5168 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 28, NO.

9, SEPTEMBER 2024

EEGDepressionNet: A Novel Self

Attention-Based Gated DenseNet With Hybrid
Heuristic Adopted Mental Depression Detection
Model Using EEG Signals
Mustufa Haider Abidi , Khaja Moiduddin , Rashid Ayub , Muneer Khan Mohammed ,
Achyut Shankar , and Stavros Shiaeles

Abstract—World Health Organization (WHO) has identi- The parameters within the detection network are optimized
fied depression as a significant contributor to global dis- with the assistance of the same COIWSO. Finally, imple-
ability, creating a complex thread in both public and private mentation results are analyzed in comparison to existing
health. Electroencephalogram (EEG) can accurately reveal detection models. The experimentation findings of the de-
the working condition of the human brain, and it is consid- veloped model show 96% of accuracy. Throughout the em-
ered an effective tool for analyzing depression. However, pirical result, the findings of the developed model show
manual depression detection using EEG signals is time- better performance than traditional approaches.
consuming and tedious. To address this, fully automatic
depression identification models have been designed using Index Terms—Depression analysis, mental depression
EEG signals to assist clinicians. In this study, we propose detection, electroencephalogram, self-attention-based gat-
a novel automated deep learning-based depression detec- ed densenet, chaotic owl invasive weed search optimi-
tion system using EEG signals. The required EEG signals zation, convolutional neural networks, spectral features.
are gathered from publicly available databases, and three
sets of features are extracted from the original EEG signal.
Firstly, spectrogram images are generated from the orig-
inal EEG signal, and 3-dimensional Convolutional Neural
I. INTRODUCTION
Networks (3D-CNN) are employed to extract deep features. EPRESSION is defined as a pervasive feeling of sadness
Secondly, 1D-CNN is utilized to extract deep features from
the collected EEG signal. Thirdly, spectral features are ex-
tracted from the collected EEG signal. Following feature
D and hopelessness that significantly impacts individuals in
their day-to-day activities. Typically, depression may persist
extraction, optimal weights are fused with the three sets for up to two weeks, and if it endures for more than one
of features. The selection of optimal features is carried out week, it is diagnosed as Major Depressive Disorder (MDD)
using the developed Chaotic Owl Invasive Weed Search [1]. According to the World Health Organization (WHO), a
Optimization (COIWSO) algorithm. Subsequently, the fused staggering 322 million people worldwide are severely affected
features undergo analysis using the Self-Attention-based
Gated Densenet (SA-GDensenet) for depression detection. by depression, leading to substantial disability [2]. Individuals
grappling with depression exhibit various symptoms, including
guilt, sadness, hopelessness, loss of interest, energy, sleep dis-
Manuscript received 31 December 2023; revised 19 March 2024 and
24 April 2024; accepted 11 May 2024. Date of publication 15 May 2024; turbances, concentration issues, and disruptions to other basic
date of current version 6 September 2024. This work was supported by routines [3]. Various factors contribute to the onset of depression,
King Saud University, Riyadh, Saudi Arabia through Researchers Sup- including physical disorders, unemployment, poverty, substance
porting Project under Grant RSPD2024R1076. (Corresponding author:
Mustufa Haider Abidi.) abuse, traumatic life events, and other life-related challenges
Mustufa Haider Abidi, Khaja Moiduddin, and Muneer Khan Mo- [4]. The recent COVID-19 pandemic has also induced depres-
hammed are with the Advanced Manufacturing Institute, King Saud sion in many individuals, attributed to factors like quarantine,
University, Riyadh 11421, Saudi Arabia (e-mail: [email protected];
[email protected]; [email protected]). lockdowns, and social distancing measures [5]. In this context,
Rashid Ayub is with the Department of Science Technology and depression poses unpredictable threats to human health, with
Innovation, King Saud University, Riyadh 11451, Saudi Arabia (e-mail: the ultimate severity potentially leading to suicidal tendencies
[email protected]).
Achyut Shankar is with the Department of Cyber Systems Engineer- [6]. Recognizing these challenges in early identification and
ing, WMG, University of Warwick, CV74AL Coventry, U.K., also with the implementing effective treatment measures is crucial for saving
Center of Research Impact and Outcome, Chitkara University, Rajpura lives [7]. The identification of depression can be done through
140401, India, and also with the School of Computer Science Engineer-
ing, Lovely Professional University, Phagwara 144411, India (e-mail: the analysis of electroencephalogram (EEG) signals. These sig-
[email protected]). nals are characterized as intricate, non-linear structures that are
Stavros Shiaeles is with the Center for Cybercrime and Economic non-invasive and non-stationary, reflecting the activities of the
Crime, University of Portsmouth, Po1 2UP Portsmouth, U.K. (e-mail:
[email protected]). human brain and its functional status [8]. Detecting abnormali-
Digital Object Identifier 10.1109/JBHI.2024.3401389 ties with the naked eye is deemed more complex due to various
2168-2194 © 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.
 ABIDI et al.: EEGDEPRESSIONNET: A NOVEL SELF ATTENTION-BASED GATED DENSENET WITH HYBRID HEURISTIC ADOPTED MDD MODEL 5169

