1

An EEG-based Machine Learning Framework

for Depression Detection using Effective
Connectivity Analysis
Jannatul Ferdous Srabonee, Fazla Rabbi Mashrur, Umaima Afifa, Antora Dev, N. Roy, Md. Kafiul
Islam, Chiranjeeb Biswas, Helal Uddin Ahmed, Md. Ashraful Amin, Farhana Sarker, Ravi
Vaidyanathan, and Khondaker A. Mamun

Abstract Objective: Clinical depression, often referred to as major depressive disorder (MDD), is one of the most prominent
mental illnesses with significant economic consequences. MDD is a chronic condition that needs intensive treatment and
professional supervision. One of the most challenging aspects of treating MDD is getting an early diagnosis. The standard
approaches to diagnosing MDD rely on questionnaires and other psychiatric assessment methods. However, these evaluations are
subjective and therefore may lead to misinterpretations. Researchers are now concentrating on developing the diagnostic process
using Electroencephalogram data since they are highly informative and patient-friendly. Methods: This study proposes a machine
learning-based framework for distinguishing MDD patients from healthy individuals by employing an online dataset of 64
participants (30 healthy, 34 MDD) in resting state with two different conditions (eyes open and eyes closed). After preprocessing
the raw signal, features were extracted in both time domain and time-frequency domain. The Support Vector Machine (SVM) with
10-fold cross-validation and leave one sample out (LOSO) validation techniques were used to categorize MDD and healthy groups
after feature selection using the wrapper-based Recursive feature elimination technique. Results: In both methods, 100% accuracy
was attained in all brain areas except the occipital lobe. Granger causality-based classification achieved a 99.167% accuracy rate.
Conclusion: Our proposed model outperforms the existing methods and shows that depression affects the whole brain region. No
specific lobe or hemisphere can be identified as a distinguishing trait between MDD and healthy patients.

Index Terms— Brain Connectivity Analysis, Electroencephalogram (EEG), Granger Causality (GC), Major Depressive Disorder (MDD),
Support Vector Machine (SVM).
Clinical and Translational Impact Statement— The proposed machine learning (ML) based depression classification method can
be implemented as an MDD diagnosis tool.
I. INTRODUCTION1

1
This work was supported in part by Information and Communication
Technology (ICT) Division, Ministry of Posts, Telecommunications and
Information Technology, Bangladesh, under Grant
56.00.0000.028.33.003.21-208; and in part by the Institute for Advanced
Research (IAR), United International University (UIU), Bangladesh, under
Grant UIU/IAR/01/2021/SE/16.
J. F. Srabonee, F. R. Mashrur, U. Afifa, and K. A. Mamun are with
Advanced Intelligent Multidisciplinary Systems Lab, United International
University, Dhaka, Bangladesh (Corresponding Author: K. A. Mamun,
e-mail: [email protected]).
N. Roy is with Institute of Natural Sciences, United International
University (UIU).
Md. Kafiul Islam is with Biomedical Instrumentation and Signal
Processing Lab (BISPL), Independent University, Bangladesh (e-mail:
[email protected]).
Chiranjeeb Biswas is with Department of Psychiatry, Medical College
for Women & Hospital (MCWH).
Helal Uddin Ahmed is with Child Adolescent & Family Psychiatry,
National Institute of Mental Health (NIMH).
Md. Ashraful Amin is with Department of Computer Science &
Engineering, Independent University, Bangladesh (IUB).
Farhana Sarker is with Department of Computer Science & Engineering
University of Liberal Arts Bangladesh (ULAB).
Ravi Vaidyanathan is with Department of Mechanical Engineering and
UK Dementia Research Institute Care, Research and Technology Centre
(DRI-CRT), Imperial College London, London, UK.
C LINICAL depression, commonly known as major
depressive disorder (MDD), is one of the most severe
and pervasive mental diseases around the world. It is defined
units. The units represent groups of neurons, individual
neurons, or brain regions that are segregated by anatomy [8].
Researchers have utilized different machine learning (ML)
by chronic melancholy, lack of interest or pleasure, poor algorithms to diagnose MDD in the recent years [16], [17]. B.
energy, worsening appetite and sleep, and even suicidal Hosseinifard et al. used SVM classifier with Genetic
ideation, which interferes with everyday tasks and algorithm for feature selection and achieved an accuracy of
psychological functions. Since 2008, the World Health 88.6% on classifying depressed patients [9]. In [10] S.
Organization has ranked depression as the third leading Mantri et al. experimented on 13 depressed and 12 normal
cause of disease burden, and by 2030, it is predicted to be the subjects and employed SVM for classification. The authors
leading cause [1], [2]. Depression accounted for 34.1 million suggested that linear analysis of EEG can be an efficient
years lived with disability in 2016, according to the Global method for identifying depressed patients from normal
Burden of Diseases, Injuries, and Risk Factors Study [1]. In subjects. Bachmann et al. analyzed 30 channels EEG signals
addition, 30–35% of Major Depressive Disorder (MDD) or from 13 control and 13 depressed groups using both linear
and nonlinear methods. They used logistic regression for
chronically depressed people attempt suicide annually, and
classifying with leave-one-out cross validation method and
2–15% of them result in fatality, according to WHO data [3].
attained a maximum accuracy of 92% with a mixed
In addition, the COVID-19 epidemic exacerbated depression
combination of three linear and three non-linear measures.
among the general population, especially owing to worry [11] Automated EEG-based machine learning classification
caused by job loss, financial difficulties, the death of closed algorithms have shown their potential to properly categorize
ones due to COVID-19, and the inability to move outdoors. depressed and healthy participants [12], [13]. A diagnostic
Therefore, technology-based identification of depression is tool for depression based on nonlinear characteristics
crucial for diagnosing and treating this mental disorder. derived from EEG data [14] has recently been proposed.
Traditionally, mental health professionals employ the Using relative wavelet energy and wavelet entropy, the
clinical questionnaire-based evaluation to diagnose authors of [15] conducted linear discriminant analysis and
depression, which heavily rely on patients' responses and radial basis function with a classification accuracy of
behavioral actions [4]. Nonetheless, the questionnaire is 93.33%.
particularly susceptible to human subjectivity, which In addition to ML methods, researchers also used deep
undermines the objectivity of the diagnostic procedure [5]. learning (DL) to diagnose MDD. Acharya et al. [16]
Therefore, individuals with mood disorders often suffer from proposed a 13-layer convolutional neural network (CNN) to
incorrect medication trials or numerous trials before a classify healthy and depressed subjects obtaining
definitive diagnosis. MDD is usually diagnosed using classification accuracies of 96% and 93%, respectively, for
criteria from the Diagnostic and Statistical Manual of Mental the right and left hemispheres. Zhang et al. [17] developed a
Disorders (DSM) and as treatment response [4]. In cases one dimensional (1D) CNN fed with a fusion of EEG data
when the DSM categorization is ambiguous, and the and demographic data (age and gender) which achieved
subjective clinical impression is imprecise, an efficient 75.29% classification accuracy of healthy and depressed
diagnostic tool using objective brain imaging measures, for participants. Similarly, using 1D EEG as inputs, Mumtaz et
example, is urgently required. As a consequence, studies al. [18] suggested a combined CNN and long short-term
have been done to improve the efficacy of the conventional memory (LSTM) model that achieved 98.32% accuracy (30
model and to establish more effective replacement healthy individuals and 33 depressed patients). Nevertheless,
diagnostic tools for depression, such as visual evaluation of using both CNN and LSTM is a sophisticated design that is
facial expression [11], Magnetic Resonance Imaging (MRI) expensive to compute.
[12], Heart Rate Variability (HRV) [13], and frequent social In this research, the effective connectivity (EC) between
media use [14]. However, these approaches have a number MDD and healthy subjects has been investigated. It has been
of limitations, such as the need for long-term and attentive a prominent analytical technique in recent neuroscience
monitoring, the volatility of the HRV analysis, and the attributed to its ability to portray the causal effect that one
impact of body motions on the results. Despite the fact that neural system exerts on another, hence revealing the
MRI provides the most precise findings among these direction of influence [19], [20]. This effect can be examined
methods, it is quite expensive, hardly portable, and time directly via signals, known as data-driven EC, or
consuming [6]. On the other hand, electroencephalography constructed using models that identify causal relations
(EEG) is less expensive, portable, and provides high within separate brain areas, known as model-driven EC.
temporal resolution which records the electrical activity of Numerous researchers have devised diverse data-driven
the brain [7]. These advantages make EEG an excellent techniques for evaluating effective connections based on
candidate to identify biomarkers of depression in patients. Granger causality (GC) [21]. GC is a statistical method in
Brain connection has shown great potential in evaluating which cause precedes and helps predict the effect. In this
brain function in patients with various neurological disorders method, the concept of causality is applied to time series data
as well as monitoring various cognitive and emotional brain [22]. This strategy is a useful middle ground between highly
states in healthy subjects. It refers to a network of statistical model-dependent and fully model-free approaches [23]. First
interdependence (functional connectivity), anatomical conceived within the setting of economics, it has since been
linkages (anatomical connectivity), or causal interactions widely used in the neuroscience and other fields. In general,
(effective connectivity) between separate nervous system
 3

