IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL.

31, 2023 4725

Depression Identification Using EEG Signals via

a Hybrid of LSTM and Spiking Neural Networks
A. Sam , R. Boostani , S. Hashempour , M. Taghavi, and S. Sanei , Senior Member, IEEE

Abstract— Depression severity can be classified into I. I NTRODUCTION

distinct phases based on the Beck depression inventory
EPRESSION is a prevalent and serious mental disorder
(BDI) test scores, a subjective questionnaire. However,
quantitative assessment of depression may be attained
through the examination and categorization of electroen-
D affecting 280 million people worldwide [1], which
extensively influences an individual’s quality of life. This
cephalography (EEG) signals. Spiking neural networks significant public health concern impacts an individual’s
(SNNs), as the third generation of neural networks, incor- physical and mental welfare in various aspects, such as
porate biologically realistic algorithms, making them ideal
for mimicking internal brain activities while processing
alterations in appetite, diminished motivation and interest,
EEG signals. This study introduces a novel framework irregular sleep patterns, and in severe instances, contemplation
that for the first time, combines an SNN architecture of suicide. Early diagnosis of depression and treatment
and a long short-term memory (LSTM) structure to can prevent patients’ conditions from worsening [2]. The
model the brain’s underlying structures during different revised Beck depression inventory (BDI-II) stands out as one
stages of depression and effectively classify individual
of the extensively employed psychometric assessments for
depression levels using raw EEG signals. By employing
a brain-inspired SNN model, our research provides fresh quantifying the extent of depression [3]. Comprising a 21-
perspectives and advances knowledge of the neurological question multiple-choice self-report inventory, this assessment
mechanisms underlying different levels of depression. The method examines characteristic attitudes and symptoms
methodology employed in this study includes the utilization associated with depression. This test’s overall score, which
of the synaptic time dependent plasticity (STDP) learning ranges from 0 to 63, can classify the severity of depression
rule within a 3-dimensional brain-template structured
SNN model. Furthermore, it encompasses the tasks of into four groups [4]. It’s important to acknowledge that the
classifying and predicting individual outcomes, visually BDI-II test lacks a robust physiological underpinning and
representing the structural alterations in the brain linked is qualitative in essence. Given that depression influences
to the anticipated outcomes, and offering interpretations the neurotransmitter release within the human brain, it is
of the findings. Notably, our method achieves exceptional reasonable to hypothesize that it also impacts the electrical
accuracy in classification, with average rates of 98% and
96% for eyes-closed and eyes-open states, respectively.
neuronal activity captured through electroencephalography
These results significantly outperform state-of-the-art deep (EEG). EEG captures rich temporal data and offers reasonable
learning methods. spatial resolution, particularly when recorded with a larger
number of electrodes, such as 64, although it doesn’t match the
Index Terms— Beck depression inventory, LSTM, spiking
neural network, synaptic time dependent plasticity. precision of MRI images [5]. This brain activity-related dataset
is amenable to evaluation and interpretation through diverse
machine learning methodologies. Analysis of EEG is used for
the diagnosis of different neuropsychiatric disorders such as
Manuscript received 19 August 2023; revised 4 November 2023; Schizophrenia [6], [7], [8], Alzheimer’s [9], ADHD [10], [11],
accepted 21 November 2023. Date of publication 23 November 2023; dementia [12], brain fatigue [13], [14], sleep disorders [15],
date of current version 7 December 2023. (Corresponding author: [16], bipolar manic depression (BMD) [17], and Seizure [18].
A. Sam.)
This work involved human subjects or animals in its research. Approval To differentiate between individuals with depression and
of all ethical and experimental procedures and protocols was granted by those without, some studies first extract characteristics from
the University of Arizona’s ethics review process. the raw data and then input these features into machine
A. Sam, R. Boostani, and S. Hashempour are with the CSE&IT
Department of Electrical and Computer Engineering, Shiraz learning (ML) and artificial neural networks (ANN). For
University, Shiraz 71454, Iran (e-mail: [email protected]; instance, in [19], authors extracted Higuchi’s fractal dimension
[email protected]; [email protected]). (HFD) and sample entropy (SampEn) features, applying them
M. Taghavi is with the Department of Psychiatry, School of Medicine,
Islamic Azad University, Kazerun Branch, Kazerun 7319866451, Iran to seven machine learning algorithms, including Multilayer
(e-mail: [email protected]). Perceptron and Logistic Regression. Their reported average
S. Sanei is with the Department of Electrical and Electronic classification accuracy was 93.5% for distinguishing depressed
Engineering, Imperial College London, SW7 2BT London, U.K. (e-mail:
[email protected]). and healthy individuals. In another study conducted by
Digital Object Identifier 10.1109/TNSRE.2023.3336467 Puthankattil and Joseph [20] several EEG features like relative