limitations [9]. Depression primarily disrupts brain activity, and In Section II, various research works on existing depression
the EEG signal serves as a validation of the brain’s electrical detection frameworks are discussed. Section III delves into the
activity, relying on synaptic relationships. Researchers assert explanation of the developed EEG-based depression detection
that EEG signals can effectively assess the level of depression framework and deep learning models. Section IV introduces
in individuals. Visual observations conducted over the long- optimal weighted feature fusion for depression analysis. The
term EEG signal reveal notable contrasts between depressed presentation of the Self-attention-based Gated DenseNet for
individuals and those without depression, introducing complex- depression detection can be found in Section V. Section VI
ity and the potential for human errors [10]. Consequently, a covers different analyses and observations conducted for the
computer-based depression analysis model has been developed developed depression detection model, while the conclusion is
using deep learning techniques. These methods have demon- presented in Section VII.
strated commendable performance rates compared to computer-
aided techniques derived from intricate concepts [5], [11], [12]. II. LITERATURE REVIEW
These well-organized concepts play a pivotal role in shaping
the layers of depression detection approaches. Additionally, the This section presents the existing literature within the field of
layer involving multiple processing plays a crucial role in data machine learning applications for EEG.
and pattern structure identification [13]. Moreover, multi-layer
techniques find widespread application across various fields, A. Related Works
including the automotive industry, agriculture, and the medical Seal et al. [20] proposed an innovative framework, the Deep
sector [14], [15], [16]. Most deep learning structures engage in Network (DeprNet), which utilizes EEG signals for identifying
automatic learning and extract features from raw input data [17]. depression. Fu et al. [21] introduced the Symmetric Convolu-
The intricacies encountered in manually observing EEG signals tional and Adversarial Neural Network (SDCAN) to categorize
are addressed effectively by employing deep learning models to depression using EEG signals. Hamid et al. [22] suggested a
effortlessly identify essential features [18]. This deep learning fused model for depression classification, incorporating facial
eliminates the need for a specifically crafted, hand-modeled and EEG features. The Bidirectional Long Short-Term Memory
classifier to resolve classification issues [19]. Therefore, the de- (BiLSTM) deep structured model, initiated with a feature fusion
velopment of a novel automated MDD identification framework basis of facial and EEG data, outperformed other models in
becomes imperative, leveraging deep learning approaches that state-of-the-art analysis. Baek et al. [23] initiated a depression
help to solve the diverse complexities inherent in existing Major prediction model with a DNN and a multiple-regression ap-
Depressive Disorder (MDD) detection models. proach. Sharma et al. [24] developed a neural network frame-
Various contributions associated with the developed auto- work for analyzing depression using EEG signals. Akbair et al.
matic MDD identification framework are outlined below: [25] proposed a novel depression identification approach in-
r Introducing a novel MDD recognition model that employs
volving Reconstructed Phase Space (RPS) and EEG signals.
deep learning techniques and meta-heuristic approaches to Malviya et al. [26] suggested a depression recognition frame-
identify depression in individuals at the initial stages. work named Discrete Wavelet Transform (DWT), employing
r Executing optimal weighted feature fusion through the
hybrid approaches like BiLSTM and CNN to detect individual
utilization of recommended COIWSO, which determines stress levels. Souza et al. [27] proposed a DAC Stacking frame-
optimal features and optimized weights. This process work to identify anxiety and depression among individuals using
aims to achieve optimal weighted fused features, thereby standard datasets. Srinivasan et al. [38] have implemented the
maximizing the correlation coefficient between the Deep CNN model for detecting the disease at an early stage with
features. the help of three classification tasks. Here, the deep CNN model
r Implementing an effective depression detection frame-
has outperformed the efficient performance when compared with
work called SA-GDensenet. This framework incorporates other existing approaches. Sangeetha et al. [39] have developed
the developed COIWSO for parameter optimization to the Multimodal Fusion Deep Neural Network (MFDNN) model
enhance accuracy and efficacy rates compared to conven- to provide better diagnostic accuracy for different modalities in
tional techniques. the medical field. However, a different analysis was performed
r A hybrid heuristic scheme named COIWSO will be cre-
for the validation to offer better performance for the developed
ated for tuning weights and features during the optimal model. Thus, the developed MFDNN model has shown reliable
weighted feature fusion phase. Additionally, optimizing performance in degrading the misclassification error while treat-
the optimizer and activation function in Gated DenseNet ing the disorder.
is carried out to improve the overall performance of de-
pression identification.
r Conducting a performance analysis of the proposed de- B. Problem Statement
pression detection framework through multiple analy- Early recognition of illness holds significant importance for
ses involving different classifiers and heuristic tech- effective treatment. A deep learning algorithm is highly effi-
niques. The objective is to demonstrate an improved cient in discovering patterns and learning features with high
depression identification rate compared to traditional accuracy. Previous works on depression detection, as shown in
techniques. Table I, include DeprNet [20], which boasts high performance in

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.
 5170 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 28, NO. 9, SEPTEMBER 2024

TABLE I
REVIEW ON DEEP STRUCTURE-BASED DEPRESSION DETECTION TECHNIQUES

terms of speed, precision, sensitivity, and specificity. Despite its not encode the object’s position and requires extensive training
strengths, DeprNet overlooks redundancy features, potentially data.
avoiding significant aspects during feature extraction. SDCAN 1) Motivation: Numerous techniques like machine learning
quickly detects stress levels [21] with a very low false accept rate. and deep learning strategies are adapted to detect mental de-
However, it demands a substantial amount of data for detection pression. For efficient analysis, the data needs to be collected
and incurs high implementation costs. BiLSTM exhibits high from the standard datasets, but it needs to be validated with the
scalability and availability [22] while being faster than LSTM larger samples. Certainly, it often shows the misclassification
models. Nevertheless, it is ill-suited for recognition, proving error in the larger set of samples. So, the better prediction is
expensive and challenging to train. needed for the accurate performance. These challenges motivate
DNN predicts potential contextual influences on the risk the development of an efficient depression detection model using
of depression, resulting in higher detection performance [23], advanced deep learning techniques. Here, the developed model
[27]. It autonomously identifies local variations but is prone is validated with the larger datasets to handle the misclassifica-
to overfitting and struggles to extract patterns from raw data. tion issues. Moreover, the result analysis is validated with diverse
DepHNN makes model implementation straightforward [24], performance analysis to show the existing algorithms.
offering high accuracy with lower computation complexity.
However, training the classifier takes more time, and handling III. EEG-BASED DEPRESSION ANALYSIS USING AN
complex data presents challenges. SVM, RBF, and KNN facil- ADVANCED MODEL OF DEEP LEARNING WITH A SOFT
itate the easy identification of linear decision boundaries [25], COMPUTING FRAMEWORK
providing high accuracy with reduced computation complexity.
Yet, they demand high memory, exhibit low effectiveness, and A. EEG Datasets Used for Depression Analysis
possess limited generalization ability. DWT and CNN-BLSTM EEG signals are sourced from the Depression-Rest Dataset
enhance low-resolution image components to high, boasting [28], encompassing approximately 122 subjects with EEG sig-
a high compression ratio [26]. Despite these advantages, they nals. The dataset exclusively comprises pre-processed files for
lack phase transformation, leading to potential misdiagnoses analysis. Moreover, data bias occurs when there is limited in-
and poor directionality. Ensemble deep learning emerges as a formation in the datasets which shows an inaccurate represen-
promising objective approach for diagnosing mild depression tation. Here, a total of 122 subjects are taken in this analysis,
[7], automatically detecting crucial features. However, it does which seems to be considered less information in this dataset.