if a signal is anticipated more successfully by the prior II. MATERIALS AND METHODS
formation from a second signal then the second signal can be
deemed causative to the first [24]. While GC was previously A. Dataset Description
used to forecast the causality of bivariate signals in the time For this study, we used the dataset contributed by Mumtaz
domain, it has now modified to include multivariate signals et al. [34] and is publicly available at [35]. 34 MDD patients
[25], [26]. (17 females and 17 males) participated in this study and
Recent research has investigated brain connection to fulfilled the globally accepted Diagnostic and Statistical
comprehend the impact of depression on brain networks. Manual-IV (DSM-IV) depression criteria. In addition, a
Compared to healthy individuals, depressed people exhibit control group of 30 healthy individuals (9 females and 21
abnormal brain connections [27]–[29]. Compared to healthy men) were recruited. The healthy people were examined for
controls (HC), Olbrich et al. [30] found impaired functional signs of mental disorders and found to be free of such
connectivity across distinct scalp sites in the frontal and problems. The data was collected from the participants
temporal regions. Xie et al. combined phase lag index based (MDD patients and HC during the eyes closed (EC) and eyes
brain functional network with a standard two-dimensional open (EO) conditions for 5 minutes), with minimum eye
(2D) CNN [31] classified 10 HC and 10 patients with blinks and head movements at 256 Hz sampling rate. The
67.67 % accuracy. Li et al. [32] employed a similar researchers employed a 19 electrode EEG cap, which were
combination in which functional connectivity from distinct placed according to the international 10-20 on-scalp
EEG bands was converted into pictures and trained on placements scheme [36] as portrayed in Figure 2. Among the
2-stacked CNN to obtain an accuracy of 80.74 % for patients sensors, the frontal (F3, F4, Fp1, F7, Fz, F8, and Fp2)
with moderate depression and HC. Combined feature included seven electrodes, parietal (Pz, P3, and P4), central
matrices derived from inter-hemispheric asymmetry and (Cz, C3 and C4), occipital (O2, O1), and temporal areas (T5,
cross-correlation of 64 channel EEG data were utilized to T3, T4, T6) were also covered. A band-pass filter (0.1-70 Hz)
train a 2D CNN with an accuracy of 94.13% for 16 MDD and a notch filter (50 Hz) were applied.
and 16 HC. EEG recordings from 30 MDD patients and 30
HC (HC) were utilized by Danish M. Khan et al. [33] to B. Pre-processing
evaluate the effective connectivity across the brain default We use EEGLAB toolbox [38] and MATLAB 2020a
mode network (DMN). Then, a three-dimensional (3D) CNN (MathWorks, Natick, MA) for customized scripts to
was trained and evaluated using effective connections within preprocess the data. At first, we perform z-score
six key sections of the DMN. Half of the participants' normalization of the signal for each participant. To reduce
connection samples were utilized for training, while the low frequency drift and high frequency noise, a band pass
other half were used for testing. In categorizing MDD and filter with a low cut-off frequency of 0.5 Hz is applied. After
HC test participants, the accuracy, sensitivity, and specificity filtering, artifact subspace reconstruction (ASR) [39], an
of their suggested method for the diagnosis of MDD were all automated artifact rejection method, is used to interpolate
100%. the flawed regions of the data without eliminating them.
In this study, we propose a machine learning based ASR identifies clean data (calibration data) and computes
framework for the classification of MDD and healthy the standard deviation of PCA-extracted components. It
individuals using resting state EEG signals. The discards data regions that are more than 20 times the
contributions of this article are summarized as follows: calibration data's standard deviation. Finally, independent
• To the best of the author's knowledge, this is the first component analysis (ICA) [40] is performed to remove
approach achieving an accuracy of 100% utilizing ML artifacts (muscle, eye blinks, or eye movements) from the
framework (SVM) for differentiating healthy persons and data without eliminating the impacted data segments. The
multiple artifact rejection algorithm (MARA) is used to
depressed subjects from time and time-frequency domain
detect artifactual components automatically. MARA is a
EEG features.
linear classifier that extracts six features from the spatial,
• A detailed experimental assessment (by tuning
spectral and temporal domains to learn from expert
hyperparameters) is performed to determine the viability of assessments [41].
the presented strategy.
• Granger Causality is used in the study to conduct causal C. Feature Extraction
analysis, and the results reveal a substantial causal difference In this subsection, we describe the varieties of features
between healthy and depressed individuals that may serve as specifically, the time-frequency domain and time domain.
a biomarker for the diagnosis of depression. Here, we used two MATLAB toolboxes namely EEG
The remainder of the article is structured as follows: We Features Extraction Toolbox [42] and EMG Features
provide the study's structural diagram and a description of Extraction Toolbox [43], [44] for extracting various features.
the experimental data in Section II. Several experimental We also estimated some spectral features since in recent
findings are presented in Section III. In Section IV, the studies spectral features have shown great potential in
benefits and drawbacks of our strategy are described classifying EEG signals [45], [46].
together with pertinent experimental findings. Finally,
Section V presents the restrictions, drawbacks, and
conclusion.
 1) Time Domain anticipation of X, beyond the amount to which X is
Time domain feature is extracted from the preprocessed forecasted by its own history and of Z’s history [22].
signals. We used the toolboxes and also GC based time
domain features. 2) Time-Frequency Domain
The temporal and frequency characteristics of EEG waves
a) Granger Causality (GC) Estimation make them difficult to analyze. This study uses wavelet
EEG data is highly suited for GC analysis due to its high packet transformation (WPT) to break EEG signals into six
temporal resolution, quick sampling, and often stochastic bands, which can be used to recover frequency information
character [22]. In this work, we use Multivariate Granger while in the temporal domain. WPT is extensively used in
Causality (MVGC) [22] toolbox for GC estimation. The the literature to identify EEG frequency bands [49]–[51].
algorithm computes MVGC using time series data in both
frequency and time domains. It eliminates separate full and a) Wavelet-Packet Transform
reduced regressions, which is a typical cause of statistical WPT breaks down the signal into its component coefficients
inaccuracy and computer inefficiency, and so improves on in two ways: in detail and approximatively. All of the
traditional techniques for GC inference [22]. The GC coefficients that were found to depend on the EEG signals'
formalism relies on VAR (Vector Auto Regressive) temporal and frequency domain properties can be termed
modeling for forecasting. VAR(p) is calculated as following: features. WPT recursively applies low-pass and high-pass
filters to generate a signal's subspace tree including
�
information about its frequency components [52]. Let,
�� = �� ⋅ ��−� + �� (1) Qa,b(k), n = 0, .., 2a − 1, denote the WPT coefficients at level
�=1 a. There are two orthogonal bases for wavelet packet
coefficients, and they are used to compute them