© 2023 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
For more information, see https://creativecommons.org/licenses/by/4.0/
4726 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 31, 2023

wavelet energy (RWE) and sample entropy were extracted SNNs, as computational models, encompass spiking neurons
and given to a two-layer feedforward ANN, achieving a as processing components, interconnected by biologically
classification accuracy rate of 98.11%. Acharya et al. [21] feasible learning algorithms [32], [33], [34]. SNNs inspired by
extracted a number of nonlinear features, such as detrended the brain have found utility across diverse domains, including
fluctuation analysis (DFA), fractal dimension, higher order but not limited to forecasting [35], simulation of the impact
spectra (HOS), Hurst’s exponent (HE), largest Lyapunov of mindfulness on individuals with depression [36], real-
exponent (LLE), recurrence quantification analysis (RQA), and world data classification, image recognition, odor recognition,
SampEn and fed them into five different classifiers. Using motor control, trajectory tracking, and more. In 2014,
a support vector machine (SVM) classifier, they reported Kasabov et al. introduced an SNN architecture called Neucube
an average accuracy of 98%. Raw EEG signals have been [37], designed to facilitate effective learning, modeling,
widely employed for classification using Artificial Neural and classification of spatiotemporal brain data (STBD).
Networks. For example, Acharya et al. [22] applied raw EEG Shah et al. [38] employed the SNN Neucube architecture to
signals to a 13-layer convolutional neural network (CNN) model and visualize brain activity in individuals displaying
and obtained classification accuracies of 93.5% and 96% over symptoms of depression. They utilized the dynamic evolving
the left and right hemispheres, respectively. Xia et al. [23] spiking neural network method (deSNN) for classification,
proposed an end-to-end integrated deep-learning model for achieving an accuracy of 68.18% for eyes-open state and
classifying major depressive disorder (MDD) patients and 72.13% for eyes-closed state. Despite the utilization of
healthy controls using raw EEG data. They achieved 91.06% brain-inspired SNNs for diverse spatiotemporal brain data
average accuracy. In another approach, Hashempour et al. [24] (STBD) modeling applications, a proficient supervised model
introduced a hybrid convolutional and temporal-convolutional for classifying Neucube’s output results has yet to be
neural network (CNN-TCN) to estimate the BDI score from introduced.
raw EEG signals in a continuous manner. Their method Despite growing interest in processing of EEG patterns for
achieved a mean squared error (MSE) of 5.64±1.6 and mean assessing depression, not many studies have scored the degree
absolute error (MAE) of 1.73±0.27 for the eyes-open state, of depression using EEG and more importantly, identifying
as well as an MSE of 9.53±2.94 and MAE of 2.32±0.35 for the brain mechanisms associated with different degrees
the eyes-closed state. of depression. However, such a research motivates further
Moreover, in [25], researchers introduced “DeprNet,” a deep development of this field by addressing these critical aspects.
learning-based CNN, achieving an accuracy of 0.91% and an To achieve this goal, we have utilized an extensive dataset
AUC of 0.95 in the classification of EEG data from both comprising EEG signals from 119 participants who underwent
depressed and normal subjects. Notably, their analysis of the the Beck test and were stratified into four depression levels:
final CNN layer visualization revealed that right electrodes minimum, mild, moderate, and severe. To estimate the
had higher prominence in depressed subjects, while left depression level, we present a novel methodology, combining a
electrodes exhibited greater prominence in normal subjects. brain-inspired SNN architecture with an LSTM neural network
In [26], an innovative automatic feature extraction method to model, visualize, learn, compare, and classify the subjects’
was employed using the Node2vec framework. This approach EEG signals. To compare the results of our method, we have
offered three fusion strategies: graph-level, feature-level, and also applied the raw EEG to a CNN-TCN, a CNN-LSTM, and
decision-level fusion, with a peak accuracy of 93.3% attained a 13-layer CNN network.
in the decision-level fusion process. In another study [27], The organization of this paper is outlined as follows:
a novel technique for extracting features from EEG signal Section II provides an overview of the dataset employed
channels has been developed. These features were integrated in this study, followed by a concise introduction to SNNs.
using a fuzzy ensemble strategy. In their method, they used Subsequently, we introduce a hybrid network that merges an
K-Nearest Neighbor classifier, which delivered the highest SNN architecture with an LSTM model. Section III presents
classification accuracy among the three datasets, with accuracy the visualization of the simulated network, accompanied by
scores of 91%, 96%, and 94%. Furthermore, [28] introduced a an analysis of the underlying brain structures associated with
dataset and employed traditional supervised machine learning depression. We present the empirical results, compare them to
algorithms to differentiate between healthy subjects and those state-of-the-art methods, and evaluate their respective strengths
with depression. Notably, the XGBoost classifier demonstrated and limitations. Finally, in Section IV, we conclude our study
the best performance, achieving an 87% accuracy rate for and outline avenues for future research.
the eyes-open (EO) state. These studies collectively exemplify
inventive methodologies and robust classification accuracy in
II. M ATERIALS AND M ETHODS
the domain of depression detection.
Although the mentioned methodologies have achieved A. Dataset Description and Preparation
commendable classification accuracy, none of them offered 1) Participants: We utilized an openly available dataset
model interpretations that could facilitate the identification from the PRED+CT website [39], initially comprising EEG
and comprehension of the brain mechanisms linked to recordings from 121 participants, including 72 females and
depression. While deep learning techniques draw inspiration 49 males. Subsequently, two subjects with incomplete practical
from certain observed properties in brain research [29], [30], information were identified and excluded from the analysis.
the latest generation of ANNs, called spiking neural networks All participants granted written informed consent, a protocol
(SNN) exhibits a greater degree of biological realism [31]. duly sanctioned by the University of Arizona’s ethics
SAM et al.: DEPRESSION IDENTIFICATION USING EEG SIGNALS VIA A HYBRID OF LSTM AND SNNs 4727