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.
 ABIDI et al.: EEGDEPRESSIONNET: A NOVEL SELF ATTENTION-BASED GATED DENSENET WITH HYBRID HEURISTIC ADOPTED MDD MODEL 5171

Fig. 2. Architecture of the recommended depression detection model.

status of the brain. That’s why we have used EEG signals in this
research area. Generally, the EEG offers better data privacy in
terms of user identity, emotions, and so on. Moreover, informed
consent is a procedure in which the healthcare provider informs
the patient about the risk. But, the research work is concentrated
on detecting mental depression using the developed model.
Fig. 2 illustrates the systematic layout of the proposed Major
Depressive Disorder (MDD) framework. The process begins
Fig. 1. Sample EEG signals from the considered dataset.
with obtaining EEG signals from a benchmark dataset, serving as
input to the feature extraction segment. This step yields three sets
Henceforth, the developed SA-GDenseNet model is performed of extracted features. The first set results from converting EEG
to classify mental depression efficiently. Thus, it shows better signals into spectrogram images, used as input for 3D-CNN,
impact on the result analysis in terms of accuracy maximiza- producing the initial set of features. The second set of features is
tion with the help of efficient tuning process by the designed acquired by inputting EEG signals into 1D-CNN. Subsequently,
COIWSO algorithm. Fig. 1 exhibits several sample signals spectral features obtained directly from the original EEG signals
extracted from the dataset. Here, the acquired EEG signals constitute the third set.
from the dataset are termed as SIGinph , where h = 1, 2, … …,
In the optimal weighted feature fusion stage, a meticulous
H, and H denote the total number of EEG signals for model feature selection process is executed across the three sets—3D-
analysis. CNN, 1D-CNN, and spectral features. Ten optimal features from
each set are computed using the developed COIWSO. These
optimal features are then introduced into the feature fusion
B. Proposed Depression Analysis stage, where the COIWSO-derived optimal weights maximize
Depression, a persistent mental health condition, can disrupt the correlation coefficient in the fused features.
daily routines for extended periods—years, months, or weeks. The resulting optimal weighted fused features undergo a
Addressing the complexities inherent in current approaches, deep learning-based depression detection phase. The employed
there’s an assured need for a novel framework to identify depres- SA-GDensenet is designed to classify depression. The param-
sion. The mental depression generally occurs through feelings or eters, such as the optimizer and activation function, are opti-
thoughts so; it is necessary to diagnose the mental disorder using mized by the COIWSO to enhance the depression detection
EEG signals. Here, the EEG signal is used to record the electrical accuracy rate within the proposed depression identification
signals, which tend to diagnose depression based on the working framework.

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.
 5172 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 28, NO. 9, SEPTEMBER 2024

C. Proposed Depression Analysis owl validates the distance information Ds to get the prey easily
and it is provided in (5).
The diverse machine learning approaches have seemed to
be adapted for the efficient optimization process. However, Ds = Qs , X2 (5)
there is a significant need to improve the models. Due to the
complex parameters, the existing optimization algorithms are Here, the prey location found by the fitness owl is referred
not efficient to attain a higher performance rate. In order to to as X, where X = QBst . The modified intensity based on the
solve this issue, the COS and IWO algorithms are considered for inverse square law for sound intensity is provided in (6).
their significant performance. Here, the COS ensures a balance IT Ns
Jds = + RN (6)
between the exploitation and exploration phases [19]. However, Ds2
it doesn’t address discrete-time variant-based issues. To address
Here, the term RN indicates random noise, which is utilized
these concerns, an alternative model, IWO, is integrated with
to offer a more realistic outcome. The new position allocation is
the COS algorithm to create the COIWSO algorithm. IWO
given in (7).
accurately identifies minima and effectively resolves prema-  v
ture convergence issues [29]. However, it lacks computational Qs + β ∗ Jds ∗ |αU − Qvs | Rp < 0.5
Qv+1 = (7)
complexity, which degrades the system’s performance. A novel s Qvs − β ∗ Jds ∗ |αU − Qvs | Rp ≥ 0.5
depression detection model utilizes the suggested COIWSO
The prey movement probability is termed as Rp a random
to enhance the detection rate by optimizing parameters such
number in which the limit [0, 0.5] is given α and also the constant
as the optimizer and activation function in Gated DenseNet,
utilized to minimize the linearity rate from the limit [1.9,0] in the
along with the weights in optimal features. In the proposed
iterations is offered as β that effectively allows better exploration
COIWSO algorithm, the ultimate update occurs when satisfying
in the search space.
the condition, D< 0.5. Should D be less than 0.5, the final update
IWO [29] is a nature-inspired optimization approach, and it
follows the COS approach; otherwise, the update is carried out
also holds different phases to offer a better performance rate.
following IWO. In this context, D represents a random number,
The population initialization of weed in the Dth dimension is
a newly introduced parameter provided by the computation
offered in (8).
in (1).
BF − CF Q = (Q1 , Q2 , . . . , QD ) (8)
D= (1)
W F − CF Here, the weeds position is given as Q that presents the trial
In the provided parameter equation for determining fitness solution of the optimization issue, which spreads over the Dth
levels, the terms are designated as follows: BF represents the dimension issue with some random position. Every member
best fitness, CF denotes current fitness, and WF stands for the of the population is allowed to generate a seed based on their
worst fitness. COS algorithm [19] is based on the hunting traits abilities. The seed generated by the weeds improves linearly,
of barn owls that are known for its adept hearing in nocturnal starting from the smallest possible seed for the weed with the
prey detection. At the algorithm’s inception, the primary task poorest fitness function up to the increased number of seeds in
involves establishing the initial positions of owls. The population the plant with the optimal solution. Further, the produced seeds
consists of m owls, each represented in the dimensional vector are distributed randomly in the Dth dimension search area with
D as outlined in (2). a generally offered random number and its mean is equal to zero.
The non-linearity rate gets minimized rapidly which results in
Qs = (Qs1 , Qs2 , . . . . . . .., QsD ) (2) categorizing the fitter plants together and the unwanted plants
The position of the sth owl in wth dimension is referred to as are neglected. The random function of standard deviation σ
Qsw . In a forest, to assign the initial position of every owl (3) is effectively minimizes the iteration count from the pre-defined
utilized. initial value σItl and the final value σF nl is validated in entire
time steps by utilizing (9).
Qs = QN + rand (0, 1) ∗ (QM − QN ) (3)
(ItrM x − Itr)a
Here, the upper range is termed as QM and the lower range σItr = (σItl − σF nl ) + σF nl (9)
(ItrM x )a
is referred to as QN and rand(0, 1) the random number that is
in the limit of [0, 1]. Next, the location of the owl is validated Here, the term ItrM x denotes the maximum iteration, σItr is
based on a fitness function F. Based on the fitness function, the the standard deviation of the current time step, and finally, the
intensity of information received by the owl’s ear is computed. non-linearity modulation index is presented as a in the above
Moreover, the good owl QBst is allocated to the owl which gains equation.
the enhanced intensity. Similarly, the bad owl is termed as QW st . Based on rapid reproduction, the iteration numbers of repro-
duced plants in the colony reached the maximum, that is qM ax .
F (Qs ) − QW st
IT N = (4) In this process, the competitive model kicks in to eliminate
QBst − QW st undesired plants exhibiting poor fitness functions. It also allows
Here, the range of the best owl is given QBst = the fitter plants to produce a greater number of seeds, which
M ax({F (Qs )) and also the worst owl is presented QW st = are accepted. This sequence is followed until the termination
M in({F (Qs ) : s = 1, . . . , m}) for the above equation. Every condition is met, and the plant with superior fitness is ultimately