here, p = model order. In this experiment, the Akaike (�) =

�−1
�(�) (2� + 1 − lmod ��−1) (7)
information criterion (AIC) was used to establish the proper �,2� �=0
�−1
�−1,�

model order, which helps to avoid under-representation or �,2�+1

(�) =
�=0
�(�)
�−1,�
(2� + 1 − lmod ��−1) (8)

overfitting of the model [47], [48]. The AIC for n variables is

defined as where H(s) and G(s) are the impulse responses of the wavelet
2��2 packets' highpass and lowpass filters, respectively, and I =
���(�) = ��(���(∑)) + (2)
� 1... N and Nb = N/2b, according to [52]. In this paper, we
employ the Meyer wavelet to compute the sub-bands, due to
where Σ represents the prediction error covariance matrix of its previous superior performance utilizing EEG signals by
the bivariate autoregressive model, p is the model order, and Mamun et al. in [53].
T is the signal length. In order to extract time-frequency domain features, EEG
To demonstrate the conditional situation, the universe U, data was split into five levels yielding six bands: delta
variables that are known is considered to be divided into (0-4Hz), theta(4-8Hz), alpha (8- 12Hz), beta1 (12-20Hz),
three interdependent multivariate processes. beta2 (20-32Hz), and gamma (32-64Hz). Then, a total of 500
�� features, all of which were retrieved from each band.
�� = �� (3) Relative and average power ratios were also estimated as
�� distinct features.

and any combined influence of Z on the determination of Y to

X G-causality needs to be eliminated. The VAR(p) is D. Feature Selection
analogously split full and reduced regressions that needs to In this work, the wrapper-based SVM-Recursive Feature
be considered. Elimination with correlation bias reduction [54] was utilized
� � �
to choose the optimal features. The wrapper-based technique
�� = ���,� . ��−� + ���,� . ��−� + ���,� . ��−�
(4) assesses the significance of a subset of features by training a
�=1 �=1
+ ��,�
�=1
model on them. The method is based on greedy search and
� � evaluation criteria. Tt is greedy because the technique selects
�� = �'��,� . ��−� + �'��,� . ��−� + �'�,� (5) the locally optimum subset of features at each iteration. The
�=1 �=1 evaluation criterion then takes over as the judge [55]. In
bioinformatics, SVM-RFE is a strong feature selection
the causality Y → X conditioned on Z, which is written technique. Additionally, when there are strongly associated
FY→X|Z: features, it may be biased, and the correlation bias reduction
|�'�� | approach is employed to tackle it [54]. The nonlinear
ℱ�⟶�|� ≡ �� (6)
|��� | SVM-RFE methodology is chosen for non-linear optimal
decision since it includes a cutting-edge kernel technique.
But in this case, the addition of Z in both regressions is Moreover, we applied the radial basis function (RBF) kernel
regarded as its combined influence. FY→X|Z is referred to as which is non-linear. SVM-RFE may depict feature
- the extent to which the previous values of Y assists dependencies for its backward elimination strategy. This
 5