review process. The recruitment process involved enrolling mixtures and ICA for decomposition. It is important to note
individuals from introductory psychology courses, with the that HEOG and VEOG channels are initially dropped from the
selection based on their scores in the BDI mass survey. The analysis as they are not used for artifact removal. The artifacts,
eligibility criteria encompassed factors such as (a) age range which mainly include eye artifacts (such as blinks and eye
of 18 to 25 years, (b) absence of any history of head trauma movements), muscle artifacts, heart artifacts (ECG), and other
or seizures, and (c) no ongoing utilization of psychoactive non-neural artifacts, are eliminated using FastICA. The final
medications [40]. The recruited subjects exhibited diverse step involves back-projecting the remaining ICA components
levels of depression. Among the enrolled participants, 76 had into the channel space.
a Beck score ranging from 0 to 13, classifying them In the featured dataset, the training samples are not equally
into the control group (minimum depression). Furthermore, distributed across the target classes. Therefore, we employ
14 participants received scores ranging from 14 to 19 an undersampling technique in order to prevent the model
(indicating mild depression), 24 subjects attained scores from being biased toward the class that has a larger number
between 20 and 28 (reflecting moderate depression), and of training cases, which would reduce the model’s predictive
5 individuals scored within the range of 29 to 63 (indicative ability. To achieve this, an initial selection involves opting
of severe depression) [41]. for two minutes (equivalent to 30000 samples) of the EEG
2) Data Acquisition and Preprocessing: The dataset com- signal from both the eyes-open and eyes-closed states for
prised 500 seconds of recorded signals, acquired through each subject. Secondly, the signals are divided into five-
64 channels along with two additional channels, HEOG second windows (1250 samples), with each window having
and VEOG, following electrode settings aligned with the a 90% overlap. The 5-second window size has been chosen
10-20 standard EEG recording system [42]. The signals to accommodate SNNs that learn from spike occurrences.
were recorded during a resting state, utilizing a sampling In the absence of specific cognitive tasks during data
frequency of 500 Hz. The last two channels along with acquisition, this extended window supports more effective
the ‘CB1’ and ‘CB2’ channels are dropped, and ‘M1’ and unsupervised learning through spike-time dependent plasticity
‘M2’ are set as reference channels. This results in having (STDP) within the SNN reservoir. It allows for the capturing
a total number of 62 proper scalp channels. The recording of subtle temporal patterns and enhances the modeling of
paradigm encompassed events of both eyes-open and eyes- spatiotemporal EEG patterns, aligning with the network’s
closed conditions, exhibiting varying durations for different spike-driven processing. The data is then balanced across all
individuals. Consequently, the EEG data has been segregated depression levels based on the number of individuals. As a
into two distinct datasets: eyes-open and eyes-closed resting result, within each window, the data point count for depression
states. classes is normalized to align with the count of the class
In the initial stage, the EEG signal of each individual is possessing the smallest data point size. A data matrix with
partitioned into distinct event points. There are a total of the dimensions (4554, 1250, 62) and float type values is the
12 unique events within the signal. Due to variations in the result of this windowing operation.
number of occurrences for each event across different subjects,
differing quantities of segments are generated for each unique
event. In pursuit of dataset balance, we homogenize the B. Spiking Neural Networks
segment count for each distinct event to align with the In accordance with the all-or-none principle, information
minimum segment count of 120. The raw EEG signals within human brains is encoded through distinct events
are first preprocessed before being fed to the model. referred to as action potentials or spikes. A neuron generates a
To accomplish this, our methodology involves several steps. spike when its cumulative potential surpasses a predetermined
In the preliminary phase, we apply a downsampling of EEG threshold; otherwise, it continues to be inactive. Information
signals by a factor of two to reduce the data volume without regarding external stimuli and other internal computations
significantly violating the Nyquist rate. Subsequently, the is carried by the timing of spiking, the neuron’s location,
signal baselines are eliminated. Following this, a 50 Hz notch the neurons’ firing rate, and the temporal patterns. Due to
filter is employed, as outlined in [43] and [44], to counteract its binary information processing capability, SNN maintains
the power line interference. The signals then undergo bandpass its advantage in terms of energy efficiency and effectiveness
filtering, with cutoff frequencies set at 0.2 Hz and 50 Hz. over conventional ANNs [46], [47]. Incorporating a more
Lastly, the signals are processed using a Butterworth filter biologically realistic neuron model compared to traditional
with a fifth-order configuration, incorporating a high-cut ANNs [48], SNNs, as the third generation of neural network
at 50 Hz and a low-cut at 1 Hz. The filtered EEGs models, uniquely mimic the intricate mechanisms of the
are passed through independent component analysis (ICA) brain’s neurons. This inherent similarity to the brain’s neuron
in the last stage to remove any remaining undesirable mechanisms makes SNNs particularly well-suited for the
components. This study utilizes the MNE-python software analysis and modeling of EEG data. Operating across a
[45] to mitigate data contamination, primarily through a semi- multitude of spiking neurons, this model effectively processes
automated independent component analysis (ICA) approach. dynamic input information. The leaky-integrate and fire
In this context, we employ FastICA due to its notable model (LIF), a representation of a spiking neuron, can be
speed advantages over traditional ICA methods and its employed to emulate each neuron within the SNN model [49].
capability to accommodate non-Gaussianity. The procedure Notably, temporal dynamics are integrated into the operations,
involves principal component analysis (PCA) for whitening the alongside the synaptic states of the neurons. This temporal
4728 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 31, 2023

tuned to address the specific requirements of the task at

hand, particularly focusing on enhancing the classification
component.

