 5 showed p-tuning mistakes. For correct correlation, we
where n is the number of training data and p is an unfixed variable. If p
employed the Hommel procedure to obtained correct p values as noted
= 1, then the Minkowski and city block metric will be equivalent, and if
in Table 6.
p = 2, then the Minkowski metric and Euclidean distance will be
This shows the importance of the features used, however the number
equivalent.
of features are selected by binary PSO to reduce the computation of the
classifier.
4. Performance evaluation metrics

5.2. Classification outcome

In leave one out cross-validation strategy, accuracy (ACC), sensi­
tivity (SEN), and specificity (SPE) measures are computed with the KNN.
The binary PSO is used as a feature selection algorithm to reduce
In a binary classification task, the ACC measures the ability of the pro­
classifier complexity. In the present study, the fitness function is fixed
posed framework in a correct classification on normal and abnormal
for classification performance of the KNN in the ten-fold cross-validation
EEG signals. The ACC is introduced as follow [54]:
strategy. The acceleration factors (i.e., c1 and c2) are 1 and 2, vmax and
TP + TN vmin have a range between 6 and -6. The preliminary N-particle popu­
ACC = × 100 (26)
TP + TN + FP + FN lation is varied between 1 and 10.
Table 7 gives the selected features for each database. These features
the SEN and SPE measure the capability of the proposed framework for are fed into the KNN classifier with Euclidian distance. Table 8 gives the
detecting abnormal and normal EEG signals, respectively. These are attained classification ACC, SEN, and SPE for the classification tasks.
formulated as follow:
TP 5.2.1. Outcomes for depression database
SEN = × 100 (27) In the depression database, the average classification ACC of 93.85
TP + FN
± 1.2, SEN of 92.16 ± 0.9, and SPE of 95.71 ± 0.8 are reached by EEG
SPE =
TN
× 100 (28) signals in the left half of the brain, and average classification ACC of
TN + FP 94.75 ± 0.9, SEN of 91.83 ± 0.7, and SPE of 96.82 ± 0.8 are attained by
In the normal against depression classification tasks for the left and EEG signals in the right half of the brain. In order to evaluate the per­
formance of our proposed framework, subject-wise experiments are

10
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

Table 3
Mean and standard deviation (std) of features in depression database.
Left Depression Left Normal

SDC SRA STA SDC SRA STA

Mean Std Mean Std Mean Std Mean Std Mean Std Mean Std

IMF1 1.86E-03 2.87E-06 5.36E-06 6.43E-08 5.53E-06 5.64E-08 3.46E-03 3.17E-04 1.04E-03 3.17E-04 1.09E-03 3.19E-04
IMF2 7.05E-04 1.16E-06 3.92E-07 2.79E-09 4.17E-07 2.90E-09 2.21E-03 3.16E-04 1.00E-03 3.16E-04 1.00E-03 3.16E-04
IMF3 9.45E-04 1.62E-06 8.16E-07 8.70E-09 4.10E-07 3.46E-09 2.27E-03 3.16E-04 1.00E-03 3.16E-04 1.00E-03 3.16E-04
IMF4 2.70E-03 4.30E-06 1.27E-05 6.32E-08 4.41E-06 3.00E-08 4.10E-03 3.20E-04 1.12E-03 3.20E-04 1.08E-03 3.18E-04
IMF5 7.28E-03 1.32E-05 1.07E-04 6.21E-07 3.66E-05 2.05E-07 7.37E-03 3.39E-04 1.25E-03 3.14E-04 1.34E-03 3.34E-04
IMF6 1.33E-02 3.16E-05 2.14E-04 5.21E-06 6.44E-05 1.91E-06 9.90E-03 3.82E-04 7.32E-04 2.32E-04 6.76E-04 2.29E-04
IMF7 6.01E-03 1.22E-05 1.23E-04 8.83E-07 9.64E-06 7.94E-08 4.00E-03 3.30E-04 1.47E-03 3.77E-04 9.23E-04 2.58E-04
IMF8 9.56E-03 3.17E-05 1.03E-04 1.71E-06 1.15E-06 2.71E-08 4.42E-03 3.50E-04 5.72E-04 2.25E-04 5.03E-04 2.24E-04
IMF9 5.06E-03 3.13E-05 1.04E-03 2.47E-05 8.30E-05 3.41E-06 1.57E-03 6.49E-05 7.90E-05 2.64E-05 5.36E-04 2.24E-04
IMF10 4.18E-03 2.92E-05 7.75E-04 2.27E-05 1.42E-07 4.11E-09 2.08E-03 1.27E-04 1.23E-04 2.10E-05 1.77E-03 3.39E-04

Right Depression Right Normal

SDC SRA STA SDC SRA STA
Mean Std Mean Std Mean Std Mean Std Mean Std Mean Std

IMF1 4.04E-04 1.14E-06 2.88E-06 7.89E-08 3.02E-06 8.10E-08 1.55E-03 3.16E-04 1.00E-03 3.16E-04 1.00E-03 3.16E-04
IMF2 3.97E-03 8.09E-06 1.21E-04 2.31E-06 9.40E-05 1.90E-06 1.15E-02 3.32E-04 1.41E-03 3.24E-04 1.28E-03 3.21E-04
IMF3 5.52E-04 1.91E-06 3.08E-06 1.02E-07 2.40E-06 8.49E-08 1.79E-03 3.16E-04 1.00E-03 3.16E-04 1.00E-03 3.16E-04
IMF4 7.29E-04 2.81E-06 7.23E-06 2.73E-07 5.24E-06 2.14E-07 1.82E-03 3.16E-04 1.00E-03 3.16E-04 1.00E-03 3.16E-04
IMF5 1.95E-03 1.29E-05 2.25E-04 9.58E-06 1.76E-04 7.73E-06 2.43E-03 3.16E-04 1.00E-03 3.16E-04 1.00E-03 3.16E-04
IMF6 2.52E-03 8.48E-06 2.75E-04 9.08E-06 1.65E-04 6.60E-06 2.61E-03 3.17E-04 1.02E-03 3.16E-04 1.01E-03 3.16E-04
IMF7 2.77E-03 7.43E-06 6.97E-05 1.12E-06 7.31E-06 1.58E-07 2.32E-03 3.17E-04 1.16E-03 3.24E-04 1.10E-03 3.19E-04
IMF8 1.07E-03 6.80E-06 4.28E-05 1.16E-06 2.27E-06 7.59E-08 1.26E-03 3.16E-04 1.00E-03 3.16E-04 1.00E-03 3.16E-04
IMF9 1.29E-03 1.07E-05 5.78E-05 1.73E-06 1.92E-06 8.46E-08 1.21E-03 3.16E-04 1.00E-03 3.16E-04 1.00E-03 3.16E-04
IMF10 3.08E-03 2.64E-05 6.75E-04 2.25E-05 6.36E-07 1.87E-08 6.81E-04 3.45E-05 1.10E-05 1.80E-06 1.88E-03 3.55E-04