technique is less susceptible to overfitting, can maximize the an empirical technique to discover the minimal quantity of
use of training data, is faster, particularly for many candidate features appropriate for training the classification model in a
features. This is because it does not employ cross-validation way that avoids over-fitting. According to the process, we
(cv) accuracy on the training data as a selection criterion. establish the optimum quantity of features by iteratively
evaluating the model performance for each subgroup of
E. Classification
features chosen from the top 30 features. Figure 3 shows the
For classification, SVM is used with Gaussian Radial influence of hyper-parameters such as C and gamma on our
Basis kernel function (RBF). SVM uses a similarity function model's performance (accuracy).
to add features to tackle non-linear problems, measuring We apply k-fold cross validation and
how much each occurrence matches a certain landmark. The leave-one-sample-out (LOSO) approaches to explore the
similarity function utilized in this experiment is RBF, which feasibility and generalizability of our suggested strategy. In
is a bell-shaped function ranging from 0 (extremely far away both instances, we achieved 100% accuracy almost within
from the landmark) to 1 (at the landmark). This type of the whole brain, as shown in Table 1. Log energy entropy,
kernel works effectively in situations when the training set is cardinality, Hjorth complexity, kurtosis, detrended
not too broad and it also performed well in earlier research fluctuations analysis, the absolute value of summation root
based on EEG signals [56]. This kernel can be expressed square, Renyi entropy, skewness, and normalized second
mathematically as shown in the equation: difference are some of the prominent features in the
||�1 − �2 ||2 time-frequency domain. New zero-crossings were helpful in
�(�1 , �2 ) = ���( − ) (9)
2�2 classifying MDD and healthy groups in the time domain.
However, in the occipital lobe, the accuracy of both
hyperparameters: where, σ is the variance and our procedures is slightly lower than in other brain areas. The
hyperparameter, ||X1-X2 || is the Euclidean Distance between categorization was also done using Granger causality, which
two points X1 and X2 [57]. The SVM classifier's yielded a 99.17% accuracy within the complete brain.
hyper-parameters, particularly the regularization parameter We also compute the average of Granger matrices for all
C and the gamma, are tweaked while training to improve MDD and HC participants in both eyes open and eyes closed
classification performance, as proved most efficiently and state to provide a more comprehensive picture of the
accurately by Hsu et al. in [58]. The LIBSVM function [59] connectivity for each class. We used the F-test to perform a
is used to categorize in the proposed study. significance test where if the probability value is less than
In this study, three types of validation procedures were the significance value of 0.05, the null hypothesis is rejected.
utilized to distinguish MDD patients from HC: 10-fold cv, This is equivalent to a signal y Granger causing x. The
random split (80% train – 20% test), and leaving one subject averaged Granger matrices with significant connectivity are
out (LOSO). illustrated in Figure 5.
F. Evaluation Metrics
IV. DISCUSSION
For measuring the performance of the classification model,
the metrics- accuracy in case of 10-fold cross validation, In this study, we present a ML framework for classifying
LOSO and accuracy, sensitivity, specificity, and f1-score in depressed and healthy subjects from resting state EEG
random splitting method were considered. signals. Along with various time and time-frequency
domain features, we also compute Granger connectivity to
analyze the causal relationships within different brain
�� + ��
�������� = (10) regions. Our proposed strategy has shown great results in
�� + �� + �� + �� classifying MDD patients from HC with an accuracy of
�� 100%. Our study followed a similar approach with a
����������� = (11)
�� + �� different algorithm of the original research work by Mumtaz
et al. in [34]. The comparison between this study and the
��
����������� = (12) original paper is shown in Table 2. From Table 1 it can be
�� + �� observed that maximum of the features in any region is from
2�� the delta and theta bands.
�1 − ����� = (13)
2�� + �� + ��
A. Effects of Depression on Different Brain Regions
where FP, FN, TP and TN stand for the number of false
From the classification result, it can be seen that none of the
positive, false negative, true positive, and true negative
validation methods produce inter-hemispheric result
respectively.
differences. All brain lobes except the occipital lobe
obtained 100% accuracy in both validation methods. This
observation suggests that depression affects the entire brain
III. RESULTS and that no specific region or lobe is responsible for it. The
In this experiment, we implement the protocol in different relatively lower accuracy in the occipital lobe indicates that
brain regions such as the left and right hemispheres, and the it is not that effective in the case of classifying MDD groups
four lobes- frontal, parietal, occipital, and temporal. We use from other brain regions. The fundamental role of the
occipital lobe is to regulate vision and visual processing [60]. parameter C and the gamma, are adjusted during training to
So, the result suggests that MDD does not affect one’s vision enhance classification performance.
that much.
E. Comparison with Previous Work
B. Different EEG Rhythms in Depression Mumtaz et al. [33] achieved an accuracy of 90% by
Diverse features were derived from distinct frequency bands. combining wavelet-based characteristics with a Logistic
As depicted in Figure 4 and Figure 5, the delta and theta band Regression classifier. The authors of [60] utilized an SVM
comprises the majority of the best features among the classifier to reach an accuracy of almost 98 percent. In a
frequency bands which is similar to previous work [34]. different research [32], Khan et al. used the CNN classifier
They also found the best results using the delta band (0.5 Hz and achieved 100% accuracy. To get a more reliable
to 4 Hz). Uyulan et al. [61] built a ResNet-50 architecture outcome, we employed SVM in addition to tuning the
based CNN model with a delta band accuracy of 90.22%. hyperparameters. Even though CNN was able to attain 100%
The delta frequency band is predominantly detected in accuracy, the computational complexity is higher. With the
infants and adults during sleep. Children as young as 13 suggested framework, we only used a machine learning
years old and adults during sleep can both show theta band model that is especially suited for small datasets to obtain the
activity, which is classified as sluggish activity. This can be best results, whereas deep learning approaches need larger
observed in some cases of hydrocephalus as well as diffuse datasets to provide good results in real-world experiments.
disorders such as metabolic encephalopathy [62].
Koller-Schlaud et al. used the Variance model analysis and
F. Limitations
discovered that theta activity at the central electrode location
is highly distinct to identify HC and bipolar depressed Even though the classification accuracy was 100%, no
controls in the resting [63]. Since the data was collected in a specific set of features can be suggested for classification
resting state and both delta and theta band differed between since they were different for each region as well as each
those with MDD and HC, it can be inferred that the validation methods. So, for exploring all the features must be
depressed brain exhibits altered resting activity. The taken into account. Furthermore, the GC pairs that are
topographical plot in Figure 7 portrays the most visible significant in this particular dataset, might not be same for
differences in delta and theta bands. This also implies in other datasets. Mumtaz et al. [34] faced some issues during
support of the accuracy found from the study. data collection which can also contribute to getting different
result in case of applying the proposed method in other
datasets. The dataset also contained task related recordings
C. Granger Connectivity in Depression which were not used in this framework. The proposed
Granger connectivity matrices also performed good framework was implemented on an online dataset. It has not
classification results as feature set. As illustrated in Figure 6, been applied in practical environments such as in hospitals
the grand average of all the MDD and healthy Granger with real MDD patients. The results might vary depending
connections was determined for a comprehensive overview. on subjects and data collection protocol.
Both with eyes closed and with eyes open, MDD patients
had less causal activity than healthy people. This finding V. CONCLUSION
suggests that MDD patients' brain areas are less in harmony According to the findings, MDD affects the majority of the
with one another. The variation is more noticeable in eyes brain, and no specific lobe or hemisphere can be identified as
closed condition when the subjects have no distraction and a distinguishing trait between MDD and healthy patients.
are in a complete resting state. Figure 7 depicts the best The delta and theta bands play vital role in classification if
causal connections that helped achieve this good result. The the data is collected in resting state. As causes must occur
illustration shows visible difference in causal connectivity in before consequences in time, systematic perturbations of the
healthy and MDD subjects. From the causal analysis, a whole brain may be monitored using Granger causality. In
neural pathway can be determined and a pattern can be found this investigation, the effective connection was discovered
that can effectively distinguish MDD individuals from among many brain areas, resulting in a useful classification
healthy ones that can be treated as a biomarker for strategy for distinguishing healthy from depressed groups
identifying MDD persons. which can be a potential biomarker. Using the resulted
pattern, in the future, depression can be diagnosed with EEG
D. Experimental Evaluation data successfully and quickly. The future researchers may
In this experiment, SVM and the Gaussian Radial Basis deploy the framework using datasets from other EEG
kernel function are utilized for classification (RBF). SVM is resources with different protocols such as using stimuli.
a complex and adaptable machine learning model that can They can also analyze further to find a number of optimal
carry out regression, outlier detection, and even linear and channels for depression detection. Lastly, the proposed
nonlinear classification [55]. It is excellent at classifying method has been optimized for classifying depressed
intricate small- to medium-sized datasets. As demonstrated subjects. Future researchers can apply the same framework
most effectively and precisely by Hsu et al. in [58], the SVM to diagnose other psychological disorders.
classifier's hyper-parameters, in particular the regularization
 7