C. The Proposed SNN Architecture in Combination With

an LSTM Network
Our proposed SNN architecture functions as a spa-
tiotemporal machine, employing a brain-inspired spiking
neural network design. Its overarching objectives encompass
knowledge extraction, STBD learning modeling, and inves-
Fig. 1. LIF neuron model. tigation into the neurological mechanisms underpinning data
generation [52], [53]. In our proposed approach, the 3D SNNr
consideration aligns well with scenarios where the timing of module is merged with an LSTM network, enhancing the
input signals is the main concern [50]. Consequently, SNNs comprehension and classification of depression. In Fig. 2,
emerge as an apt approach for applications involving STBD we present a diagram illustrating the consecutive steps of
analysis, including EEG and fMRI [51]. our proposed method. The subsequent steps pertain to the
LIF neurons represent the predominant neuronal model modeling phase.
employed within SNNs. It describes the behavior of a neuron 1) Spike Encoding: SNN-based architecture processes infor-
as it integrates incoming signals and, when a certain membrane mation through binary spiking events. Accordingly, the initial
potential threshold is reached, fires an action potential. The step is to encode all continuous variables into spike trains.
LIF neuron model can be mathematically described by the In this context, our focus lies specifically on temporal
following equation: spike encoding techniques, wherein spike timings signify
alterations in the signal’s value over time [54]. This strategy
d V (t)
τ = −[V (t) − Vrest ] + R I (t) (1) is motivated by the biologically tenable hypothesis that
dt information is encoded by accurate relative spike timing [55].
where V (t) is the membrane potential of the neuron at time The majority of widely used encoding algorithms [56] revolve
t, τ is the time constant of the neuron’s membrane,Vrest is the around monitoring temporal signal changes, subsequently
resting membrane potential, R is the membrane resistance, and represented through the exact timing of spikes. Examples
I (t) is the input current. When the membrane potential V (t) include threshold-based representation (TBR) algorithm, step-
crosses a predefined threshold, the neuron fires, resetting V (t). forward (SF) encoding, moving-window (MW) encoding, and
This model provides a basic representation of how neurons the Bens spiker algorithm (BSA) [57].
integrate and transmit signals in the brain. An illustration of In this research, we employ the address event representation
LIF neuron’s function is shown in Fig. 1. (AER) approach, a simplified adaptation of the TBR technique,
The STDP rule [64] is a fundamental learning mechanism to transform EEG data into spike trains [58], [59]. This
in SNNs. It governs the adjustment of synaptic weights approach proves particularly effective for data streams, as is
based on the precise timing of spikes between pre-synaptic the case with EEG signals. It hinges on the principle of
and post-synaptic neurons. STDP is inspired by biological thresholding the rate of alteration in an input variable over time
mechanism of synaptic plasticity, where the strength of [60]. Notably, each of the 62 input data channels is assigned an
synaptic connections is modified in response to the timing individualized variable threshold value, forming the core of the
of neuronal spikes. The STDP rule can be mathematically algorithm. Acknowledging the potential divergence in signal
represented as follows: dynamics and value ranges across input channels, a distinctive
variable threshold array is computed for each channel in the
− τ 1t

LTP · e LTP , if 1t > 0
A
following way:
1w = 1t (2)
 −A
LTD· e LTD , if 1t < 0
τ
x ′ ( j, k, i) = |X ( j + 1, k, i) − X ( j, k, i)| (3)
T −1
where 1w represents the change in synaptic weight, 1t is 1 X
the time interval between the pre-synaptic and post-synaptic µ(k, i) = x ′ ( j, k, i) (4)
T
spikes, ALTP and ALTD are positive constants that control j=1
sP
the magnitude of long-term potentiation (LTP) and long-term T −1
j=1 (x ( j, k, i) − µ(k, i))
′ 2
depression (LTD), respectively, and τLTP and τLTD are time σ (k, i) = (5)
constants that determine the rate of weight changes during T −2
N
LTP and LTD phases. STDP is a crucial learning mechanism 1 X
in SNNs, allowing them to adapt their synaptic connections V T (k) = (µ(k, i) + σ (k, i)) (6)
N
based on the temporal order of spikes, which is essential for i=1

various cognitive and computational tasks. where k ranges from 1 to the number of channels Nchannels =
The depression identification method proposed in this study 62, T is the signal length, and N represents the number of
is built upon a customized and improved version of the samples. X is a (T × Nchannels × N ) data matrix and VT
NeuCube framework [37]. This adaptation has been finely represents the resulting variable threshold array. Upon the
SAM et al.: DEPRESSION IDENTIFICATION USING EEG SIGNALS VIA A HYBRID OF LSTM AND SNNs 4729

Fig. 2. The architecture of the proposed method comprises three primary stages: EEG data encoding into spike trains, mapping to a 3-dimensional
brain-inspired SNN reservoir with 1471 neurons, and a two-step learning process for EEG dataset–unsupervised and supervised, ultimately leading
to classification.