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.
 ABIDI et al.: EEGDEPRESSIONNET: A NOVEL SELF ATTENTION-BASED GATED DENSENET WITH HYBRID HEURISTIC ADOPTED MDD MODEL 5173

Algorithm 1: Implemented COIWSO. Convolution layer: In this layer, the convolution operation
Initiate the population of COS and IWO takes place within a specific area of the input signal to produce
Allocate the parameters corresponding one-dimensional feature maps. Various kernels
For all the solutions are employed to extract diverse features from the input signal.
Validate the fitness function for an entire solution Equation (10) represents the 1-D convolution layer.
Initialize the newly developed variable D N 

Renew the random variable yb = p
a
yc ∗ lcb + ub
a−1 a a
(10)
If D<0.5 c=1
Renew the final solution by COS
Here, the convolution kernels are indicated as l, kernel count
Else
is given as b, channel number presented in the input is offered
Renew the final solution by IWO
as N, bias performed in the kernel is provided as u, convolution
End If
operator is termed as∗, and the activation function is referred to
End For
as p() in the above-mentioned equation.
Return the optimal good solution
Pooling layer: Average pooling process is utilized to analyze
the entire feature mapping that validates the signals based on
their size in the predefined pooling window. Moreover, the max-
selected as the optimal solution. The pseudocode for the COI- imal pooling window procedures choose the maximal parameter
WSO optimization model developed is provided in Algorithm 1. within a certain limit as the outcome value. The acquired features
from 1D-CNN are presented F El1D−CN N and they are provided
1) Algorithmic Description: Initially, the number of popula-
to the optimal weighted feature fusion phase.
tions of existing COS and IWO algorithms is taken into account
for the evaluation process. Then, the parameters are initiated,
C. Feature 3: Spectral Features
and the fitness function is evaluated for all the solutions in order
to attain the optimal solution. However, the term D taken as The obtained EEG signals SIGinp h from the standard re-
the random variable and it also lies in the range of 0.5. Here, sources are utilized to attain the spectral features. Here, the STFT
the iterative process is performed for validating the solutions method is employed to capture the features of EEG signals.
optimally. If the random variable D is less than the value of 0.5 STFT analyses the variations in the signal by observing its
then, the solution gets updated in the existing COS algorithm; frequency components and referencing the intensity value of
otherwise, the solution is updated in the IWO algorithm. In each component. The STFT is designed to perform localized
the end, the optimal solution is attained using the developed observations within the signal, utilizing a constant time win-
COIWSO algorithm. dow and multiplication on the input signal. The mathematical
representation of STFT is provided in (11).
IV. DEPRESSION ANALYSIS USING EEG SIGNALS BY ∞
stf t (ω, j, γ) = ∫−∞ p (d) b ∗ (d − j) exp (−iωj) md (11)
EXTRACTING THREE SETS OF FEATURES WITH OPTIMAL
WEIGHTED FUSED FEATURES The function of the window is offered as b ∗ (d − j), a conju-
A. Feature 1: Spectrogram With 3D-CNN Features gate of complexity is presented as∗, and the analytical signal
is provided as p(j) in the above STFT equation. The major
The acquired EGG signal SIGinp h is utilized to attain the spec- utilization of STFT is to process the signal based on time and
trogram images. The spectrogram images undergo processing frequency domain. The acquired spectral features are termed as
with a 3D-CNN to extract deep features. The 3D-CNN validates F Etstf t and they are further offered to the optimal weighted
the image volume using a specified kernel set [30]. This struc- feature fusion phase.
ture comprises four convolution layers with associated kernel
elements. Following this, a max-pooling layer is applied to each D. Optimal Weighted Fused Features
convolution layer, reducing spatial invariance and parameters.
Subsequently, the ReLU activation function is utilized. The The attained three sets of features from 3D-CNN
result obtained from the deepest convolutional layers is flattened F Eh3D−CN N , 1D-CNN F El1D−CN N , and spectral features
and fed into the fully connected layer, serving as the classifier and F Etstf t are utilized in the optimal feature selection with the
presenting the extracted features from the convolutional layer. help of developed COIWSO. Here, the attained optimal fea-
The attained feature from 3D-CNN is termed as F Eh3D−CN N tures from 3D-CNN, 1D-CNN, and spectral features are termed
and offered to optimal weighted feature fusion phase. as OFw3D−CN N , OFi1D−CN N and OFpstf t , correspondingly.
Then, the acquired three sets of optimal features are subjected
B. Feature 2: 1D-CNN Features as the input to the optimal weighted feature fusion phase. Here,
the optimal weighted feature fusion is performed based on (12).
The acquired EEG signals SIGinp h from the standard re-
3D−CN N
sources are offered as the input to 1D-CNN to obtain the deep g = W g1 ∗ OFw
F F utn + W g2 ∗ OFi1D−CN N
features. In 1D-CNN, time-series signals are employed as the
input in the input layer [31]. + W g3 ∗ OFpstf t (12)

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.
 5174 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 28, NO. 9, SEPTEMBER 2024

Fig. 5. Pictorial view of recommended SA-GDensenet-based depres-

Fig. 3. Pictorial representation of optimal weighted feature fusion sion detection model.
based on developed COIWSO.