ACKNOWLEDGMENTS (MDD): a review,” Nanoelectron. Circuits Commun. Syst.,

This work was supported in part by Information and pp. 323–335, 2019.
Communication Technology (ICT) Division, Ministry of [13] D. Librenza-Garcia et al., “The impact of machine
Posts, Telecommunications and Information Technology, learning techniques in the study of bipolar disorder: a
Bangladesh, under Grant 56.00.0000.028.33.003.21-208; systematic review,” Neurosci. Biobehav. Rev., vol. 80, pp.
and in part by the Institute for Advanced Research (IAR), 538–554, 2017.
United International University (UIU), Bangladesh, under [14] U. R. Acharya et al., “A Novel Depression Diagnosis
Grant UIU/IAR/01/2021/SE/16. Index Using Nonlinear Features in EEG Signals,” Eur.
Neurol., vol. 74, no. 1–2, pp. 79–83, 2015, doi:
10.1159/000438457.
REFERENCES [15] S. Mahato and S. Paul, “Detection of major depressive
disorder using linear and non-linear features from EEG
[1] A. F. Dadi, E. R. Miller, T. A. Bisetegn, and L. Mwanri, signals,” Microsyst. Technol., vol. 25, no. 3, pp. 1065–1076,
“Global burden of antenatal depression and its association Mar. 2019, doi: 10.1007/s00542-018-4075-z.
with adverse birth outcomes: an umbrella review,” BMC [16] U. R. Acharya, S. L. Oh, Y. Hagiwara, J. H. Tan, H.
Public Health, vol. 20, p. 173, Feb. 2020, doi: Adeli, and D. P. Subha, “Automated EEG-based screening
10.1186/s12889-020-8293-9. of depression using deep convolutional neural network,”
[2] Z. Li, M. Ruan, J. Chen, and Y. Fang, “Major Comput. Methods Programs Biomed., vol. 161, pp. 103–113,
Depressive Disorder: Advances in Neuroscience Research Jul. 2018, doi: 10.1016/j.cmpb.2018.04.012.
and Translational Applications,” Neurosci. Bull., vol. 37, no. [17] X. Zhang et al., “EEG-based Depression Detection
6, pp. 863–880, Feb. 2021, doi: Using Convolutional Neural Network with Demographic
10.1007/s12264-021-00638-3. Attention Mechanism,” in 2020 42nd Annual International
[3] “Depression.” Conference of the IEEE Engineering in Medicine Biology
https://www.who.int/news-room/fact-sheets/detail/depressio Society (EMBC), Jul. 2020, pp. 128–133. doi:
n (accessed Aug. 20, 2022). 10.1109/EMBC44109.2020.9175956.
[4] “Depression (major depressive disorder) - Diagnosis [18] W. Mumtaz and A. Qayyum, “A deep learning
and treatment - Mayo Clinic.” framework for automatic diagnosis of unipolar depression,”
https://www.mayoclinic.org/diseases-conditions/depression/ Int. J. Med. Inf., vol. 132, p. 103983, Dec. 2019, doi:
diagnosis-treatment/drc-20356013 (accessed Jul. 02, 2022). 10.1016/j.ijmedinf.2019.103983.
[5] S. M. Eack, C. G. Greeno, and B.-J. Lee, “Limitations [19] L. Astolfi et al., “Comparison of different cortical
of the Patient Health Questionnaire in Identifying Anxiety connectivity estimators for high-resolution EEG
and Depression: Many Cases Are Undetected,” Res. Soc. recordings,” Hum. Brain Mapp., vol. 28, no. 2, pp. 143–157,
Work Pract., vol. 16, no. 6, pp. 625–631, Nov. 2006, doi: 2007, doi: 10.1002/hbm.20263.
10.1177/1049731506291582. [20] D. S. American Psychiatric Association and A. P.
[6] “We can now detect depression with an MRI-Health Association, Diagnostic and statistical manual of mental
News , Firstpost.” disorders: DSM-5, vol. 5. American psychiatric association
https://www.firstpost.com/health/we-can-now-detect-depres Washington, DC, 2013.
sion-with-an-mri-7744401.html (accessed May 28, 2022). [21] M. Hamedi, S.-H. Salleh, and A. M. Noor,
[7] “Benefits of EEG Testing – The NeuLine Clinic.” “Electroencephalographic Motor Imagery Brain
https://theneulineclinic.com/benefits-of-eeg-testing/ Connectivity Analysis for BCI: A Review,” Neural Comput.,
(accessed Jul. 02, 2022). vol. 28, no. 6, pp. 999–1041, Jun. 2016, doi:
[8] O. Sporns, “Brain connectivity,” Scholarpedia, vol. 2, 10.1162/NECO_a_00838.
no. 10, p. 4695, Oct. 2007, doi: 10.4249/scholarpedia.4695. [22] L. Barnett and A. K. Seth, “The MVGC multivariate
[9] B. Hosseinifard, M. H. Moradi, and R. Rostami, Granger causality toolbox: a new approach to
“Classifying depression patients and normal subjects using Granger-causal inference,” J. Neurosci. Methods, vol. 223,
machine learning techniques,” in 2011 19th Iranian pp. 50–68, Feb. 2014, doi: 10.1016/j.jneumeth.2013.10.018.
Conference on Electrical Engineering, May 2011, pp. 1–4. [23] A. K. Seth, A. B. Barrett, and L. Barnett, “Granger
[10] S. Mantri, D. Patil, P. Agrawal, and V. Wadhai, “Non Causality Analysis in Neuroscience and Neuroimaging,” J.
invasive EEG signal processing framework for real time Neurosci., vol. 35, no. 8, pp. 3293–3297, Feb. 2015, doi:
depression analysis,” in 2015 SAI Intelligent Systems 10.1523/JNEUROSCI.4399-14.2015.
Conference (IntelliSys), Nov. 2015, pp. 518–521. doi: [24] H. E. Wang, C. G. Bénar, P. P. Quilichini, K. J. Friston,
10.1109/IntelliSys.2015.7361188. V. K. Jirsa, and C. Bernard, “A systematic framework for
[11] M. Bachmann et al., “Methods for classifying functional connectivity measures,” Front. Neurosci., vol. 8,
depression in single channel EEG using linear and nonlinear 2014, Accessed: Jun. 02, 2022. [Online]. Available:
signal analysis,” Comput. Methods Programs Biomed., vol. https://www.frontiersin.org/article/10.3389/fnins.2014.0040
155, pp. 11–17, Mar. 2018, doi: 5
10.1016/j.cmpb.2017.11.023. [25] K. J. Friston, L. Harrison, and W. Penny, “Dynamic
[12] S. Mahato and S. Paul, “Electroencephalogram (EEG) causal modelling,” NeuroImage, vol. 19, no. 4, pp.
signal analysis for diagnosis of major depressive disorder 1273–1302, Aug. 2003, doi:
10.1016/S1053-8119(03)00202-7.
[26] J. Geweke, “Measurement of Linear Dependence and no. 1, pp. 48–63, May 2007, doi:
Feedback between Multiple Time Series,” J. Am. Stat. 10.1016/j.neuroimage.2007.02.034.