performed by employing the Leave-one-out cross-validation (LOO) 0.2, SEN of 95.19 ± 0.4, and SPE of 99.48 ± 0.3 are obtained for
method. As we have a total of 44 subjects, one of the 44 subjects in interictal vs. ictal EEG signal classification task. In the Indian database,
normal or depressed class was used as a testing set and the rest of the 43 average classification ACC of 96 ± 0.3, SEN of 96 ± 0.4, and SPE of 96 ±
subjects were used as a training set and the same procedure was 0.3 are given for normal vs. ictal EEG signal classification task and
repeated 44 times for both depression or normal subjects to ensure that average classification ACC of 99 ± 0.2, SEN of 99 ± 0.3, and SPE of 98 ±
each subject was used as test sample once. 0.4 are obtained for interictal vs. ictal EEG signal classification task.
Table 9 presents a subject-wise accuracy score for the proposed The proposed algorithm achieved average classification ACC of
framework. This table shows the experimental results of the LOO cross 99.33 ± 0.3 and 98.33 ± 0.2 in normal vs. ictal and intrictal vs. ictal EEG
validation method. For each of 44 normal or depression subjects, the signals classification task in leave one out (LOO) cross validation strat­
obtained accuracy scores (in percentage) for the testing set were pro­ egy for Bonn database which are very near to 100%. Although, there are
vided in Table .9. It is apparent from this table that 100% classification some works in literature, such as time–frequency representation using
accuracy is obtained by subjects 4, 7, 11, 13, 16, 19, 20 and 21 in IEVDHM–HT, which achieved 100% classification ACC [56], yet need
depression group. For normal category, 100% classification accuracy is heavy computations which limit its application for real-time applica­
deliver by subjects 6, 10, 12, 15, 17 and 22. In addition, for both normal tions. On contrary the proposed CAD system is less complex due to
and depression groups, most of the subjects provide around 90% clas­ simplicity of proposed 2D approach and graphical features. In addition,
sification results indicating the suitability of proposed framework for most of the available methods for seizure detection have been evaluated
subject-wise analysis. just by evaluating Bonn database while performance of our proposed
We have also reported feature-wise and subject-wise classification method has been evaluated with both Bonn and Indian databases,
outcomes in Fig. 10 and Fig. 11 respectively. Due to the different mental indicating that proposed method is more robust in comparison with
and physical nature of subjects distinct combinations are obtained. For others.
each subject, a combination is suggested after a trial and error manner. It
is observed that for most of the cases in left or right brain hemispheres, 5.2.3. Computational complexity of recommended framework
most of the subjects obtained more than 80% of classification accuracy The algorithm run times for depression, Bonn, and Indian databases
by using a single feature and IMF. These obtained results are much better were 0.7, 1.8, and 0.5 s respectively. These run times were related to the
than the 70% of the accuracy required to control a computer-aided processing of every EEG signal considering 2000, 4097, and 1024
diagnosis framework as mentioned in Ref. [49]. Table 10 and samples in depression, Bonn, and Indian datasets, respectively. The
Table 11 represent the best combination of single feature and IMF in processing steps were comprised of preprocessing, generation of EWT
terms of best classification performance results (%) for each subject left filter bank, decomposition of EEG signal to ten IMFs, 2D modeling, and
(L) and right hemispheres (R). It is noted from these tables that right computation of their STA, SDC, and SRA as graphical features. In this
hemispheres (R) yields better classification results for both classes. study, a computer system with specifications of i5-M480 CPU (2.67
We conclude that whether it’s a trials based, subject-wise or feature- GHz), 6 GB RAM were employed for implementing the process in
wise experiments, the proposed framework is suitable for classification MATLAB 2014. The small run time indicates the simplicity of the pro­
of normal and un-healthy subjects. posed method. Also, the proposed method requires less than 0.001 s for
classifying an input test signal in all classification tasks. This method
5.2.2. Outcomes for epilepsy databases contains a small number of arrays in the feature vector, leading to its fast
In the Bonn database, the average classification ACC of 99.33 ± 0.3, implementation. If the proposed method is developed by C++ codes, it
SEN of 99.00 ± 0.5, and SPE of 99.50 ± 0.4 are given for normal vs. ictal can be implement faster than the current situation.
EEG signal classification task and average classification ACC of 98.33 ±

11
 M.T. Sadiq et al.
Table 4
Mean and standard deviation (std) of features in epilepsy databases.
Bonn database

Normal Interictal Ictal

SDC SRA STA SDC SRA STA SDC SRA STA

Mean Std Mean Std Mean Std Mean Std Mean Std Mean Std Mean Std Mean Std Mean Std

IMF1 6.63E- 3.45E- 4.07E- 4.29E- 3.95E- 4.19E- 8.11E- 1.82E- 1.51E- 1.31E- 1.40E- 1.18E- 2.81E- 2.62E- 1.48E- 2.38E- 1.49E- 2.41E-
03 03 05 05 05 05 03 02 03 02 03 02 01 01 01 01 01 01
IMF2 2.59E- 3.15E- 1.52E- 5.75E- 1.47E- 5.03E- 1.60E- 6.08E- 2.19E- 2.43E- 1.41E- 1.41E- 1.12E- 1.84E- 4.09E- 1.26E- 4.19E- 1.27E-
03 03 05 05 05 05 03 03 05 04 05 04 01 01 02 01 02 01
IMF3 2.28E- 3.37E- 1.49E- 5.26E- 1.49E- 5.20E- 1.14E- 2.92E- 6.61E- 6.21E- 4.13E- 3.92E- 9.36E- 1.82E- 4.04E- 1.60E- 4.03E- 1.57E-
03 03 05 05 05 05 03 03 06 05 06 05 02 01 02 01 02 01
IMF4 3.50E- 7.41E- 4.52E- 1.87E- 4.50E- 1.85E- 1.06E- 2.30E- 4.02E- 2.66E- 1.61E- 7.23E- 8.25E- 1.41E- 2.50E- 1.08E- 2.56E- 1.08E-
03 03 05 04 05 04 03 03 06 05 06 06 02 01 02 01 02 01
IMF5 3.57E- 6.96E- 5.17E- 1.87E- 5.24E- 1.89E- 1.01E- 2.41E- 4.71E- 4.27E- 7.96E- 2.43E- 7.51E- 1.38E- 2.41E- 1.13E- 2.60E- 1.18E-
03 03 05 04 05 04 03 03 06 05 07 06 02 01 02 01 02 01
IMF6 3.46E- 6.03E- 4.52E- 1.36E- 4.70E- 1.42E- 8.98E- 1.70E- 2.98E- 1.69E- 8.51E- 3.23E- 9.37E- 1.65E- 3.43E- 1.24E- 3.70E- 1.32E-
03 03 05 04 05 04 04 03 06 05 07 06 02 01 02 01 02 01
IMF7 2.65E- 4.85E- 3.74E- 1.29E- 4.00E- 1.42E- 9.05E- 2.35E- 6.48E- 4.91E- 1.37E- 7.53E- 7.15E- 1.31E- 2.32E- 1.06E- 2.58E- 1.09E-
03 03 05 04 05 04 04 03 06 05 06 06 02 01 02 01 02 01
IMF8 1.46E- 3.21E- 1.05E- 4.90E- 1.11E- 5.04E- 4.87E- 1.31E- 1.51E- 8.15E- 1.06E- 7.78E- 5.15E- 1.19E- 1.44E- 1.01E- 1.52E- 1.02E-
03 03 05 05 05 05 04 03 06 06 06 06 02 01 02 01 02 01
IMF9 3.89E- 9.61E- 1.34E- 4.99E- 1.40E- 5.13E- 9.54E- 5.04E- 2.19E- 2.66E- 1.90E- 2.55E- 1.18E- 1.81E- 4.82E- 1.43E- 5.53E- 1.58E-
03 03 04 04 04 04 04 03 05 04 06 05 01 01 02 01 02 01
IMF10 1.05E- 3.99E- 1.49E- 8.58E- 1.61E- 9.94E- 2.55E- 7.34E- 5.26E- 2.44E- 4.72E- 4.09E- 2.23E- 1.39E- 1.85E- 1.30E- 1.99E- 1.40E-
12

03 03 05 05 05 05 05 05 09 08 12 11 02 01 02 01 02 01

Indian database
Normal Interictal Ictal
SDC SRA STA SDC SRA STA SDC SRA STA
Mean Std Mean Std Mean Std Mean Std Mean Std Mean Std Mean Std Mean Std Mean Std

IMF1 2.10E- 3.17E- 8.91E- 3.44E- 8.85E- 3.18E- 1.49E- 7.55E- 8.39E- 8.83E- 8.86E- 9.84E- 2.72E- 2.01E- 8.88E- 1.61E- 9.49E- 1.66E-
02 02 04 03 04 03 02 03 05 05 05 05 01 01 02 01 02 01
IMF2 1.24E- 1.46E- 2.99E- 6.65E- 3.62E- 8.35E- 1.10E- 1.20E- 1.65E- 4.20E- 1.74E- 4.03E- 1.97E- 2.36E- 8.84E- 2.37E- 8.83E- 2.27E-
02 02 04 04 04 04 02 02 04 04 04 04 01 01 02 01 02 01
IMF3 1.33E- 1.74E- 5.99E- 1.77E- 6.24E- 1.68E- 1.05E- 7.39E- 1.71E- 2.24E- 1.85E- 2.50E- 1.67E- 1.79E- 7.09E- 1.59E- 7.10E- 1.58E-
02 02 04 03 04 03 02 03 04 04 04 04 01 01 02 01 02 01
IMF4 1.40E- 2.05E- 6.79E- 2.22E- 4.95E- 1.52E- 1.08E- 9.74E- 2.05E- 4.88E- 1.71E- 4.31E- 1.88E- 2.23E- 8.94E- 2.15E- 7.67E- 2.04E-