Here, the term CRR indicates the correlation coefficient. The

correlation coefficient is determined using (14).

hv − h̄ (yv − ȳ)
CRR =  (14)
2
hv − h̄ (yv − ȳ)2

The worth of h the variable attained in the features is referred

to as hv and their mean is termed as h̄, sample value of y is
offered as yv and also the mean is provided as ȳ, in the above
equation.

V. DEPRESSION ANALYSIS USING EEG SIGNALS BY

Fig. 4. Structural presentation of gated densenet architecture. EXTRACTING THREE SETS OF FEATURES WITH OPTIMAL
WEIGHTED FUSED FEATURES
A. Gated DenseNet Model
Here, the term F F utng indicates the fused weighted features
and the optimized weights are presented as W g1 , W g2 and W g3 DenseNet architecture is based on densely connected patterns
by the suggested COIWSO. Here, the weights are tuned in the utilized to enhance the throughput of information flow in the
range [0.01, 0.99] and ten optimal features are selected from each network structure [33]. The architectural representation of the
set of extracted features are tuned in the range [0, 29]. Further, gated DenseNet is depicted in Fig. 4.
the attained optimal weighted fused features F F utn g are given
DenseNet connects entire layers directly in the feature map
as the input to the depression detection phase. The structural with equal size. A DenseNet can minimize the parameter count
representation of optimal weighted feature fusion is presented in the network, making training easier and eliminating redundant
in Fig. 3. feature map learning. The DenseNet structure holds Z layers.
The fitness function of the optimal weighted feature fusion is Every layer denotes the composite operation set Ks , which
to maximize the correlation among the weighted features offered holds different layers such as convolution, activation layer, batch
in (13). normalization, and pooling layer. Here, the DenseNet links the
entire layer with the succeeding layer, and also z th layer gains a
  feature map of an existing layer as the input. The output of z th
1
F t1 = arg min layer is formulated in (15).
{ 3D−CN N
OFw ,OFi1D−CN N ,OFpstf t ,W g1 , W g2 , W g3 }
CRR
(13) fs = Ks [f0 , f1 , . . . . . . .., fs−1 ] (15)

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.
 ABIDI et al.: EEGDEPRESSIONNET: A NOVEL SELF ATTENTION-BASED GATED DENSENET WITH HYBRID HEURISTIC ADOPTED MDD MODEL 5175

The output zth is offered as fs , the concatenation of the are tuned within the range of (including ReLU, softmax, linear,
feature map from the entire previous map, and is termed as and sigmoid) using the developed COIWSO. The objective
[f0 , f1 , . . . , fs−1 ], the composite functions like activation func- function of the recommended SA-GDensenet aims to enhance
tion, batch normalization, and 3 × 3 convolutions are presented the accuracy rate, as specified in (18).
as Ks in-layer output. Additionally, the gating mechanism  
1
controls the flow of network information paths. Without the F t2 = arg min (18)
use of gates, stored information easily disappears during the {OpGden
a , AFcGden } Auy
transformation at each time step.
Here, the term OpGden
a indicates the optimizers and AFcGden
B. Self-Attention-Based Gated DenseNet Model represents the activation functions. The term Auy denotes the
accuracy and it is validated by the nearness of a particular value
In general, diverse deep learning models have emerged, which and its equation is represented in (19).
promotes to provide sufficient performance. However, several
techniques are not suitable for the long-term dependency prob- (xd + xf )
Auy = (19)
lem. Therefore, the research work concentrates on the gated (xd + xf + xg + xh)
DenseNet model which has been adapted by the self-attention
Here, the true negative is represented as xf, the true positive
mechanism to provide better outcomes. Thus, it captures the
is represented as xd, the false negative is represented as xh and
long-term dependency and offers a better relationship in the
the false positive is represented as xg, respectively in the above
input sequence. In this phase, the optimal fused features F F utn
g
equation.
are given as the input to the depression detection phase. Em-
ploying the SA-GDenseNet model, the focus is on achieving an
effective depression detection rate. By integrating an attention VI. RESULTS AND DISCUSSIONS
mechanism into the identification model, the aim is to detect de- A. Experimental Setup
pression in individuals with distinctive features. Widely adopted
for establishing intra-dependency within sequence data [34], the Various experimental analyses were conducted for the pro-
self-attention scheme also serves to merge inner dependencies posed depression detection model in Python. The developed
in the actual source and target source. The mathematical repre- model has experimented with a population count of 10 and a
sentation of the self-attention model is presented in (16). maximum iteration rate of 25 for effective analysis. Comparative
  analyses were performed against techniques such as the Jaya
θ (Z) ∅(Z)U Algorithm (JA) [35], Deer Hunting Optimization Algorithms
HP = sof t max √ (16)
k (DHOA) [36], COS [19], and IWO [29]. Additionally, compar-
isons were made with various classifiers, including Bidirectional
Y R = HP y (Z) (17) Long Short-Term Memory (BiLSTM) [22], DNN [23], Support
Vector Machines (SVM) [25], and SAB-DenseNet [33].
Here, the input feature map is termed as Z, the affinity ma-
trix is given as HP, and the feature amp output is presented
as YR. Here, the entire features of HP trace the similarities B. Performance Metrics
between the two positions. Moreover, the self-attentions utilize The performance measure used to analyze the implemented
two transform functions such as θ and φ transform the input to depression detection model is provided in [37].
the minimal dimensional space. The internal products presented
in the minimal dimensional space are utilized to validate the C. Cost Function Analysis in the Developed Model
matrix of dense affinity HP. The term k indicates the scaling
factor, which is used to resolve the small gradient issues in the Examining Fig. 6 reveals an analysis of the cost function for
softmax function, and the function g is employed to learn the the proposed depression detection framework, COIWSO-SA-
effective embedding in the features. The diagrammatic view of GDenseNet. The convergence analysis carried out in this model
developed SA-GDenseNet-based depression detection is offered yielded improvements of 19.04%, 15.01%, 26.08%, and 39.18%
in Fig. 5. compared to JA-SA-GDensenet, DHOA-SA-GDensenet, COS-
SA-GDenseNet, and IWO-SA-GDenseNet, respectively. Con-
C. Self-Attention-Based Gated DenseNet Model With sequently, the suggested depression detection model demon-
Parameter Optimization strates a superior convergence rate when compared to alternative
methods.
In the depression detection model, the developed SA-
GDenseNet is employed to detect depression among individuals
using the input optimal features. The parameters in the gated D. Performance Analysis on Suggested Depression
Detection Framework
densenet, such as the optimizer and activation functions, are
tuned using the developed COIWSO. To improve the accuracy In Figs. 7 and 8, various analyses are presented for the
of depression identification, optimizers are adjusted within the proposed depression detection scheme based on COIWSO-
range of (including Adam, SGD, Adadelta, and Ftrl) through the SA-GDenseNet. The accuracy analysis of the developed tech-
suggested COIWSO. Similarly, optimized activation functions nique, utilizing 3D-CNN features, demonstrated improvements