occurrence of a rate of signal change surpassing the specific Algorithm 1 AER Spike Encoding
variable threshold in the kth input channel, a positive spike Require: X input ∈ RT ×Nchannels
is emitted. Conversely, when the rate of change breaches the Ensure: X Spike ∈ {0, 1}T ×Nchannels
variable threshold in a descending, or negative, direction, this 1: N ← #(X input ) ▷ Number of data samples in the dataset
triggers the generation of negative spikes. Algorithm 1 offers 2: for k = 1 to Nchannels do
a detailed exposition of this process, meticulously outlining 3: V T (k) ← 0
each step. 4: for i = 1 to N do
5: x ← channel k of the ith sample in X input
2) The 3D SNNr and Input Mapping: The 3D SNN 6: 1x T ×1 ← |δx|
reservoir (SNNr) module essentially constitutes an assemblage 7: µ ← mean(1x T ×1 )
of spiking neurons positioned spatially, with well-defined 8: σ ← st.dev.(1x T ×1 )
coordinates for the input neurons. This structure is designed 9: V Tk ← V Tk + (µ + σ )
to mimic the configuration of neurons within the brain. Here, 10: end for
we have implemented an SNNr with Nr eser voir = 1471 leaky 11: V Tk ← V Tk /N
12: end for
integrate and fire (LIF) model. These neurons are situated
13: for k = 1 to Nchannels do
in accordance with the Talairach Atlas [61], [62], forming 14: for i = 1 to N do
a cuboid shape resembling the human brain. Each neuron 15: x ← channel k of the ith sample in X input
represents 1 cm 3 brain area. The number of channels resulting 16: 1x T ×1 ← δx
from the loaded dataset, which in this case is Nchannels = 62, 17: xspike ← 0T ×1
defines the number of input neurons. The coordinates of these 18: for j = 2 to T do
input neurons are a subset of the SNNr coordinates. Using 19: if 1x j > V Tk then
the Koessler et al. mapping method [63], the nearest neuron 20: xspike ( j) ← 1
in the Talairach Atlas is allocated to the associated channel 21: end if
based on measurements of electrode placements, as shown in 22: end for
Fig. 3. Through the input neurons, the spike trains acquired 23: Store xspike in X Spike for channel k, sample i
following data encoding with the AER method are fed into 24: end for
25: end for
the SNNr.
An Nr eser voir × Nr eser voir matrix of distances called L dist
is created where an L2 norm is computed to determine
the distances between pairs of neurons. The “small world” connectivity (SWC) involves introducing a parameter for
connectedness tenet was selected based on the biological linking neurons within a defined range, referred to as the
process. Neighboring neurons become potentially coupled small-world radius (SWR). At the outset, all connections C(i j)
to one another as a result. The technique of small-world among the neurons within the entire reservoir are initialized to
4730 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 31, 2023

Fig. 3. Utilizing the Talairach template coordinates, brain’s three-dimensional coordinates are assigned to the designated spiking neurons within
the SNNr. This procedure yields a three-dimensional SNNr structure that mimics the shape of the brain. Notably, red neurons (right) are designated
as input neurons and correspondingly map to the positions of 62 EEG channels (left).

1. If L dist (i j) > SW R, the connection status between the two Algorithm 2 SNNr Connection and Weight Initialization
neurons is marked as zero (disconnected). Each connection Require: X brain ∈ R1×3 , X input ⊂ X brain , C : 1 Nreservoir ×Nreservoir
between neurons i and j signifies i as the pre-synaptic neuron ▷ Hyperparameters: SW R
and j as the post-synaptic neuron. When a connection is Ensure: C : {0, 1} Nreservoir ×Nreservoir , W : R Nreservoir ×Nreservoir
deemed bidirectional, we randomly assign a value of 1 to one 1: Nr ← #(X brain ) ▷ Number of reservoir neurons
connection and a value of 0 to the other, thereby preserving 2: L dist is a matrix of distances between all pairs of neurons
only one of the two options. As a result, an SNNr with 3: for i = 1toNreservoir do
sparsely connected neurons is created. Our model’s SNN 4: for j = 1toNreservoir do
initialization is carried out by using the SWC connection 5: if L dist (i, j) > SW R then
rule with SW R = 2.5. After initialization, the weights of 6: Ci j ← 0
connections Wi j between the connected pairs of neurons (ij) 7: else
are established using the subsequent equation: 8: Wi j ← sgn(rand − 0.3) · rand · L dist1(i, j)
1 9: end if
Wi j = sgn (rand(1) − 0.3) . rand(1) . (7) 10: if Ci j = 1andC ji = 1 then
L dist (i j)
11: if rand < 0.5 then
where rand(1) generates pseudorandom values drawn from the 12: Ci j ← 0
standard uniform distribution within the open interval of 0 to 1. 13: Wi j ← 0
According to (7), the matrix W is expected to contain around 14: else
70% positive and 30% negative weights. In Algorithm 2, 15: C ji ← 0
we present an in-depth account of this procedure, leaving no 16: W ji ← 0
room for ambiguity. 17: end if
3) Unsupervised Learning in SNNr and Visualization: This 18: end if
methodology divides the learning process into two phases: 19: end for
unsupervised learning and supervised learning. The process 20: end for
of unsupervised learning is employed to adapt the initial
connection weights of the SNNr model as the model learns
from the continuous EEG data presented in the form of presynaptic neuron fires and neuron j is not in the refractory
spikes. The STDP rule, an unsupervised learning technique time:
with biological plausibility, is employed for this learning
process. The STDP mechanism regulates the synaptic strength P j (t) = P j (t − 1) + Wi j (8)
based on the temporal relationship between presynaptic and
During any given moment t, if a neuron’s potential surpasses
postsynaptic action potentials. This algorithm operates with
the firing threshold potential β, the neuron fires, leading to its
the utilization of the subsequent parameters:
potential Pk (t) being reset to 0. Concurrently, its refractory
• N iter : number of training iterations
counter Rk is established at R (refractory time). Conversely,
• β: spike generation threshold
if neuron k fails to achieve the firing threshold potential at
• η: learning rate
time t, its potential is diminished by the leak rate D, and its
• R: the resting period between spikes
refractory counter is updated:
• D: the leakage rate of neurons’ potential while inactive
Nchannels spike states are sent to the associated input Pk (t) = Pk (t − 1) − D (9)
neurons within the SNNr at each time step and potential Rk = Rk − 1 (10)
propagations are computed. Take the (i, j) neuron pair as
an example, where i represents the pre-synaptic neuron and Based on the STDP model grounded in the Hebbian learning
j represents the post-synaptic neuron. In the event when a rule, an increase in the connection weight between two neurons
SAM et al.: DEPRESSION IDENTIFICATION USING EEG SIGNALS VIA A HYBRID OF LSTM AND SNNs 4731