Computers in Biology and Medicine 138 (2021) 104922

02 02 04 03 04 03 02 03 04 04 04 04 01 01 02 01 02 01
IMF5 1.69E- 3.21E- 1.64E- 7.96E- 1.83E- 9.71E- 1.10E- 9.42E- 2.45E- 4.42E- 2.20E- 3.72E- 2.02E- 2.13E- 9.41E- 1.77E- 8.21E- 1.66E-
02 02 03 03 03 03 02 03 04 04 04 04 01 01 02 01 02 01
IMF6 5.16E- 9.78E- 2.20E- 8.73E- 7.24E- 3.24E- 2.32E- 2.31E- 1.70E- 3.50E- 6.42E- 1.39E- 8.52E- 1.48E- 3.23E- 1.41E- 5.28E- 1.49E-
03 03 04 04 04 03 03 03 05 05 05 04 02 01 02 01 02 01
IMF7 7.21E- 8.58E- 1.13E- 2.70E- 7.45E- 2.15E- 3.77E- 3.64E- 2.03E- 4.90E- 1.49E- 3.40E- 3.00E- 3.23E- 1.39E- 2.49E- 9.13E- 2.10E-
03 03 04 04 05 04 03 03 05 05 05 05 01 01 01 01 02 01
IMF8 9.46E- 1.19E- 4.60E- 1.20E- 4.03E- 1.09E- 7.41E- 5.30E- 1.47E- 2.97E- 2.43E- 7.82E- 9.08E- 1.86E- 4.64E- 1.68E- 5.06E- 1.69E-
04 03 06 05 06 05 04 04 06 06 06 06 02 01 02 01 02 01
IMF9 1.76E- 5.53E- 1.06E- 6.88E- 1.37E- 7.48E- 5.96E- 5.69E- 1.19E- 2.29E- 7.04E- 2.01E- 4.68E- 1.46E- 2.41E- 1.42E- 2.75E- 1.43E-
03 03 04 04 05 05 04 04 06 06 07 06 02 01 02 01 02 01
IMF10 1.81E- 3.79E- 1.04E- 5.14E- 3.75E- 2.05E- 2.09E- 5.75E- 3.49E- 1.78E- 3.74E- 2.05E- 1.55E- 2.28E- 6.28E- 1.75E- 2.27E- 1.41E-
02 02 03 03 05 04 02 02 03 02 03 02 01 01 02 01 02 01
 Table 5
P-values of features in depression, Bonn and Indian databases.
IMF Depression database Bonn database Indian database
number
Depression vs. Normal Depression vs. Normal Ictal vs. Normal Ictal vs. Interictal Ictal vs. Normal Ictal vs. Interictal
M.T. Sadiq et al.

(Right halve) (Left halve)

SDC SRA STA SDC SRA STA SDC SRA STA SDC SRA STA SDC SRA STA SDC SRA STA

IMF1 3.63E- 4.31E- 4.17E- 3.23E- 8.73E- 5.70E- 4.18E- 3.22E- 3.22E- 2.37E- 7.23E- 7.09E- 9.25E- 1.56E- 1.47E- 5.56E- 1.24E- 5.12E-
176 134 118 186 147 131 45 45 45 43 42 42 17 16 16 08 07 04
IMF2 1.40E- 1.10E- 3.73E- 4.25E- 1.25E- 5.99E- 5.94E- 8.43E- 1.47E- 2.41E- 3.16E- 1.24E- 3.47E- 1.59E- 2.08E- 2.40E- 1.41E- 8.33E-
112 115 131 199 204 223 37 35 33 42 42 41 15 14 14 17 17 15
IMF3 2.49E-69 4.44E-71 3.10E- 1.15E-97 1.31E- 2.53E- 6.86E- 1.43E- 2.22E- 3.04E- 6.44E- 1.27E- 1.08E- 2.63E- 2.23E- 1.10E- 1.39E- 4.13E-
116 100 151 37 35 33 42 42 40 15 15 15 16 16 16
IMF4 1.55E-38 1.01E-39 1.01E- 5.38E-54 7.69E-54 4.13E- 2.00E- 1.18E- 1.41E- 9.11E- 1.29E- 3.32E- 3.10E- 1.09E- 5.38E- 1.04E- 1.25E- 2.55E-
109 131 33 32 30 42 41 40 15 14 15 17 17 17
IMF5 9.21E-11 2.37E-09 5.80E-76 7.38E-15 3.32E-13 2.21E-91 1.87E- 5.48E- 1.46E- 1.57E- 2.45E- 7.81E- 6.70E- 1.28E- 8.79E- 1.65E- 1.47E- 1.65E-
33 33 30 42 42 42 15 14 15 16 16 16
IMF6 3.11E-03 8.65E-03 4.24E-64 2.79E-04 8.36E-03 5.18E-67 5.54E- 1.69E- 3.63E- 2.47E- 2.83E- 1.01E- 2.45E- 3.19E- 1.09E- 6.52E- 7.33E- 5.16E-
35 34 32 43 43 42 14 14 14 17 17 17
IMF7 2.32E-06 8.66E-08 3.69E-46 2.37E-06 1.36E-07 2.11E-50 9.36E- 4.55E- 6.99E- 1.22E- 2.18E- 9.91E- 2.20E- 4.89E- 1.21E- 6.52E- 1.17E- 9.81E-
35 34 31 42 42 41 16 16 15 17 16 17
IMF8 4.45E-22 1.45E-23 1.06E-31 5.73E-23 2.27E-23 3.62E-34 5.01E- 2.10E- 5.76E- 9.12E- 8.95E- 7.37E- 1.75E- 3.10E- 1.14E- 1.31E- 2.20E- 1.47E-
36 35 33 43 43 42 16 16 15 16 16 16
IMF9 2.28E-12 9.95E-12 2.44E-31 3.02E-12 2.67E-11 5.06E-30 6.09E- 7.15E- 2.92E- 5.72E- 4.89E- 5.63E- 1.68E- 3.19E- 3.33E- 1.08E- 1.43E- 1.51E-
36 36 33 43 43 42 14 14 12 15 15 15
IMF10 1.66E-21 5.90E-11 1.53E-02 2.46E-39 1.58E-22 7.66E-02 2.95E- 2.83E- 3.64E- 1.20E- 1.10E- 3.19E- 7.43E- 2.33E- 1.16E- 7.73E- 7.73E- 7.28E-
12 12 12 18 18 20 07 06 02 18 18 18

13
Table 6
Corrected P-values of features in depression, Bonn and Indian databases.
IMF Depression database Bonn database Indian database
number
Depression vs. Normal Depression vs. Normal Ictal vs. Normal Ictal vs. Interictal Ictal vs. Normal Ictal vs. Interictal

(Right halve) (Left halve)

SDC SRA STA SDC SRA STA SDC SRA STA SDC SRA STA SDC SRA STA SDC SRA STA

IMF1 3.63E- 4.31E- 3.75E- 2.91E- 7.85E- 3.99E- 4.18E- 3.22E- 3.22E- 2.13E- 1.93E- 3.72E- 8.32E- 1.55E- 1.47E- 5.56E- 1.24E- 5.12E-
175 133 117 185 146 130 44 44 44 42 41 41 16 15 15 08 07 04
IMF2 1.26E- 9.89E- 3.73E- 4.25E- 1.25E- 5.99E- 4.75E- 5.06E- 1.17E- 7.22E- 1.26E- 6.19E- 1.67E- 4.78E- 6.25E- 1.74E- 1.13E- 1.67E-
111 115 130 198 203 222 36 34 32 42 41 41 14 14 14 16 16 14
IMF3 2.00E-68 3.55E-70 2.48E- 9.22E-97 1.04E-99 2.28E- 5.49E- 9.99E- 1.55E- 9.11E- 1.93E- 3.80E- 7.55E- 1.84E- 1.56E- 3.30E- 4.40E- 1.65E-
115 150 36 35 32 42 41 40 15 14 14 16 16 15
IMF4 1.08E-37 7.04E-39 7.08E- 3.77E-53 5.38E-53 2.89E- 4.00E- 2.35E- 2.83E- 1.82E- 2.57E- 6.65E- 1.55E- 4.26E- 2.69E- 9.39E- 9.99E- 2.29E-
109 130 33 32 30 41 41 40 14 14 14 17 17 16
IMF5 2.76E-10 7.11E-09 3.48E-75 2.95E-14 1.33E-12 1.33E-90 3.74E- 1.64E- 2.92E- 6.07E- 1.21E- 3.90E- 2.68E- 4.26E- 3.52E- 4.95E- 4.41E- 8.25E-
33 32 30 42 41 41 14 14 14 16 16 16
IMF6 3.11E-01 8.65E-01 2.12E-63 2.79E-01 8.36E-01 2.59E-66 2.34E- 8.47E- 1.82E- 2.22E- 2.55E- 1.01E- 4.89E- 6.39E- 4.17E- 2.75E- 3.67E- 3.61E-
34 34 31 42 42 41 14 14 14 16 16 16
IMF7 4.64E-06 1.73E-07 1.48E-45 4.74E-06 2.73E-07 8.46E-50 3.75E- 1.82E- 2.10E- 5.06E- 1.09E- 2.97E- 1.76E- 3.91E- 8.91E- 2.75E- 4.40E- 4.95E-
34 33 30 42 41 40 15 15 15 16 16 16
IMF8 2.67E-21 8.73E-23 3.17E-31 2.87E-22 1.36E-22 1.09E-33 3.01E- 1.47E- 3.46E- 4.56E- 6.31E- 3.72E- 1.40E- 2.48E- 8.91E- 3.93E- 6.60E- 7.35E-
35 34 32 42 42 41 15 15 15 16 16 16
IMF9 9.10E-12 4.98E-11 4.89E-31 9.07E-12 8.01E-11 1.01E-29 3.65E- 6.31E- 2.02E- 3.67E- 4.03E- 3.38E- 3.67E- 6.39E- 6.67E- 2.16E- 2.85E- 4.53E-
35 35 32 42 42 41 14 14 12 15 15 15
IMF10 8.32E-21 2.36E-10 1.53E-02 1.48E-38 7.91E-22 7.66E-02 2.95E- 2.83E- 3.64E- 1.20E- 1.10E- 3.19E- 7.43E- 2.33E- 1.16E- 6.96E- 6.33E- 7.28E-
12 12 12 18 18 20 07 06 02 17 17 17
Computers in Biology and Medicine 138 (2021) 104922
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