Assoc., vol. 77, no. 378, pp. 304–313, Jun. 1982, doi: [38] A. Delorme and S. Makeig, “EEGLAB: an open source
10.1080/01621459.1982.10477803. toolbox for analysis of single-trial EEG dynamics including
[27] A. E. Whitton, S. Deccy, M. L. Ironside, P. Kumar, M. independent component analysis,” J. Neurosci. Methods, vol.
Beltzer, and D. A. Pizzagalli, “Electroencephalography 134, no. 1, pp. 9–21, Mar. 2004, doi:
Source Functional Connectivity Reveals Abnormal 10.1016/j.jneumeth.2003.10.009.
High-Frequency Communication Among Large-Scale [39] C. A. Kothe, “The Artifact Subspace Reconstruction
Functional Networks in Depression,” Biol. Psychiatry Cogn. Method,” p. 23.
Neurosci. Neuroimaging, vol. 3, no. 1, pp. 50–58, Jan. 2018, [40] “Independent component analysis: algorithms and
doi: 10.1016/j.bpsc.2017.07.001. applications - ScienceDirect.”
[28] W. Mumtaz, S. S. A. Ali, M. A. M. Yasin, and A. S. https://www.sciencedirect.com/science/article/abs/pii/S089
Malik, “A machine learning framework involving 3608000000265 (accessed Jun. 04, 2022).
EEG-based functional connectivity to diagnose major [41] I. Winkler, S. Haufe, and M. Tangermann, “Automatic
depressive disorder (MDD),” Med. Biol. Eng. Comput., vol. Classification of Artifactual ICA-Components for Artifact
56, no. 2, pp. 233–246, Feb. 2018, doi: Removal in EEG Signals,” Behav. Brain Funct., vol. 7, no. 1,
10.1007/s11517-017-1685-z. p. 30, Aug. 2011, doi: 10.1186/1744-9081-7-30.
[29] N. Fogelson, P. Diaz-Brage, L. Li, A. Peled, and E. [42] J. Too, “Jx-EEGT : Electroencephalogram ( EEG )
Klein, “Functional connectivity abnormalities during Feature Extraction Toolbox.” May 31, 2022. Accessed: Jun.
processing of predictive stimuli in patients with major 04, 2022. [Online]. Available:
depressive disorder,” Brain Res., vol. 1727, p. 146543, Jan. https://github.com/JingweiToo/EEG-Feature-Extraction-To
2020, doi: 10.1016/j.brainres.2019.146543. olbox
[30] S. Olbrich, A. Tränkner, T. Chittka, U. Hegerl, and P. [43] J. Too, “Jx-EMGT : Electromyography ( EMG )
Schönknecht, “Functional connectivity in major depression: Feature Extraction Toolbox.” Jun. 23, 2022. Accessed: Jul.
Increased phase synchronization between frontal cortical 02, 2022. [Online]. Available:
EEG-source estimates,” Psychiatry Res. Neuroimaging, vol. https://github.com/JingweiToo/EMG-Feature-Extraction-To
222, no. 1, pp. 91–99, Apr. 2014, doi: olbox
10.1016/j.pscychresns.2014.02.010. [44] J. Too, A. R. Abdullah, N. Mohd Saad, and W. Tee,
[31] Y. Xie et al., “Anxiety and Depression Diagnosis “EMG Feature Selection and Classification Using a
Method Based on Brain Networks and Convolutional Neural Pbest-Guide Binary Particle Swarm Optimization,”
Networks,” in 2020 42nd Annual International Conference Computation, vol. 7, no. 1, Art. no. 1, Mar. 2019, doi:
of the IEEE Engineering in Medicine Biology Society 10.3390/computation7010012.
(EMBC), Jul. 2020, pp. 1503–1506. doi: [45] M. Rashid, N. Sulaiman, M. Mustafa, S. Khatun, and B.
10.1109/EMBC44109.2020.9176471. S. Bari, “The Classification of EEG Signal Using Different
[32] X. Li, R. La, Y. Wang, B. Hu, and X. Zhang, “A Deep Machine Learning Techniques for BCI Application,” in
Learning Approach for Mild Depression Recognition Based Robot Intelligence Technology and Applications, Singapore,
on Functional Connectivity Using Electroencephalography,” 2019, pp. 207–221. doi: 10.1007/978-981-13-7780-8_17.
Front. Neurosci., vol. 14, 2020, Accessed: May 28, 2022. [46] A. R. Hassan and A. Subasi, “Automatic identification
[Online]. Available: of epileptic seizures from EEG signals using linear
https://www.frontiersin.org/article/10.3389/fnins.2020.0019 programming boosting,” Comput. Methods Programs
2 Biomed., vol. 136, pp. 65–77, Nov. 2016, doi:
[33] D. M. Khan, N. Yahya, N. Kamel, and I. Faye, 10.1016/j.cmpb.2016.08.013.
“Automated Diagnosis of Major Depressive Disorder Using [47] S. Wang et al., “Revealing the dynamic causal
Brain Effective Connectivity and 3D Convolutional Neural interdependence between neural and muscular signals in
Network,” IEEE Access, vol. 9, pp. 8835–8846, 2021, doi: Parkinsonian tremor,” J. Frankl. Inst., vol. 344, no. 3, pp.
10.1109/ACCESS.2021.3049427. 180–195, May 2007, doi: 10.1016/j.jfranklin.2006.06.003.
[34] W. Mumtaz, L. Xia, M. A. M. Yasin, S. S. A. Ali, and A. [48] A. K. Seth, “A MATLAB toolbox for Granger causal
S. Malik, “A wavelet-based technique to predict treatment connectivity analysis,” J. Neurosci. Methods, vol. 186, no. 2,
outcome for Major Depressive Disorder,” PLOS ONE, vol. pp. 262–273, Feb. 2010, doi:
12, no. 2, p. e0171409, Feb. 2017, doi: 10.1016/j.jneumeth.2009.11.020.
10.1371/journal.pone.0171409. [49] M. K. Wali, M. Murugappan, and B. Ahmmad,
[35] “MDD Patients and HC EEG Data (New).” figshare, “Wavelet Packet Transform Based Driver Distraction Level
Nov. 23, 2016. doi: 10.6084/m9.figshare.4244171.v2. Classification Using EEG,” Math. Probl. Eng., vol. 2013, p.
[36] K. G. H, “The ten-twenty electrode system of the e297587, Nov. 2013, doi: 10.1155/2013/297587.
international federation. the internanional federation of [50] L. S. Vidyaratne and K. M. Iftekharuddin, “Real-Time
clinical nenrophysiology,” Electroencephalogr Clin Epileptic Seizure Detection Using EEG,” IEEE Trans.
Neurophysiol Suppl, vol. 52, pp. 3–6, 1999. Neural Syst. Rehabil. Eng., vol. 25, no. 11, pp. 2146–2156,
[37] L. Marzetti, G. Nolte, M. G. Perrucci, G. L. Romani, Nov. 2017, doi: 10.1109/TNSRE.2017.2697920.
and C. Del Gratta, “The use of standardized infinity [51] V. P. B and S. Chinara, “Automatic classification
reference in EEG coherency studies,” NeuroImage, vol. 36, methods for detecting drowsiness using wavelet packet
 9