TABLE I Algorithm 3 unsupervised SNNr Weight Learning: STDP

T HE STDP M ODEL PARAMETERS AND S ETTINGS Require: Wr ∈ R Nr ×Nr , X Spike ∈ {0, 1}T ×Nchannels
▷ Hyperparameters: D, R, η, β, Niter
Ensure: Wout ∈ R Nr ×Nr
1: N ← #(X Spike ) ▷ number of data samples in the dataset
2: κ ← [1, 2, . . . , Nr ] ▷ neuron indices
3: Pk , Rk ← 0
4: Find input neuron indices ι ← κ
5: for n iter = 1toNiter do
occurs when a presynaptic neuron fires immediately before a 6: η′ = √nη
iter
postsynaptic neuron, and conversely. This study employs a 7: for i = 1toN do
modified version of the STDP model. 8: x ← all Nchannels spikes of the ith sample in X Spike
Following the general STDP principle, the modification of 9: for t = 1toT do
the connection weight occurs when a presynaptic neuron i fires 10: α = {firing neurons in κ} ∪ {k ∈ κ\ι Pk > β}
at time t, and the postsynaptic neuron j has most recently fired 11: for all j ∈ α do
f
at t j : 12: Find post-synaptic neuron indices λ
13: for all k ∈ λ and Rk = 0 do
η 14: Pk ← Pk + w jk
Wi j = Wi j − f
(11)
(t − t j + 1) 15: end for
16: end for
On the other hand, if a post-synaptic neuron j fires at time 17: Pk ← max(0, Pk − D)
f
t and a pre-synaptic neuron i has most recently fired at ti : 18: Rk ← max(0, Rk − 1)
η 19: Pk ← 0, ∀k ∈ α
Wi j = Wi j + f
(12) 20: Rk ← R, ∀k ∈ α
(t − ti + 1) 21: for all j ∈ α do
At the start of each subsequent training cycle niter (of Niter ), 22: Find post-synaptic neuron indices λ
The learning rate η is adjusted to √nη . It is essential to 23: for all k ∈ λ do
f
iter
emphasize that all time instances involved in the learning 24: w jk ← w jk − η′ (t − tk )
algorithm are discrete. The parameter settings for the STDP 25: end for
method, as presented in Table I, were determined by grid 26: Find pre-synaptic neuron indices λ
search. We established a parameter range guided by [65], 27: for all k ∈ λ do
f
aligning with the physiological ranges of actual neurons. Grid 28: w jk ← w jk + η′ (t − tk )
search was then employed to systematically investigate various 29: end for
parameter combinations, evaluating network performance. 30: end for
As network performance metrics, we employed statistical 31: end for
metrics, including mean, standard deviation, and coefficient of 32: end for
variation, to gauge synaptic weight and spiking rate balance 33: end for
during grid search. These specific values have been selected
as they yielded the most favorable results. After the training
of SNNr is done, an output matrix with boolean type values 4) Supervised Learning Using LSTM: Long short-term
and a size of (4554, 1250, 1471) is obtained which is used as memory (LSTM) networks, which were introduced by
input for the next part of our model. The output data shows Hochreiter and Schmidhuber in 1997 [66], have demonstrated
that the number of time steps is the same as the input EEG their effectiveness in analyzing and interpreting EEG signals.
data. Nevertheless, the number of channels has changed from These networks excel at capturing the temporal dependencies
62 electrodes to 1471 neurons. Every neuron acts just like in EEG signals, effectively modeling both short-term and long-
an electrode in this situation. Consequently, this data might term patterns. By employing LSTM layers to process the
be regarded as advanced EEG data. Algorithm 3 provides a sequential EEG data, the model can capture the temporal
comprehensive breakdown of this procedure, elucidating the dynamics within each channel and the interdependencies
intricacies step by step. among different channels. This capability allows LSTMs
Through unsupervised STDP learning, the unique spike to leverage the spatial patterns and relationships in EEG
trains in EEG STBD data are transformed into connections signals, resulting in improved classification performance by
between neurons. These connections effectively capture the considering the holistic information in the multichannel
recurring patterns within the EEG data. Subsequently, these EEG data. LSTMs, as a special type of recurrent neural
learned connections can be visually observed, represented networks (RNN), were explicitly designed to address the
graphically, and further scrutinized, enabling us to delve challenge of long-term dependency in RNNs. Traditional
deeper into the data’s underlying structure. Additionally, they RNNs trained through back-propagation through time (BPTT)
empower us to perform comparative analyses of EEG data often encounter the vanishing/exploding gradient problem
across diverse subject groups. when learning from extended sequences. In order to overcome
4732 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 31, 2023