Table 7 used the Bonn database reported that the classification ACC was more
The selected features by the binary PSO algorithm in various classification tasks. than 95%. In other words, it seems that Bonn EEG signals alone cannot
Classification Normal vs. depression Normal vs. depression be a trustworthy database in epilepsy EEG signals classification. In this
task regard, the classification performance of the proposed framework was
(Depression -Left half) (Depression - Right half)
examined by employing the Indian database. In the Indian database, the
Selected features STA of IMF6 and 7 STA of IMF5, 8 and 10 average classification ACC of 96% was obtained for both normal against
SDC of IMF6,7 and 10 SDC of IMF5, 8, 9 and 10
SRA of IMF1, 3 and 8 SRA of IMF1, 2, 3, 7, 8 and 10
ictal, and interictal vs. ictal EEG signal classification. Besides, SPE of
100% was achieved for the interictal vs. ictal classification tasks.
Classification Normal vs. Ictal Intrictal vs. Ictal
task (Bonn database) (Bonn database)
6.1. Comparison of depression detection with other studies
Selected features STA of IMF1, 2, 3, 6 and 7 STA of IMF3, 5 and 6
SDC of IMF2, 3, 5, 8, 9 and 10 SDC of IMF2 and 10
SRA of IMF1, 3, 5, 6 and 7 SRA of IMF1, 6, 8 and 9 To evaluate the proposed framework for depression detection, 22
normal and 22 depressed subjects were considered to record the EEG
Classification Normal vs. Ictal Intrictal vs. Ictal
task (Indian database) (Indian database)
signals from the left and right half of the brain. This process is performed
by using a bipolar device. Table 12 provides a comparison between the
Selected features STA of IMF6, 7, 8 and 9 STA of IMF1, 3, 4, 6, 7 and 10
proposed framework and the existing methods for detecting depression
SDC of IMF4, 5, 7, 8 and 10 SDC of IMF3, 4, 6, 9 and 10
SRA of IMF3, 4, 6, 8, 9 and 10 SRA of IMF1, 2, 3, 7, 8 and 10 EEG signals.
Knott et al. [58] utilized linear methods to compute the relative
power, absolute power, frequency, and coherence measures for
depression and healthy subjects EEG signals. The results have presented
Table 8
The performance of the proposed framework in various classification tasks. a profile for frontal activation and described aberrant inter-hemispheric
synchrony/asymmetry as a pattern in male depression.
Classification Task ACC (%) SEN (%) SPE (%) No. of
Ahmadlou et al. [28] developed a nonlinear method based on the
K
wavelet-chaos methodology presented by Adeli et al. [38]. In this
Normal vs. Depression (Left 93.85 ± 92.16 ± 95.71 ± 9
method, the EEG signals decompose by wavelet transform to five sub
half) 1.2 0.9 0.8
Normal vs. Depression (Right 94.75 ± 91.83 ± 96.82 ± 6 bands and Katz’s and Higuchi’s fractal dimensions (KFD and HFD)
half) 0.9 0.7 0.8 computed as feature and fed to enhanced probabilistic neural network
Normal vs. Ictal (Bonn 99.33 ± 99.00 ± 99.50 ± 5 (EPNN) classifier which resulted in 91.3% of classification ACC. By
database) 0.3 0.5 0.4 Puthankattil et al. [43], EEG signals were decomposed to eight levels
Intrictal vs. Ictal (Bonn 98.33 ± 95.19 ± 99.48 ± 3
database) 0.2 0.4 0.3
and relative wavelet energy (RWE) computed as a feature and ACC of
Normal vs. Ictal (Indian 96 ± 0.3 96 ± 0.4 96 ± 0.3 5 98.11% was reported for the feed-forward neural network (FFNN)
database) classifier.
Intrictal vs. Ictal (Indian 99 ± 0.2 99 ± 0.3 98 ± 0.4 6 In proposed method by Hosseinifard et al. [36], the power, HFD,
database)
correlation dimension (CD), detrended fluctuation analysis (DFA), and
large Lyapunov exponent (LLE) features were extracted from EEG sig­
6. Discussions nals. Then, significant features were obtained by genetic algorithm and
used in logistic regression (LR) classifier. Their results showed that a
The proposed method, for detecting depression disorder needs to combination of nonlinear features and LR classifier provides an ACC of
record the EEG signals for around 8 s with sampling frequency of 256 90% in depression detection. Ahmadlou et al. [59] presented a method
Hz. Besides for detecting ictal EEG signals, it needs to record EEG signals based on spatiotemporal analysis of relative convergence (STARC) in
for around 23.6 and 2.56 s with sampling frequency of 173.61Hz and which EEG signals decomposed to five sub-bands and STRARC
400 Hz, respectively. Our recommended system can detect the depressed computed for each sub-band as discrimination feature.
and ictal EEG signals in both open and closed eyes situations, but they Bachmaan et al. [27] have developed a automated system for
should by awake. It is already proven in many previous studies [7–10, detecting depression EEG signals by combining the spectral asymmetry
12,17,18,22,23,30,31,34,37] that 4097 and 1024 samples are suitable index (SASI) with HFD. They reported best classification ACC of 94% by
samples in one EEG signal for Bonn and Indian epilepsy datasets for the SVM classifier. Faust et al. [33] proposed a entropy based method by
effective classification. In addition, for depression database 2000 sam­ computing bispectral entropy (Ph), rényi entropy (REN), approximate
ples are recommended in available literature [4,29,33]. In our previous entropy (ApEn), and sample entropy (SampEn) as features from
studies [35,35,40], we have tested many samples and found that 2000 sub-bands of wavelet packet decomposition (WPD) of EEG signals. They
samples for each signal provide accurate classification. The proposed reported best classification ACC of 99.50% for depression detection.
framework needed at least five trials/epochs to achieve classification A new depression diagnosis index (DDI) was introduced by Acharya
accuracy of 85% ± 5%, which is much better than the minimum clas­ et al. [29] using nonlinear methods, including FD, LLE, SampEn, DFA,
sification accuracy i.e. 70% required to control an automated system. Hurst’s exponent (HE), higher order spectra (HOS), and recurrence
The average classification ACC of 92.95% and 93.35% was resulted quantification analysis (RQA) which resulted in classification ACC of
in normal vs. depression EEG signal classification tasks in the left and 98% by the SVM classifier. A new EEG-based method was suggested by
right half of the brains, respectively. It is believed that the EEG re­ Liao et al. [60] and called kernel eigen-filter-bank common spatial
cordings collected from the right half of brain are better than the left one pattern (KEFB-CSP)s. The KEFB-CSP was employed to decompose the
for detecting depression. The previous studies have reported similar raw EEG signals into sub-bands. Each sub-band was spatially trans­
results with the same technique (i.e., bipolar strategy for recording the formed into the new space. In the new space, new signals (CSPs) were
EEG signals) [4,29,33,41–43,57], but we cannot generalize this concept optimized for classifying healthy subjects and major depressive disorder
to the studies with different data recording techniques [27,28,36, (MDD) patients. Then, kernel principal component analysis (kernel PCA)
58–61]. was used for the reduction of dimension. The SVM classifier was used to
In the Bonn database, the average classification ACC of 99.33%, and distinguish between two groups. This classifier gave the classification
97.33% were given for regular against ictal, and interictal versus regular ACC of 81.23%.
EEG signals identification tasks, accordingly. Most of the articles that Mumtaz et al. [61] used synchronization likelihood (SL) measure to
extract features from EEG signals. The receiver operating curve (ROC)

14
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

Table 9
Subject-wise classification accuracy results for normal and depressed groups.
Depressed group Normal group

Subject Age/gender Tr. No H CC UC ACC (%) Subject Age/gender Tr. No H CC UC ACC (%)