transform extracted time-domain features from neighbor method and EEG signals,” Phys. Eng. Sci. Med.,
single-channel EEG signal,” J. Neurosci. Methods, vol. 347, vol. 43, no. 3, pp. 1007–1018, Sep. 2020, doi:
p. 108927, Jan. 2021, doi: 10.1016/j.jneumeth.2020.108927. 10.1007/s13246-020-00897-w.
[52] D. B. Percival and A. T. Walden, Wavelet methods for
time series analysis, vol. 4. Cambridge university press,
2000.
[53] K. A. Mamun et al., “Decoding movement and
laterality from local field potentials in the subthalamic
nucleus,” in 2011 5th International IEEE/EMBS Conference
on Neural Engineering, Apr. 2011, pp. 128–131. doi:
10.1109/NER.2011.5910505.
[54] K. Yan and D. Zhang, “Feature selection and analysis
on correlated gas sensor data with recursive feature
elimination,” Sens. Actuators B Chem., vol. 212, pp.
353–363, Jun. 2015, doi: 10.1016/j.snb.2015.02.025.
[55] E. Tzini, “Feature Selection: Wrapper Methods,”
Analytics Vidhya, May 20, 2022.
https://medium.com/analytics-vidhya/feature-selection-855
39d6a2a88 (accessed Jun. 05, 2022).
[56] A. Anuragi and D. Singh Sisodia, “Alcohol use disorder
detection using EEG Signal features and flexible analytical
wavelet transform,” Biomed. Signal Process. Control, vol.
52, pp. 384–393, Jul. 2019, doi: 10.1016/j.bspc.2018.10.017.
[57] S. Sreenivasa, “Radial Basis Function (RBF) Kernel:
The Go-To Kernel,” Medium, Oct. 12, 2020.
https://towardsdatascience.com/radial-basis-function-rbf-ke
rnel-the-go-to-kernel-acf0d22c798a (accessed Jun. 05,
2022).
[58] C.-W. Hsu, C.-C. Chang, and C.-J. Lin, A practical
guide to support vector classification. Taipei, 2003.
[59] C.-C. Chang and C.-J. Lin, “LIBSVM: a library for
support vector machines,” ACM Trans. Intell. Syst. Technol.
TIST, vol. 2, no. 3, pp. 1–27, 2011.
[60] “Function of the Occipital Lobe,” Bodytomy, Jul. 15,
2010. https://bodytomy.com/occipital-lobe-function
(accessed Jul. 02, 2022).
[61] C. Uyulan et al., “Major Depressive Disorder
Classification Based on Different Convolutional Neural
Network Models: Deep Learning Approach,” Clin. EEG
Neurosci., vol. 52, no. 1, pp. 38–51, Jan. 2021, doi:
10.1177/1550059420916634.
[62] “Increased Theta and Alpha EEG Activity During
Nondirective Meditation | The Journal of Alternative and
Complementary Medicine.”
https://www.liebertpub.com/doi/full/10.1089/acm.2009.011
3 (accessed Jul. 02, 2022).
[63] K. Koller-Schlaud, A. Ströhle, E. Bärwolf, J. Behr, and
J. Rentzsch, “EEG Frontal Asymmetry and Theta Power in
Unipolar and Bipolar Depression,” J. Affect. Disord., vol.
276, pp. 501–510, Nov. 2020, doi:
10.1016/j.jad.2020.07.011.
[64] M. Saeedi, A. Saeedi, and A. Maghsoudi, “Major
depressive disorder assessment via enhanced k-nearest
 .
Figure 1. Workflow of the proposed framework to diagnoses healthy and MMD individuals. Initially, we collect raw EEG signals in resting state. Then, the signal
is preprocessed by bandpass filtering, artifact rejection, and independent component analysis (ICA). After that, we extract three types of features from segmented
signals, namely, time domain, frequency domain, and granger causality. The best feature is selected from wrapper based Recursive Feature Elimination along with
support vector machine (SVM) classifier. Lastly, we classify the healthy and depressed individuals using SVM- radial basis function (RBF) classifier using
hyperparameter tuning.
 11

Figure 2 Electrode positioning according to the international 10–20 electrode placement standard of Brain Master Discover.
Figure 3 Effect of the hyperparameters (C and Gamma) following the number of features and classification accuracy while training the model on the full brain. The

color bar indicates the level of accuracy. With only the 10 best features the model got 100% accuracy.
 13

Figure 4 Best features in delta band that were common in both 10-fold cv and LOSO methods.
Figure 5 Best features in theta band that were common in both 10-fold cv and LOSO methods.
 15

Figure 6 Depiction of Granger causal matrices of healthy and MDD subjects in two conditions. The leftmost is the visual representation of the calculated pairwise
Granger causality matrix, the middle one is the matrix of the probability value (p-value) and finally, the last matrix represents the significance of each
connectivity. A. Healthy subject in eyes closed state. B. Healthy subject in eyes open state. C. MDD subject in eyes closed state. D. MDD subject in eyes open
state.
Figure 7. Brain topography of healthy and MDD subjects in different frequency bands. The delta and theta band shows the most significant differences. (A) Brain

topography of healthy subjects in different frequency bands. (B) Brain topography of MDD subjects in different frequency bands.
 17

a. Connectivity map of healthy subjects. b. Connectivity map of MDD subjects.