III. R ESULTS AND D ISCUSSION

This study involves two analysis steps. In the first step,
we investigate brain connectivity and patterns associated with
depression through visualization and interpretation of the SNN
model. In the subsequent step, we evaluate the classification
accuracy of our proposed model and compare it to other
existing methods employed by fellow researchers in the field
of depression recognition.

A. Pattern Discovery of Dynamic Brain Activities

Fig. 4. The structure of an LSTM cell. Associated With Depression Through Visualization of the
SNN Models
TABLE II
In this section, we explore the functional connections
T HE LSTM M ODEL PARAMETERS AND S ETTINGS within the brain by analyzing the insights obtained from
the learned SNNr models. To compare the underlying brain
functions across different states of depression, we conduct
separate STDP training of the SNN models using samples
from each group. For the analysis of the trained networks,
we construct graphs [67], which effectively depict the extent
of interactions among distinct brain regions. For quantifying
the extent of interaction among the input neurons within the
SNN models, we construct an N × N affinity matrix within
this challenge, LSTMs employ a gated cell structure as a the confines of the SNN model. This matrix captures the
replacement for the traditional RNN cell. Fig. 4 illustrates the aggregated spikes exchanged between neurons i and j via the
basic architecture of an LSTM cell. connection Wi j . Each input neuron establishes a cluster of
Due to the temporal nature, large receptive field size, and surrounding neurons, signifying those that receive the most
large number of channels of the SNNr output, we employ spikes from that particular input neuron relative to others.
an LSTM module to classify depression levels. The binary The level of spike interaction between any two groups of
output matrix of the previous module with the size of neurons is calculated in terms of the spikes exchanged. The
(Nsamples , Timesteps, Nr eser voir ) is fed to an LSTM layer strength of connections is visually depicted by the line’s
which is configured with 64 memory cells. Subsequently, thickness connecting nodes, symbolizing the intensity of spike
in order to prevent over-fitting and enhance the model’s transmission between different segments of the brain model.
generalizability, we utilize a dropout layer. This layer is In Fig. 5 and 6, we present the 500 strongest connections
connected to a fully-connected linear layer with 32 units and for each level of depression during both eyes-closed
a ReLU activation function. Lastly, we employ another fully- and eyes-open states. These connections are represented
connected linear layer with 4 units and a softmax activation by blue lines (indicating excitatory connections) and red
function to carry out the classification. In Table II, we present lines (indicating inhibitory connections). Additionally, the
the network’s model parameters and settings, while Table III brightness of neurons represents their spike emission. In both
provides a more comprehensive description of the network’s the eyes-closed and eyes-open states, there is a notable trend of
layers. Algorithm 4 exemplifies the conclusive stage of the the strongest connections in the brain shifting towards higher
model. regions as depression severity increases. This pattern suggests
disturbances in the normal connectivity patterns, potentially
Algorithm 4 Supervised Learning and Classification reflecting the impact of depression on neural communication
Require: X reservoir ∈ {0, 1}T ×Nr ▷ SNN reservoir output data and network dynamics. This scattering towards the top of the
Ensure: Yclasses ▷ Number of output classes brain, particularly involving the prefrontal cortex responsible
1: Split X reservoir into training and testing sets for cognitive functions and emotional regulation, suggests a
2: Initialize LSTM model for classification
significant impact of depression on these crucial processes.
3: for each epoch in training do
Our findings are aligned with previous research, including
4: Train LSTM model on X reservoir [68], which reported hyperconnectivity between the thalamus
5: end for
and the cortex in individuals with major depression, supporting
6: Initialize empty array Yclasses
the idea of increased connections between lower and higher
7: for each data point in testing set do
brain areas in depressed patients. Conversely, in non-depressed
states, stronger connections are concentrated in lower brain
8: Pass data through trained LSTM model
regions, such as the limbic system, which is associated with
9: Classify data into one of the classes
emotional processing and regulation, possibly indicating a
10: Append class label to Yclasses
more balanced emotional state. The observed changes in brain
11: end for
connectivity may also reflect the brain’s adaptive response
12: Return Yclasses
to depression, with neuroplasticity playing a role in forming
SAM et al.: DEPRESSION IDENTIFICATION USING EEG SIGNALS VIA A HYBRID OF LSTM AND SNNs 4733

TABLE III
D ETAILED PARAMETERS S ETTINGS OF THE LSTM M ODULE

Fig. 5. The connectivity results from four distinct SNNr modules in the eyes-closed state are depicted. For each SNNr, the top 500 strongest
connections are showcased. Positive (excitatory) connections are portrayed with blue lines, whereas negative (inhibitory) connections are
represented by red lines. The brightness of each neuron corresponds to its spike emission level: (a) minimal depression; (b) mild depression;
(c) moderate depression; (d) severe depression.