1 58/M 88 L 62 26 70.45 1 51/M 77 L 69 8 89.61

R 65 23 73.86 R 68 9 88.31
2 29/M 86 L 69 17 80.23 2 18/M 79 L 75 4 94.94
R 73 13 84.88 R 76 3 96.2
3 31/M 65 L 59 6 90.77 3 20/M 83 L 79 4 95.18
R 60 5 92.31 R 78 5 93.98
4 34/M 86 L 86 0 100 4 23/M 64 L 63 1 98.44
R 86 0 100 R 60 4 93.75
5 44/M 77 L 66 11 85.71 5 33/M 86 L 76 10 88.37
R 69 8 89.61 R 77 9 89.53
6 31/M 86 L 75 11 87.21 6 20/M 82 L 82 0 100
R 76 10 88.37 R 82 0 100
7 44/M 73 L 73 0 100 7 33/M 75 L 74 1 98.67
R 73 0 100 R 73 2 97.33
8 28/M 88 L 80 8 90.91 8 19/M 76 L 69 7 90.79
R 83 5 94.32 R 69 7 90.79
9 34/M 81 L 75 6 92.59 9 23/M 69 L 57 12 82.61
R 75 6 92.59 R 56 13 81.16
10 57/M 62 L 55 7 88.71 10 46/M 86 L 86 0 100
R 56 6 90.32 R 86 0 100
11 59/W 65 L 64 1 98.46 11 51/M 83 L 78 5 93.98
R 65 0 100 R 76 7 91.57
12 56/W 78 L 72 6 92.31 12 58/M 88 L 88 0 100
R 74 4 94.87 R 88 0 100
13 25/W 79 L 79 0 100 13 27/M 87 L 68 19 78.16
R 79 0 100 R 73 14 83.91
14 29/W 79 L 64 15 81.01 14 31/M 85 L 72 13 84.71
R 68 11 86.08 R 71 14 83.53
15 33/W 86 L 79 7 91.86 15 35/M 68 L 68 0 100
R 81 5 94.19 R 68 0 100
16 49/W 80 L 80 0 100 16 51/M 63 L 56 7 88.89
R 80 0 100 R 54 9 85.71
17 32/W 78 L 74 4 94.87 17 34/W 88 L 88 0 100
R 75 3 96.15 R 88 0 100
18 23/W 73 L 69 4 94.52 18 25/W 77 L 76 1 98.7
R 69 4 94.52 R 74 3 96.1
19 46/W 68 L 67 1 98.53 19 48/W 67 L 59 8 88.06
R 68 0 100 R 57 10 85.07
20 43/W 87 L 85 2 97.7 20 45/W 89 L 86 3 96.63
R 87 0 100 R 83 6 93.26
21 42/W 66 L 65 1 98.48 21 44/W 68 L 62 6 91.18
R 66 0 100 R 63 5 92.65
22 23/W 69 L 66 3 95.65 22 25/W 60 L 60 0 100
R 65 4 94.2 R 60 0 100

CC: correct classified, UC: un-correct classified, Tr. No: total trials, H: hemisphere.

Fig. 10. Feature-wise classification accuracy for subjects (a)Left hemisphere (b)Right hemisphere.

15
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

Fig. 11. Subject-wise classification accuracy (a)Left hemisphere (b)Right hemisphere.

Table 10
Best combination results for each normal subject for left (L) and right (R) hemisphere of brain.
Subject Left Best Combination Classification Accuracy Subject Right Best Combination Classification Accuracy
Region (%) Region (%)

L1 IMF2+STA 82 R1 IMF1+STA 88
L2 IMF7+SDC 81 R2 IMF2+SRA 88
L3 IMF10+SDC 82 R3 IMF1+SRA, IMF4+SDC & IMF7+SDC 82
L4 IMF10+STA 80 R4 IMF7+SDC 81
L5 IMF1+SRA 80 R5 IMF3+STA 89
L6 IMF5+SDC 88 R6 IMF4+SDC 91
L7 IMF7+STA 81 R7 IMF3+SRA & IMF3+SDC 81
L8 IMF2+SDC 80 R8 IMF4+STA 88
L9 IMF3+SDC & 82 R9 IMF2+SDC 81
IMF6+SDC
L10 IMF8+SDC & 82 R10 IMF5+SRA 88
IMF8+STA
L11 IMF9+STA 83 R11 IMF5+SRA 82
L12 IMF8+STA 83 R12 IMF6+SDC 88
L13 IMF1+SRA & 80 R13 IMF1+SRA & IMF4+SRA 83
IMF7+STA
L14 IMF2+STA & 80 R14 IMF7+STA, IMF8+STA, IMF9+SRA & 89
IMF5+STA IMF10+SDC
L15 IMF9+SRA 81 R15 IMF7+STA 80
L16 IMF9+SDC 82 R16 IMF1+STA 83
L17 IMF2+SRA 88 R17 IMF3+SRA 83
L18 IMF9+SRA 82 R18 IMF10+SDC 81
L19 IMF3+SDC & 83 R19 IMF3+SDC 82
IMF8+SRA
L20 IMF7+SRA 79 R20 IMF10+STA 89
L21 IMF5+STA & 82 R21 IMF7+STA 81
IMF8+STA
L22 IMF7+SRA 82 R22 IMF10+STA 89

method was employed as feature selection technique in which the SVM localized (BDL) three-channel orthogonal wavelet filter bank
classifier has given SEN of 96.90% and SPE of 91%. In suggested method (TCOWFB). The TCOWFB has decomposed EEG signals into seven
by Bairy et al. [41], the HOS features have been extracted from linear sub-bands, and the logarithm of L2 norm (LL2N) was computed as fea­
predictive coding (LPC) and fed to bag tree classifier which showed ACC tures and used in least square SVM (LS-SVM) classifier which resulted in
of 94.30% in normal and depression EEG signals classification. Bach­ classification ACC of 99.58% in the ten-fold cross-validation strategy.
maan et al. [57] have investigated the linear methods (i.e., alpha power The achieved classification ACC by the proposed framework is higher
variability, relative gamma power, and SASI) and nonlinear methods than the classification ACC represented in Refs. [28,36,57,58,60].
(DFA, Lempel-Ziv, and HFD) for detecting depression EEG signals. They Although the attained ACC in the proposed framework is lesser than the
claimed the best classification ACC of 92% by a combination of three suggested method in Ref. [33], our method filters the EEG signals by a
linear and three nonlinear measures in leave-one-out (LOO) cross vali­ simple low pass Butterworth. Whereas, in Ref. [33], the EEG signals
dation strategy. were filtered by a total variation filtering algorithm which needed
A deep neural network machine learning technique was suggested by heavier computation in comparison with our proposed system. Simi­
Acharya et al. [42]. In the proposed method, the convolutional neural larly, in Refs. [36,41,57,60], the results were reported in leave-one-out
network (CNN) was implemented for diagnosing depression. The accu­ (LOO) cross-validation strategy. In Refs. [27–29,33,43,58], the results
racy of the proposed method was 93.5%. Sharma et al. [4] have designed were reported without any cross-validation strategy. The proposed
an automated diagnosis of depression system using bandwidth-duration framework used the KNN algorithm for classifying any input EEG signal.

16
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

Table 11
Best combination results for each depression subject for left (L) and right (R) hemisphere of brain.
Subject Left Region Best Combination Classification Accuracy(%) Subject Right Region Best Combination Classification Accuracy(%)

L1 IMF10+STA 81 R1 IMF3+SDC 83
L2 IMF7+SDC 83 R2 IMF4+SDC & IMF6+STA 83
L3 IMF6+SDC 82 R3 IMF2+STA 82
L4 IMF4+SRA 82 R4 IMF2+STA 83
L5 IMF2+SRA 83 R5 IMF2+SRA 82
L6 IMF5+STA 81 R6 IMF2+SDC 83
L7 IMF3+SDC 80 R7 IMF1+SDC 77
L8 IMF6+SDC & IMF9+STA 82 R8 IMF1+SDC 82
L9 IMF9+STA 80 R9 IMF6+SDC 79
L10 IMF1+SRA & IMF7+SRA 82 R10 IMF3+SRA & IMF4+SRA 83
L11 IMF1+SDC 82 R11 IMF2+SDC 83
L12 IMF6+SRA 83 R12 IMF1+STA 82
L13 IMF10+STA 83 R13 IMF3+SDC 81
L14 IMF5+STA 81 R14 IMF3+STA 82
L15 IMF1+SDC 83 R15 IMF2+STA 82
L16 IMF6+SRA 82 R16 IMF10+SRA 83
L17 IMF6+SRA 82 R17 IMF3+SDC 81
L18 IMF1+STC 83 R18 IMF5+SDC 81
L19 IMF1+SRA & IMF9&SDC 79 R19 IMF8+SRA 82
L20 IMF5+STA 82 R20 IMF4+SDC, IMF4+SRA & IMF8+STA 81
L21 IMF4+SDC 80 R21 IMF3+SDC & IMF9+SRA 78
L22 IMF9+SRA 82 R22 IMF2+STA 81