Figure 8 Granger connectivity map with significant electrode pairs that contributed effectively in classifying healthy and MDD subjects. The lower the p-value the

more significant the connectivity.

Table 1Classification results with the best features in different brain regions.
Brain Validation No. of Best Features Accuracy
Region Method Features
Full Brain 10-fold cv T(F3_fzc), δ(Fz_le, T4_hc, T6_dfa, Fz_ass, T3_re), β_1(Cz_card,
10 100 %
O2_n2d), γ(F7_skurt, C4_skew)
11 Granger Causal Connectivity 99.17%
LOSO T(Pz_hc,F8_sh),δ(Fz_mmav2,C3_hm,Fp1_re,Fz_emav),β_1(O2_m
10 100 %
yop,Cz_card),γ(F7_n2d,T4_ldamv)
Left 10-fold cv T(F7_am,P3_bpg,P3_am),δ(F3_ssi,F7_sEntropy,F3_myop,T3_te),θ
Hemisphere 16 (Fp1_sDecrease),α(T5_sKurt,F3_asm),β_1(F3_asm,O1_ldasdv),β_ 100 %
2(F7_dfa, F7_tm, F7_max), γ(F7_dfa)
LOSO T(F7_am,P3_bpg,P3_am),δ(F3_myop,F7_sEntropy,F3_ssi,T3_te),θ
15 (Fp1_sDecrease),α(T5_sKurt,F3_asm),β_1(F3_asm),β_2(F7_dfa,F 100 %
7_tm,F7_max),γ(F7_dfa)
Right 10-fold cv T(F4_am,P4_var),δ(T4_zc,F4_myop,T6_hc,T4_am,T6_re),θ(T6_sS
Hemisphere pread,F8_ldasdv,T6_sEntropy,T6_sFlatness,T6_sCrest,T6_sCentroi
20 100 %
d,T6_sRollOffPoint),α(C4_LE),β_1(O2_ssi),β_2(),γ(T4_fzc,T4_mt
e),O(P4_rpγα,P4_avgβ2α)
LOSO T(F4_am,P4_var),δ(T4_am,T6_hc,T6_re,T4_zc,F4_myop,T4_ssi),
θ(F8_ldasdv,T6_sSpread,T6_sFlatness,T6_sEntropy,T6_sCrest,T6_
22 100 %
sCentroid,T6_sRolloffPoint),α(C4_le,F4_lrssv),β_1(O2_ssi),γ(T4_
mte,T4_fzc),O(P4_avgβ2α,P4_rpγα)
Frontal Lobe 10-fold cv T(Fz_mte), δ(Fz_mfl, F7_sEntropy, Fp1_bpg), θ(Fp1_sFlatness,
12 F8_le, F3_dfa, Fp1_sSpread), α(Fz_sCentroid), β_1(F4_zc), 100 %
β_2(F3_dfa), γ(F7_ha)
LOSO T(Fp1_sFlatness,F3_dfa,Fz_mte),δ(Fz_mfl,Fp1_bpg,F7_sEntropy),
11 100 %
θ(F8_le,Fz_n2d),β_1(F4_zc),β_2(F3_dfa),γ(F7_ha)
Parietal 10-fold cv T(Pz_bpg,P3_1d,Pz_mte),δ(P4_te),θ(Pz_sCentroid,P4_ssi,P4_dfa,P
10 100 %
Lobe 4_mfl),β_2(P3_dfa),O(P4_rpβ2β1)
LOSO T(Pz_bpg,Pz_am,P4_kurt),δ(P4_te,P3_dfa),θ(P3_sSkew,P4_dfa,P3
14 _sDecrease,P3_dfa),α(P3_sFlatness),β_2(P3_dfa),O(P3_avgβ2θ,P4 100 %
_rpγ,Pz_avgγθ)
Temporal 10-fold cv T(T5_1d),δ(T4_myop,T6_re),θ(T6_sSpread,T6_sFlatness,T4_te,T6
13 100 %
Lobe _dfa,T6_var,T3_te),α(T5_var),O(T4_avgβ2α,T6_rpθ,T4_rpγα)
LOSO T(T5_1d,T6_dfa),δ(T4_myop,T6_re),θ(T6_sSpread,T6_sFlatness,T
13 100 %
4_te,T6_var,T3_te),α(T5_var),O(T4_avgβ2α,T6_rpθ,T4_rpγα)
Occipital 10-fold cv δ(O2_sKurt,O2_le,O2_damv),θ(O1_sDecrease,O2_dfa,O1_hc,O1_
Lobe 18 sSkew,O1_dfa),α(O1_ha,O1_lrssv),β_1(O2_kurt,O2_le,O2_mfl,O1 99.167 %
_max),β_2(O2_1d),γ(O2_mfl),O(avg_β2θ,rp_β1α)
LOSO δ(O2_sKurt,O2_le,O2_damv),θ(O1_sDecrease,O2_dfa,O1_hc,O1_
18 sSkew,O1_dfa),α(O1_ha,O1_lrssv),β_1(O2_kurt,O2_le,O2_mfl,O1 99.167 %
_max),β_2(O2_1d),γ(O2_mfl),O(O1_avgβ2θ,O1_rpβ1α)
 19

Table 2Comparison with previous works.

No of Classification No of Subjects
Ref Year Features Accuracy
Channel Method MDD Healthy

Mumtaz et Logistic
2017 19 Wavelets+STFT+EMD 34 30 90.5% (±8.3)
al.[34] Regression

Saeedi et EKNN, SVM,

2020 19 Band power, WPD, 34 30 98.44%
al.[64] MLP

Khan et Effective Connectivity

2021 19 3D CNN 30 30 100%
al.[33] (PDC)-DMN

Wavelet Packet Transform 100%

Proposed 19 Granger Causal SVM 34 30

99.167%
Connectivity