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.
 5176 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 28, NO. 9, SEPTEMBER 2024

Fig. 6. Cost function analysis on developed COIWSO-SA-GDenseNet-based depression detection model with heuristic approaches over
“(a) 1D-CNN features, (b) 3D-CNN features, (c) spectral features, and (d) fused features”.

Fig. 7. Different analyses in developed COIWSO-SA-GDenseNet–based depression detection framework with heuristic approaches over
“(a) accuracy, (b) FDR, (c) FPR, (d) MCC, and (e) precision”.

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.
 ABIDI et al.: EEGDEPRESSIONNET: A NOVEL SELF ATTENTION-BASED GATED DENSENET WITH HYBRID HEURISTIC ADOPTED MDD MODEL 5177

Fig. 8. Multiple analyses in developed COIWSO-SA-GDenseNet-based depression detection framework with existing depression identification
techniques “(a) accuracy, (b) FDR, (c) FPR, (d) MCC, and (e) precision”.

TABLE II
STATISTICAL ANALYSIS OF THE RECOMMENDED DEPRESSION DETECTION FRAMEWORK WITH DIFFERENT FEATURES OVER CONVENTIONAL HEURISTIC
APPROACHES

of 15.85%, 14.45%, 13.09%, and 10.46% compared to con- classification model achieved a higher depression detection
ventional approaches, namely JA-SA-GDenseNet, DHOA-SA- rate.
GDenseNet, COS-SA-GDenseNet, and IWO-SA-GDenseNet,
respectively. Similarly, the precision analysis of the suggested
depression detection model showed enhanced outcomes of E. Statistical Analysis on Developed Depression
8.6%, 5.9%, 3.8%, and 0.1% compared to BiLSTM, DNN, Detection Model Over Baseline Heuristic Approaches
SVM, and SAB-DenseNet models, respectively, considering the Statistical analyses of the COIWSO-SA-GDenseNet-based
1D-CNN feature. Consequently, the recommended depression depression detection model are presented in Table II. The

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.
 5178 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 28, NO. 9, SEPTEMBER 2024

TABLE III has shown its higher efficacy on the early identification of de-
COMPARATIVE ANALYSIS OF THE RECOMMENDED DEPRESSION
DETECTION FRAMEWORK
pression among individuals. In real-world applications, mobile
based applications are performed to detect mental illness. Here,
the patient needs to be monitored with 24×7 hours to analyze
the mental illness from the particular individual. Moreover,
real-world applications of several mobile-based applications
and healthcare applications are performed where they can be
applicable in the clinical field.
Potential limitations include concerns about dataset diversity,
sensitivity to optimization techniques, model interpretability,
optimal analysis within the suggested depression detection and generalizability. It is required to have diverse datasets,
framework yields improved performance of 1.94%, 2.84%, robust optimization methods, improved model interpretability,
10.19%, and 1.9% compared to traditional heuristic mod- and exploration of broader applications.
els, namely JA-SA-GDenseNet, DHOA-SA-GDenseNet, COS-
SA-GDenseNet, and IWO-SA-GDenseNet, respectively. Con- A. Policy Implications
sequently, the developed depression identification framework
demonstrated a superior depression detection rate compared to The detection of mental health disorders becomes very essen-
alternative models. tial where untreated mental disorders could cause unnecessary
disability, suicide, and poor quality of life. However, the signif-
F. Comparative Analysis of the Developed Model icant policy implications in mental disorders help to promote
the mental health of every individual. Some of the political
The comparative analysis of the developed model for mental beliefs could affect the relationship between mental health and
depression is shown in Table III. Here, the performance shows exposure, and several social media sites were utilized to identify
that the developed model offers 94%, which is better than the the groups in the political environments. Moreover, the several
existing approaches. Throughout the validation, the developed youth development policies and other related programs for men-
model shows superiority over other existing approaches. tal health disorders in adolescents.