Whereas, CNN and LS-SVM were used for classifying EEG signals in Refs. these were fed into LS-SVM and artificial neural network (ANN) classi­
[4,42], respectively. In summary, the proposed method with the KNN fiers, respectively.
classifier is simpler than LS-SVM and CNN classifiers. It is revealed that the proposed graphical method has been led to the
highest classification ACC in contrast with the available studies such as
references [7, 8, 9, 10, 17, 62]]. In Refs. [13, 18, 22, 23, 56]], the re­
6.2. Comparison of epilepsy detection with other studies ported classification ACC was slightly higher than our method. In
Ref. [13], the clustering method has been used. The results demon­
For the Bonn EEG dataset, the EEG signals were considered as normal strated that this is a time-consuming method. In Ref. [18], the interictal
in A and B subsets. The EEG signals of subsets C and D are chosen as versus ictal classification task was tested, while the proposed framework
interictal whereas, for subset E was nominated as ictal. Thereafter, two was tested by normal against ictal and interictal versus ictal classifica­
binary classification tasks (i.e., normal against ictal, and interictal versus tion tasks. This procedure showed that the proposed framework was
ictal) were used for evaluating the proposed framework. In Refs. [7,8, more comprehensive than the proposed method in Ref. [18]. In Refs.
13], their proposed methods have been evaluated in four classification [23,24], SODP and RPS of IMF’s of EEG signals were used for detecting
tasks, including subsets of A against E, B against E, C against E, and D ictal EEG signals in the EMD domain, respectively. In other words, their
against E. According to Refs. [9,17], their proposed studies have been method suffers from EMD defects.
proposed with employing subsets A, B against E, and C, D against E The proposed method in Ref. [56] provides a good resolution in the
classification tasks. Also, subsets C, D against E classification task have time-frequency area, but involves measuring and combining several
been calculated in Refs. [10,18,22,23,56,62]. elements. Due to the above, machine complexity can be increased and it
For the classification of regular, interictal and ictal EEG signals, is subject to the challenge of mode-mixing and noise susceptibility. In
permutation entropy, and SVM classifier were used in Ref. [7]. In comparison, our approach is not noise-sensitive. The proposed method is
Ref. [8], the mean degree, and mean strength of HVG have been simpler and more flexible than the method used in Refs. [22,23,56]. In
considered for detecting ictal EEG signals in the KNN classifier. Ac­ the Indian database, the proposed method has given classification ACC
cording to Ref. [13], the clustering technique, and SVM classifier were of 96% for both normal against ictal and interictal versus ictal EEG
considered for classifying normal, interictal, and ictal EEG signals. In signal, while the other studies presented in Table 13 were tested by a
Ref. [9], linear prediction error energy features are fed into SVM clas­ single database (i.e., Bonn database). This demonstrates that the results
sifier for classifying EEG signals. In Ref. [10], FLP error and signal en­ of the proposed method are more reliable than previous studies.
ergy have been used as discriminant features between interictal and ictal
EEG signals.
In [17], DTCWT has decomposed the EEG signals, and then various 6.3. Advantages of the proposed study
entropies and statistics based features have been computed as inputs for
general regression neural network (GRNN) classifier. Consequently, ictal Based on the results of this study, the followings advantages can be
and non-ictal EEG signals have been discriminated. In Ref. [62], highlighted:
autoregressive modeling based EMD has been proposed for detecting
ictal. In Ref. [18], tunable Q-factor wavelet transform (TQWT) and 1. A new visual biomarker with the help of a novel 2D technique is
Kraskov entropy were utilized for classifying interictal and ictal EEG proposed for the medical team to assess brain functions in order
signals. According to Ref. [56], improved eigenvalue decomposition of to detect chaotic behaviours for both depression and epilepsy
Hankel matrix and Hilbert transform (IEVDHM–HT) has been proposed disorders.
for time-frequency representation in ictal and interictal EEG signals. 2. A novel computer-aided diagnosis framework is proposed for
Then, the rényi entropy was extracted for components as a feature, and it cross-domain EEG signals identification.
was fed into the least-square SVM (LS-SVM). In Refs. [22,23], due to the 3. The proposed framework results are verified in the leave one out
elliptical patterns of SODP and RPS of IMFs in EEG signals, 2D pro­ cross-validation strategy, while in previous studies [7,9,10,
jections have been computed as discrimination features between ictal 27–29,33,36,41,43,57,58,60,62] the results were reported
and interictal EEG signals with the confidence area of 95%. Afterward, without any cross-validation strategy.

17
 M.T. Sadiq et al.
Table 12
A comparison between the proposed method and the existing studies for detecting depression.
Ref. Sub. Method Ch. Classifier C.V ACC (%) SEN (%) SPE (%)

[43], 15 N, 15 DWT + RWE 2 FFNN No 98.11 98.73 97.5

“2012” D
[28], 12 N, 12 DWT + HFD and KFD 7 PNN No 91.3 – –
“2012” D
[59], 22 N, 22 DWT + STARC 19 – – Significant difference between the gender Significant difference between the gender Significant difference between the gender
“2013” D with depression. with depression. with depression.
[36], 45 N, 45 DFA, HFD, CD and LLE 19 LR LOO 90.05 – –
“2013” D
[33], 15 N, 15 WPD + ApEn, SampEn, REN and 2 PNN No Left 98.20 Right 99.50 Left 97.10 Right 99.20 Left 99.40 Right 99.70
“2014” D BPh
[29], 15 D, FD, LLE, SampEnt, DFA, HE, HOS 2 SVM No 98 97 98.5
“2015” 15 N and RQA
[61], 30 N, 34 synchronization likelihood 19 SVM Ten – 96.9 91
18

“2017” D fold
[60], 20 N, 20 KEFB-CSP 8 SVM LOO 81.23 – –
“2017” D
[41], 15 N, 15 LPC 2 Bagged Tree LOO 94.3 91.46 97.45
“2017” D
[58], 70 N, 23 FT + non-linear features 19 jackknife No 91.3 – –
“2001” D replication
[27], 17 N, 17 HFD, SASI 7 SVM No 94% – –
“2013” D
[57], 13 N, 13 SASI, alpha power, HFD, DFA and 2 LR LOO 92 – –
“2018” D Lempel-Ziv
[42], 15 N, 15 Deep learning 2 CNN Ten Left 93.54 Right 95.96 Left 91.89 Right 94.99 Left 95.18 Right 96.00
“2018” D fold
[4], 15 N, 15 TCWFB + entropy 2 LS-SVM Ten Left 99.02 Right 99.54 Left 98.66 Right 98.66 Left 99.38 Right 99.38

Computers in Biology and Medicine 138 (2021) 104922

“2018” D fold
This work 22 N, 22 EWT + 2D dynamics of IMFs 2 KNN Ten Left 92.95 Right 93.35 Left 91.93 Right 90.95 Left 94.60 Right 95.75
D fold
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

Table 13 syndromes that cause synchronization between synapses. These

Evaluation of proposed approach with recent findings in identification of the types of syndromes make the morphological of EEG signals to be
ictal EEG signals. more stationary.
Reference number; Year 10-fold C.V Task ACC (%)