VII. CONCLUSION
ACKNOWLEDGMENT
Advanced depression detection framework employs meta-
The authors are thankful to Researchers Supporting Project
heuristic techniques to enhance the accuracy of detecting depres-
number (RSPD2024R1076), King Saud University, Riyadh,
sion in affected individuals. EEG signals from a standard bench-
Saudi Arabia, for funding this work.
mark serve as input for analysis, leading to three sets of features
as output through the feature extraction process. The first set
of features is derived from EEG signals, converted into spec- REFERENCES
trogram images, and input to 3D-CNN. These features are then [1] R. S. Skaik and D. Inkpen, “Predicting depression in Canada by automatic
directed to the optimal weighted feature fusion region. Similarly, filling of beck’s depression inventory questionnaire,” IEEE Access, vol. 10,
the second set of features, obtained through an EEG signal input pp. 102033–102047, 2022, doi: 10.1109/ACCESS.2022.3208470.
[2] T. Chen, Y. Guo, S. Hao, and R. Hong, “Exploring self-attention graph
to 1D-CNN, is also fed into the optimal weighted feature fusion. pooling with EEG-based topological structure and soft label for depression
Lastly, the third set of spectral features, acquired from original detection,” IEEE Trans. Affect. Comput., vol. 13, no. 4, pp. 2106–2118,
EEG signals, undergoes the same process. Optimal weighted Oct.–Dec. 2022, doi: 10.1109/TAFFC.2022.3210958.
[3] Z. Wan, J. Huang, H. Zhang, H. Zhou, J. Yang, and N. Zhong, “Hy-
feature fusion is executed by the developed COIWSO using the bridEEGNet: A convolutional neural network for EEG feature learning
three sets of extracted features. The resulting optimal weighted and depression discrimination,” IEEE Access, vol. 8, pp. 30332–30342,
features are employed in the depression detection phase. The 2020, doi: 10.1109/ACCESS.2020.2971656.
[4] A. Rafiei, R. Zahedifar, C. Sitaula, and F. Marzbanrad, “Automated detec-
SA-GDensenet is utilized to identify depression among indi- tion of major depressive disorder with EEG signals: A time series clas-
viduals. Tuning of the gated densenet parameters was achieved sification using deep learning,” IEEE Access, vol. 10, pp. 73804–73817,
through the developed COIWSO to attain the maximum ac- 2022, doi: 10.1109/ACCESS.2022.3190502.
[5] U. Ahmed, J. C.-W. Lin, and G. Srivastava, “Hyper-graph attention
curate depression detection rate. Accuracy analysis within the based federated learning methods for use in mental health detection,”
developed COIWSO-SA-GDensenet-based depression identifi- IEEE J. Biomed. Health Inform., vol. 27, no. 2, pp. 768–777, Feb. 2023,
cation technique demonstrates improvements over conventional doi: 10.1109/JBHI.2022.3172269.
[6] B.-J. Lin et al., “Mental status detection for schizophrenia patients via
approaches such as JA-SA-GDensenet, DHOA-SA-GDensenet, deep visual perception,” IEEE J. Biomed. Health Inform., vol. 26, no. 11,
COS-SA-GDensenet, and IWO-SA-GDensenet, with gains of pp. 5704–5715, Nov. 2022, doi: 10.1109/jbhi.2022.3199575.
15.85%, 14.45%, 13.09%, and 10.46%, respectively. Precision [7] V. B. de Souza, J. C. Nobre, and K. Becker, “DAC stacking: A deep learning
ensemble to classify anxiety, depression, and their comorbidity from reddit
analysis in the suggested depression detection model, focus- texts,” IEEE J. Biomed. Health Inform., vol. 26, no. 7, pp. 3303–3311,
ing on 1D-CNN features, exhibits an 8.6% improvement over Jul. 2022, doi: 10.1109/jbhi.2022.3151589.
BiLSTM, 5.9% superiority to DNN, 3.80% advancement com- [8] C. Jiang, Y. Li, Y. Tang, and C. Guan, “Enhancing EEG-based
classification of depression patients using spatial information,” IEEE
pared to SVM, and a 0.10% improvement over SAB-Densenet. Trans. Neural Syst. Rehabil. Eng., vol. 29, pp. 566–575, 2021,
Therefore, the recommended depression detection framework doi: 10.1109/tnsre.2021.3059429.

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.
 ABIDI et al.: EEGDEPRESSIONNET: A NOVEL SELF ATTENTION-BASED GATED DENSENET WITH HYBRID HEURISTIC ADOPTED MDD MODEL 5179

[9] Z. Wang et al., “Automated rest EEG-based diagnosis of depression and [26] L. Malviya and S. Mal, “A novel technique for stress detection from EEG
schizophrenia using a deep convolutional neural network,” IEEE Access, signal using hybrid deep learning model,” Neural Comput. Appl., vol. 34,
vol. 10, pp. 104472–104485, 2022, doi: 10.1109/ACCESS.2022.3197645. no. 22, pp. 19819–19830, Nov. 2022, doi: 10.1007/s00521-022-07540-7.
[10] H. Chen, Y. Lin, Y. Li, W. Wang, P. Wang, and Y. Lei, “Hybrid feature [27] H. Zogan, I. Razzak, X. Wang, S. Jameel, and G. Xu, “Explainable de-
embedded sparse stacked autoencoder and manifold dimensionality reduc- pression detection with multi-aspect features using a hybrid deep learning
tion ensemble for mental health speech recognition,” IEEE Access, vol. 9, model on social media,” World Wide Web, vol. 25, no. 1, pp. 281–304,
pp. 28729–28741, 2021, doi: 10.1109/ACCESS.2021.3057382. Jan. 2022, doi: 10.1007/s11280-021-00992-2.
[11] M. H. Abidi, H. Alkhalefah, and U. Umer, “Fuzzy harmony search [28] Kaggle, “Depression-rest dataset preprocessed,” 2020. [Online].
based optimal control strategy for wireless cyber physical system with Available: https://www.kaggle.com/datasets/tocodeforsoul/depression-
Industry 4.0,” J. Intell. Manuf., vol. 33, no. 6, pp. 1795–1812, Aug. 2022, rest-preprocessed
doi: 10.1007/s10845-021-01757-4. [29] B. Xing and W.-J. Gao, “Invasive weed optimization algorithm,” in
[12] M. H. Abidi, M. K. Mohammed, and H. Alkhalefah, “Predictive main- Innovative Computational Intelligence: A Rough Guide to 134 Clever
tenance planning for Industry 4.0 using machine learning for sustainable Algorithms, B. Xing and W.-J. Gao, Eds. Berlin, Germany: Springer, 2014,
manufacturing,” Sustainability, vol. 14, no. 6, 2022, Art. no. 3387. pp. 177–181.
[13] Z. Zhao et al., “Automatic assessment of depression from speech via a [30] S. Candemir et al., “Automated coronary artery atherosclerosis de-
hierarchical attention transfer network and attention autoencoders,” IEEE tection and weakly supervised localization on coronary CT angiogra-
J. Sel. Topics Signal Process., vol. 14, no. 2, pp. 423–434, Feb. 2020, phy with a deep 3-dimensional convolutional neural network,” Com-
doi: 10.1109/JSTSP.2019.2955012. puterized Med. Imag. Graph., vol. 83, Jan. 2020, Art. no. 101721,
[14] A.-B. Meng, Y.-C. Chen, H. Yin, and S.-Z. Chen, “Crisscross optimization doi: 10.1016/j.compmedimag.2020.101721.
algorithm and its application,” Knowl.-Based Syst., vol. 67, pp. 218–229, [31] S. Huang, J. Tang, J. Dai, and Y. Wang, “Signal status recognition based on
Sep. 2014, doi: 10.1016/j.knosys.2014.05.004. 1DCNN and its feature extraction mechanism analysis,” Sensors, vol. 19,
[15] M. H. Abidi, H. Alkhalefah, M. K. Mohammed, U. Umer, and J. E. no. 9, 2019, Art. no. 2018.
A. Qudeiri, “Optimal scheduling of flexible manufacturing system using [32] H. R. Ahmadi, N. Mahdavi, and M. Bayat, “A novel damage
improved lion-based hybrid machine learning approach,” IEEE Access, identification method based on short time Fourier transform and a
vol. 8, pp. 96088–96114, 2020, doi: 10.1109/ACCESS.2020.2997663. new efficient index,” Structures, vol. 33, pp. 3605–3614, Oct. 2021,
[16] M. H. Abidi, U. Umer, M. K. Mohammed, M. K. Aboudaif, and H. doi: 10.1016/j.istruc.2021.06.081.
Alkhalefah, “Automated maintenance data classification using recurrent [33] D. Li, H. Yang, C.-A. Yuan, and X. Qin, “Leaf classification utilizing
neural network: Enhancement by spotted hyena-based whale optimiza- densely connected convolutional networks with a self-gated activation
tion,” Mathematics, vol. 8, no. 11, 2020, Art. no. 2008. function,” in Intelligent Computing Methodologies, D.-S. Huang, M. M.
[17] M. Misaghi and M. Yaghoobi, “Improved invasive weed optimization Gromiha, K. Han, and A. Hussain, Eds. Berlin, Germany: Springer, 2018,
algorithm (IWO) based on chaos theory for optimal design of PID con- pp. 370–375.
troller,” J. Comput. Des. Eng., vol. 6, no. 3, pp. 284–295, Jul. 2019, [34] L. Huang, Y. Yuan, J. Guo, C. Zhang, X. Chen, and J. Wang, “Interlaced
doi: 10.1016/j.jcde.2019.01.001. sparse self-attention for semantic segmentation,” 2019, arXiv:1907.12273.
[18] H. Göker, “Automatic detection of migraine disease from EEG signals [35] R. Motamarri and N. Bhookya, “JAYA algorithm based on Lévy flight
using bidirectional long-short term memory deep learning model,” Sig- for global MPPT under partial shading in photovoltaic system,” IEEE
nal, Image Video Process., vol. 17, no. 4, pp. 1255–1263, Jun. 2023, J. Emerg. Sel. Topics Power Electron., vol. 9, no. 4, pp. 4979–4991,
doi: 10.1007/s11760-022-02333-w. Aug. 2021, doi: 10.1109/JESTPE.2020.3036405.
[19] W. H. El-Ashmawi, D. S. A. Elminaam, A. M. Nabil, and E. El- [36] R. P. Selvam, K. Narayanasamy, and M. Ilayaraja, “Efficient deer hunting
desouky, “A chaotic owl search algorithm based bilateral negotiation optimization algorithm based spectrum sensing approach for 6G communi-
model,” Ain Shams Eng. J., vol. 11, no. 4, pp. 1163–1178, Dec. 2020, cation networks,” in AI-Enabled 6G Networks and Applications. Hoboken,
doi: 10.1016/j.asej.2020.01.005. NJ, USA: Wiley, 2023, pp. 111–129.
[20] A. Seal, R. Bajpai, J. Agnihotri, A. Yazidi, E. Herrera-Viedma, and O. [37] Wiki, “Receiver operating characteristic,” Wikipedia, 2023. Accessed:
Krejcar, “DeprNet: A deep convolution neural network framework for Nov. 10, 2023. [Online]. Available: https://en.wikipedia.org/wiki/
detecting depression using EEG,” IEEE Trans. Instrum. Meas., vol. 70, Receiver_operating_characteristic
2021, Art. no. 2505413, doi: 10.1109/TIM.2021.3053999. [38] S. Srinivasan, D. Francis, S. K. Mathivanan, H. Rajadurai, B. D. Shiv-
[21] R. Fu et al., “Symmetric convolutional and adversarial neural net- ahare, and M. A. Shah, “A hybrid deep CNN model for brain tumor
work enables improved mental stress classification from EEG,” IEEE image multi-classification,” BMC Med. Imag., vol. 24, 2024, Art. no. 21,
Trans. Neural Syst. Rehabil. Eng., vol. 30, pp. 1384–1400, 2022, doi: 10.1186/s12880-024-01195-7.
doi: 10.1109/tnsre.2022.3174821. [39] S. K. B Sangeetha et al., “An enhanced multimodal fusion deep learning
[22] D. S. B. A. Hamid, S. B. Goyal, and P. Bedi, “Integration of deep neural network for lung cancer classification,” Syst. Soft Comput., vol. 6,
learning for improved diagnosis of depression using EEG and fa- 2024, Art. no. 200068, doi: 10.1016/j.sasc.2023.200068.
cial features,” Mater. Today: Proc., vol. 80, pp. 1965–1969, Jan. 2023, [40] V. Tejaswini, K. S. Babu, and B. Sahoo, “Depression detection from
doi: 10.1016/j.matpr.2021.05.659. social media text analysis using natural language processing techniques
[23] J.-W. Baek and K. Chung, “Context deep neural network model for and hybrid deep learning model,” Assoc. Comput. Machinery Trans.
predicting depression risk using multiple regression,” IEEE Access, vol. 8, Asian Low-Resource Lang. Inf. Process., vol. 23, no. 1, pp. 1–20,
pp. 18171–18181, 2020, doi: 10.1109/ACCESS.2020.2968393. 2024.
[24] G. Sharma, A. Parashar, and A. M. Joshi, “DepHNN: A novel hybrid neural [41] M. A. Wani et al., “CoDeS: A deep learning framework for identi-
network for electroencephalogram (EEG)-based screening of depression,” fying COVID-caused depression symptoms,” Cogn. Comput., vol. 16,
Biomed. Signal Process. Control, vol. 66, Apr. 2021, Art. no. 102393, pp. 305–325, 2024.
doi: 10.1016/j.bspc.2020.102393.
[25] H. Akbari et al., “Depression recognition based on the recon-
struction of phase space of EEG signals and geometrical fea-
tures,” Appl. Acoust., vol. 179, Aug. 2021, Art. no. 108078,
doi: 10.1016/j.apacoust.2021.108078.

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on December 27,2024 at 11:20:22 UTC from IEEE Xplore. Restrictions apply.