[7]; “2012” Not A against E 93.6

The proposed work is the first attempt for successful binary classi­
B against E 82.9 fication of epilepsy vs. normal and depression vs. normal EEG signals. In
C against E 88 the future, we plan to extend the proposed work for multi-class classi­
D against E 79.9 fication with a large EEG dataset.
[8]; “2014” Yes A against E 99
B against E 97
C against E 98 7. Conclusion
D against E 93
[13]; “2011” Yes A against E 99.7 These days, depression and epilepsy are two common brain disor­
B against E 96.8 ders. Visual diagnosis of these disorders entails the physician’s high
C against E 97.7
D against E 93.9
disassembly along with tedious and time-consuming practice. To deal
[9]; “2010” Yes A, B against E 96 with these problems, this paper proposes a novel method based on 2D
C, D against E 94 plotting of IMFs of EEG signals in the EWT domain with graphical fea­
[10]; “2014” Not C, D against E 95.3 tures for the detection of depression and epilepsy EEG signals. The bi­
[17]; “2016” Yes A, B against E 99.2
nary PSO was developed to select the significant features and classify
C, D against E 95.2
[62]; “2019” Not A against E 98.1 them with the KNN. The proposed method classifies the EEG signals into
C against E 95.5 normal vs. depression task (self-collected dataset 1) with an average
D against E 93.8 classification ACC of 93.35%, for Bonn epilepsy dataset (dataset 2), the
C, D against E 95.1 normal vs. ictal task with an average classification ACC of 99.33%, and
[18]; “2017” Yes C, D against E 97.5
[56]; “2018” Yes C, D against E 100
interictal vs. ictal task with an average classification ACC of 97.33%. In
[22]; “2014” Yes C, D against E 97.8 the Indian epilepsy database (dataset 3), the proposed method identifies
[23]; “2015” Yes C, D against E 98.7 the EEG signals into both normal vs. ictal and interictal vs. ictal tasks
This work; Yes A, B against E 99.3 with an average classification ACC of 96%. The suggested system is
C, D against E 97.3
flexible since it is effective in determining EEG depression and epilepsy
activities across multiple databases. The proposed method measured the
4. The proposed framework delivers average classification ACC of degree of complexity for the 2D dynamics of IMFs of EEG signals as a
93.35%, 99.33%, and 97.33% for normal against depression, feature so in the future the proposed technique can be implement for the
normal against ictal, and interictal versus ictal EEG signals, which detection of other mental disorders.
is better than or comparable with available studies.
5. The proposed framework is relatively simpler and approximately CRediT authorship contribution statement
requires 0.7, 1.8, and 0.5 s for preprocessing, and processing.
These time intervals are involved in the generation of EWT filter Muhammad Tariq Sadiq: Conceptualization, Methodology, Soft­
bank, decomposition of EEG signals to 10 IMFs, 2D plots, and ware, Validation, Formal analysis, Investigation, Writing - original draft,
extraction of three graphical features from input test signal with Writing - review & editing, Visualization, Project adminstration. Hesam
2000, 4097, and 1024 samples. Moreover, the proposed method Akbari: Conceptualization, Methodology, Software, Validation, Formal
requires less than 0.001 s for classifying an input test signal in all analysis, Investigation, Writing - original draft, Writing - review &
classification tasks due to the small number of feature vector editing, Visualization. Siuly Siuly: Writing - review & editing, Visuali­
arrays. zation. Adnan Yousaf: Writing - review & editing, Visualization. Ateeq
6. The proposed depression method employed two channels to ac­ Ur Rehman: Writing - review & editing, Visualization.
quire the EEG signals for detecting depression, while the 7, 19,
and 8 number of channels have been used in Refs. [27,28,58–61] Declaration of competing interest
respectively.
7. The PSO algorithm is applied to the 30 graphical feature arrays The authors declare no conflict of interest.
derived through EEG signals IMF’s. Thereafter, the 8 and 12
feature arrays are selected from the left and right half of the brain References
for the depression database. Whereas, the 20, 30, and 100 feature
arrays have been used for depression detection in Refs. [36,43, [1] World health organization depression key facts, URL, https://www.who.int/new
s-room/fact-sheets/detail/depression, 2021.
61]. [2] World health organization epilepsy key facts, URL, https://www.who.int/news-roo
8. The proposed 2D visual inspection approach illustrates the m/fact-sheets/detail/epilepsy, 2021.
brain’s neuron activity that cause synchronization between syn­ [3] G. Ray, An algorithm to separate nonstationary part of a signal using mid-
prediction filter, IEEE Trans. Signal Process. 42 (9) (1994) 2276–2279.
apses in epilepsy and depression disorders. As seen in Fig. 7, the
[4] M. Sharma, P. Achuth, D. Deb, S.D. Puthankattil, U.R. Acharya, An automated
proposed 2D approach displays stationary behavior for depres­ diagnosis of depression using three-channel bandwidth-duration localized wavelet
sion and highly non-stationary characteristics or complex filter bank with eeg signals, Cognit. Syst. Res. 52 (2018) 508–520.
[5] M.T. Sadiq, X. Yu, Z. Yuan, M.Z. Aziz, S. Siuly, W. Ding, A matrix determinant
morphology for normal EEG signals. In addition to Fig. 8, EEG
feature extraction approach for decoding motor and mental imagery eeg in subject
signals have led to a substantial increase synapse activity which specific tasks, IEEE. Trans. Cognit. Dev. Syst., Early Access, DOI: 10.1109/
finally contributes to ictal components in EEG signals. TCDS.2020.3040438.
9. It should also be noted that, compared with the available studies, [6] M.T. Sadiq, X. Yu, Z. Yuan, M.Z. Aziz, Identification of motor and mental imagery
eeg in two and multiclass subject-dependent tasks using successive decomposition
the proposed framework is robust, because it is efficient in index, Sensors 20 (18) (2020) 5283.
detecting the EEG tasks of depression and epilepsy taken from [7] N. Nicolaou, J. Georgiou, Detection of epileptic electroencephalogram based on
different datasets. permutation entropy and support vector machines, Expert Syst. Appl. 39 (1) (2012)
202–209.
10. For both epilepsy and depression databases, the proposed [8] G. Zhu, Y. Li, P.P. Wen, Epileptic seizure detection in eegs signals using a fast
framework can illustrate the brain’s neuron activity for weighted horizontal visibility algorithm, Comput. Methods Progr. Biomed. 115 (2)
(2014) 64–75.

19
M.T. Sadiq et al. Computers in Biology and Medicine 138 (2021) 104922

[9] S. Altunay, Z. Telatar, O. Erogul, Epileptic eeg detection using the linear prediction [37] G.R.S. Reddy, R. Rao, Automated identification system for seizure eeg signals using
error energy, Expert Syst. Appl. 37 (8) (2010) 5661–5665. tunable-q wavelet transform, Engineering science and technology, Int. J. 20 (5)
[10] V. Joshi, R.B. Pachori, A. Vijesh, Classification of ictal and seizure-free eeg signals (2017) 1486–1493.
using fractional linear prediction, Biomed. Signal Process Control 9 (2014) 1–5. [38] H. Adeli, S. Ghosh-Dastidar, N. Dadmehr, A wavelet-chaos methodology for
[11] T.N. Alotaiby, S.A. Alshebeili, F.E. Abd El-Samie, A. Alabdulrazak, E. Alkhnaian, analysis of eegs and eeg subbands to detect seizure and epilepsy, IEEE (Inst. Electr.
Channel selection and seizure detection using a statistical approach, in: 2016 5th Electron. Eng.) Trans. Biomed. Eng. 54 (2) (2007) 205–211.
International Conference on Electronic Devices, Systems and Applications [39] V. Bajaj, R.B. Pachori, Epileptic seizure detection based on the instantaneous area
(ICEDSA), IEEE, 2016, pp. 1–4. of analytic intrinsic mode functions of eeg signals, Biomedical Engineering Letters
[12] M. Mursalin, Y. Zhang, Y. Chen, N.V. Chawla, Automated epileptic seizure 3 (1) (2013) 17–21.
detection using improved correlation-based feature selection with random forest [40] H. Akbari, M.T. Sadiq, A.U. Rehman, M. Ghazvini, R.A. Naqvi, M. Payan,
classifier, Neurocomputing 241 (2017) 204–214. H. Bagheri, H. Bagheri, Depression recognition based on the reconstruction of
[13] Y. Li, P.P. Wen, et al., Clustering technique-based least square support vector phase space of eeg signals and geometrical features, Appl. Acoust. 179 (2021)
machine for eeg signal classification, Comput. Methods Progr. Biomed. 104 (3) 108078.
(2011) 358–372. [41] G.M. Bairy, O.S. Lih, Y. Hagiwara, S.D. Puthankattil, O. Faust, U. Niranjan, U.
[14] Z. Fan, M. Jamil, M.T. Sadiq, X. Huang, X. Yu, Exploiting multiple optimizers with R. Acharya, Automated diagnosis of depression electroencephalograph signals
transfer learning techniques for the identification of covid-19 patients, J.Healthc. using linear prediction coding and higher order spectra features, J. Med. Imag.
Eng. 2020 (2020), 8889412, https://doi.org/10.1155/2020/8889412. Health Inf. 7 (8) (2017) 1857–1862.
[15] P. Singh, R.B. Pachori, Classification of focal and nonfocal eeg signals using [42] U.R. Acharya, S.L. Oh, Y. Hagiwara, J.H. Tan, H. Adeli, D.P. Subha, Automated eeg-
features derived from fourier-based rhythms, J. Mech. Med. Biol. 17 (2017) based screening of depression using deep convolutional neural network, Comput.
1740002, 07. Methods Progr. Biomed. 161 (2018) 103–113.
[16] W. Zeng, M. Li, C. Yuan, Q. Wang, F. Liu, Y. Wang, Identification of epileptic [43] S.D. Puthankattil, P.K. Joseph, Classification of eeg signals in normal and
seizures in eeg signals using time-scale decomposition (itd), discrete wavelet depression conditions by ann using rwe and signal entropy, J. Mech. Med. Biol. 12
transform (dwt), phase space reconstruction (psr) and neural networks, Artif. (2012) 1240019, 04.
Intell. Rev. 53 (4) (2020) 3059–3088. [44] R.G. Andrzejak, K. Lehnertz, F. Mormann, C. Rieke, P. David, C.E. Elger,
[17] P. Swami, T.K. Gandhi, B.K. Panigrahi, M. Tripathi, S. Anand, A novel robust Indications of nonlinear deterministic and finite-dimensional structures in time
diagnostic model to detect seizures in electroencephalography, Expert Syst. Appl. series of brain electrical activity: dependence on recording region and brain state,
56 (2016) 116–130. Phys. Rev. 64 (6) (2001), 061907.
[18] S. Patidar, T. Panigrahi, Detection of epileptic seizure using kraskov entropy [45] T. Gandhi, B.K. Panigrahi, M. Bhatia, S. Anand, Expert model for detection of
applied on tunable-q wavelet transform of eeg signals, Biomed. Signal Process epileptic activity in eeg signature, Expert Syst. Appl. 37 (4) (2010) 3513–3520.
Control 34 (2017) 74–80. [46] T.K. Gandhi, P. Chakraborty, G.G. Roy, B.K. Panigrahi, Discrete harmony search
[19] S. Taran, P.C. Sharma, V. Bajaj, Automatic sleep stages classification using based expert model for epileptic seizure detection in electroencephalography,
optimize flexible analytic wavelet transform, Knowl. Base Syst. 192 (2020) Expert Syst. Appl. 39 (4) (2012) 4055–4062.
105367. [47] T. Gandhi, B.K. Panigrahi, S. Anand, A comparative study of wavelet families for
[20] J. Gilles, Empirical wavelet transform, IEEE Trans. Signal Process. 61 (16) (2013) eeg signal classification, Neurocomputing 74 (17) (2011) 3051–3057.
3999–4010. [48] M.T. Sadiq, X. Yu, Z. Yuan, F. Zeming, A.U. Rehman, I. Ullah, G. Li, G. Xiao, Motor
[21] M.T. Sadiq, X. Yu, Z. Yuan, Exploiting dimensionality reduction and neural imagery eeg signals decoding by multivariate empirical wavelet transform-based
network techniques for the development of expert brain–computer interfaces, framework for robust brain–computer interfaces, IEEE Access 7 (2019)
Expert Syst. Appl. 164 (2020) 114031. 171431–171451.
[22] R.B. Pachori, S. Patidar, Epileptic seizure classification in eeg signals using second- [49] M.T. Sadiq, X. Yu, Z. Yuan, Z. Fan, A.U. Rehman, G. Li, G. Xiao, Motor imagery eeg
order difference plot of intrinsic mode functions, Comput. Methods Progr. Biomed. signals classification based on mode amplitude and frequency components using
113 (2) (2014) 494–502. empirical wavelet transform, IEEE Access 7 (2019) 127678–127692.
[23] R. Sharma, R.B. Pachori, Classification of epileptic seizures in eeg signals based on [50] Sadiq M. T., Yu X., Yuan Z., Aziz M. Z., Motor imagery bci classification based on
phase space representation of intrinsic mode functions, Expert Syst. Appl. 42 (3) novel two-dimensional modelling in empirical wavelet transform, Electron. Lett.
(2015) 1106–1117. 56 (25), doi:10.1049/el.2020.2509.
[24] A. Bhattacharyya, M. Sharma, R.B. Pachori, P. Sircar, U.R. Acharya, A novel [51] M.K. Moridani, S.K. Setarehdan, A.M. Nasrabadi, E. Hajinasrollah, A novel
approach for automated detection of focal eeg signals using empirical wavelet approach to mortality prediction of icu cardiovascular patient based on fuzzy logic
transform, Neural Comput. Appl. 29 (8) (2018) 47–57. method, Biomed. Signal Process Control 45 (2018) 160–173.
[25] S. Mahato, N. Goyal, D. Ram, S. Paul, Detection of depression and scaling of [52] J. Too, A.R. Abdullah, N. Mohd Saad, W. Tee, Emg feature selection and
severity using six channel eeg data, J. Med. Syst. 44 (2020) 1–12. classification using a pbest-guide binary particle swarm optimization, Computation
[26] S. Deivasigamani, C. Senthilpari, W.H. Yong, Classification of focal and nonfocal 7 (1) (2019) 12.
eeg signals using anfis classifier for epilepsy detection, Int. J. Imag. Syst. Technol. [53] H. Akbari, S. Ghofrani, Fast and accurate classification f and nf eeg by using sodp
26 (4) (2016) 277–283. and ewt, Int. J. Image Graph. Signal Process. 11 (11) (2019) 29–35.
[27] M. Bachmann, J. Lass, A. Suhhova, H. Hinrikus, Spectral Asymmetry and Higuchi’s [54] H. Akbari, S. Esmaili, A novel geometrical method for discrimination of normal,
Fractal Dimension Measures of Depression Electroencephalogram, Computational interictal and ictal eeg signals, Trait. Du. Signal 37 (1) (2020) 59–68.
and Mathematical Methods in Medicine, 2013. [55] Y.-H. Kung, P.-S. Lin, C.-H. Kao, An optimal k-nearest neighbor for density
[28] M. Ahmadlou, H. Adeli, A. Adeli, Fractality analysis of frontal brain in major estimation, Stat. Probab. Lett. 82 (10) (2012) 1786–1791.
depressive disorder, Int. J. Psychophysiol. 85 (2) (2012) 206–211. [56] R.R. Sharma, R.B. Pachori, Time–frequency representation using ievdhm–ht with
[29] U.R. Acharya, V.K. Sudarshan, H. Adeli, J. Santhosh, J.E. Koh, S.D. Puthankatti, application to classification of epileptic eeg signals, IET Sci. Meas. Technol. 12 (1)
A. Adeli, A novel depression diagnosis index using nonlinear features in eeg (2017) 72–82.
signals, Eur. Neurol. 74 (1–2) (2015) 79–83. [57] M. Bachmann, L. Päeske, K. Kalev, K. Aarma, A. Lehtmets, P. Ööpik, J. Lass,
[30] M. Sharma, R.B. Pachori, A novel approach to detect epileptic seizures using a H. Hinrikus, Methods for classifying depression in single channel eeg using linear
combination of tunable-q wavelet transform and fractal dimension, J. Mech. Med. and nonlinear signal analysis, Comput. Methods Progr. Biomed. 155 (2018) 11–17.
Biol. 17 (2017) 1740003, 07. [58] V. Knott, C. Mahoney, S. Kennedy, K. Evans, Eeg power, frequency, asymmetry and
[31] S.S. Alam, M. Bhuiyan, Detection of epileptic seizures using chaotic and statistical coherence in male depression, Psychiatr. Res. Neuroimaging 106 (2) (2001)
features in the emd domain, in: 2011 Annual IEEE India Conference, IEEE, 2011, 123–140.
pp. 1–4. [59] M. Ahmadlou, H. Adeli, A. Adeli, Spatiotemporal analysis of relative convergence
[32] H. Akbari, M.T. Sadiq, Detection of focal and non-focal eeg signals using non-linear of eegs reveals differences between brain dynamics of depressive women and men,
features derived from empirical wavelet transform rhythms, Phys. Eng. Sci.Med. 44 Clin. EEG Neurosci. 44 (3) (2013) 175–181.
(1) (2021) 157–171. [60] S.-C. Liao, C.-T. Wu, H.-C. Huang, W.-T. Cheng, Y.-H. Liu, Major depression
[33] O. Faust, P.C.A. Ang, S.D. Puthankattil, P.K. Joseph, Depression diagnosis support detection from eeg signals using kernel eigen-filter-bank common spatial patterns,
system based on eeg signal entropies, J. Mech. Med. Biol. 14 (2014) 1450035, 03. Sensors 17 (6) (2017) 1385.
[34] U.R. Acharya, F. Molinari, S.V. Sree, S. Chattopadhyay, K.-H. Ng, J.S. Suri, [61] W. Mumtaz, L. Xia, S.S.A. Ali, M.A.M. Yasin, M. Hussain, A.S. Malik,
Automated diagnosis of epileptic eeg using entropies, Biomed. Signal Process Electroencephalogram (eeg)-based computer-aided technique to diagnose major
Control 7 (4) (2012) 401–408. depressive disorder (mdd), Biomed. Signal Process Control 31 (2017) 108–115.
[35] H. Akbari, M.T. Sadiq, A.U. Rehman, Classification of normal and depressed eeg [62] D. Rafik, B. Larbi, Autoregressive modeling based empirical mode decomposition
signals based on centered correntropy of rhythms in empirical wavelet transform (emd) for epileptic seizures detection using eeg signals autoregressive modeling
domain, Health Inf. Sci. Syst. 9 (1) (2021) 1–15. based empirical mode decomposition (emd) for epileptic seizures detection using
[36] B. Hosseinifard, M.H. Moradi, R. Rostami, Classifying depression patients and eeg signals.
normal subjects using machine learning techniques and nonlinear features from
eeg signal, Comput. Methods Progr. Biomed. 109 (3) (2013) 339–345.

20