Fig. 6. The connectivity results from four distinct SNNr modules in the eyes-open state are depicted. For each SNNr, the top 500 strongest
connections are showcased. Positive (excitatory) connections are portrayed with blue lines, whereas negative (inhibitory) connections are
represented by red lines. The brightness of each neuron corresponds to its spike emission level: (a) minimal depression; (b) mild depression;
(c) moderate depression; (d) severe depression.

new connections to cope with the challenges posed by the and C3 and PO3 in both eyes-closed and eyes-open states.
condition. In contrast, for individuals with severe depression, strong
In Fig. 7 and 8 we present graphical representations of connections were identified between F5 and T7, FT8 and F6,
the brain’s 62 electrodes, based on the standard 10-20 EEG FT8 and T8, as well as T8 and PO8.
electrode system, for four levels of depression during both The results align with the findings regarding the differences
eyes-closed and eyes-open states. Notably, our results reveal a between eyes-closed and eyes-open states. In the eyes-open
remarkable increase in connections related to the frontal and state, there are more sparse connections, including long-range
prefrontal cortex regions as depression becomes more severe. connections. This alignment is consistent with the notion
This finding suggests that depression may have a specific that the brain integrates information from distant regions to
impact on the connectivity within these brain areas, which process sensory inputs. Thus, during eyes-open states, the
are known for their involvement in cognitive and emotional brain’s functional connectivity involves a broader network of
processing. Moreover, as the depression level increases, the brain regions communicating over longer distances to handle
connections tend to become less sparse but rather stronger. external sensory information. Conversely, during eyes-open
This indicates that more connections are formed, and existing states, the sparsity observed is in line with the understanding
connections become more robust, potentially reflecting a that synchronization and coherence may be weaker due
reorganization of neural communication during depressive to the brain’s active engagement in processing sensory
states. In the minimally depressed group, robust connections information. The weaker synchronization and coherence
were consistently observed between F6 and PO3, F6 and FT8, during these periods could lead to more isolated and less
4734 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 31, 2023

Fig. 7. The graphs effectively encapsulated the overall spike interaction during the eyes-closed state across regions within the SNN models,
symbolizing the 62 EEG channels as input neurons, throughout the STDP learning process for: (a) minimal depression; (b) mild depression;
(c) moderate depression; (d) severe depression. The nodes within the graphs depict the areas of input neurons in the SNN model, while the
thickness of lines represents the degree of spike transmission between these neuron areas (clusters). These clusters correspond to the input
neurons (EEG channels).

Fig. 8. The graphs effectively encapsulated the overall spike interaction during the eyes-open state across regions within the SNN models,
symbolizing the 62 EEG channels as input neurons, throughout the STDP learning process for: (a) minimal depression; (b) mild depression;
(c) moderate depression; (d) severe depression. The nodes within the graphs depict the areas of input neurons in the SNN model, while the
thickness of lines represents the degree of spike transmission between these neuron areas (clusters). These clusters correspond to the input
neurons (EEG channels).

Fig. 9. Correlation between each channel’s weighted degree centrality and depression severity for eyes-closed state.

coordinated neural activities, resulting in a sparser connectivity frontal and prefrontal cortex regions. Channels with low
pattern. correlation in the non-depressed group might play a crucial
In Fig. 9 and 10, we show the correlation between each role in maintaining emotional balance, while their weakening
channel’s weighted degree centrality and depression severity. connectivity may contribute to depressive symptoms. Channels
Channels FP1, Fpz, F3, PO5, and CP2 demonstrated the with higher correlation may be more directly related to
lowest correlation, suggesting that their connectivity weakens depressive symptoms, indicating their potential relevance in
as depression worsens. Conversely, channels AF3, AF4, F5, depression manifestation and progression.
F1, FT8, POz, and CP4 exhibited the highest correlation,
indicating a stronger association with depression severity. B. Classification Results
These findings suggest that depression level significantly To assess the classification efficacy of our proposed model,
impacts brain network connectivity, particularly involving the we conduct a comparative analysis with three well-established
SAM et al.: DEPRESSION IDENTIFICATION USING EEG SIGNALS VIA A HYBRID OF LSTM AND SNNs 4735

Fig. 10. Correlation between each channel’s weighted degree centrality and depression severity for eyes-open state.

Fig. 11. The averaged confusion matrixes for eyes-closed state along with the averaged accuracies for different models: (a) Proposed Model;
(b) Hashempoor et al., 2021; (c) Ay et al., 2019; (d) Achariya et al., 2018.

Fig. 12. The averaged confusion matrixes for eyes-open state along with the averaged accuracies for different models: (a) Proposed Model;
(b) Hashempoor et al., 2021; (c) Ay et al., 2019; (d) Achariya et al., 2018.

deep learning-based models using the same dataset. The first can be attributed to the model’s effective feature extraction,
model under consideration is the hybrid deep CNN-TCN specifically related to potential connectivity relationships
network, introduced by Hashempour et al. [24]. The second among different EEG channels, achieved through STDP
model combines a convolutional neural network with long unsupervised learning. The model’s biological plausibility
short-term memory (CNN-LSTM), as proposed by Ay et al. is a key advantage, as it is well-suited for processing
[69]. Lastly, we examine a 13-layer deep CNN presented biological EEG signals. This biologically-inspired approach
by Achariya et al. [22]. By juxtaposing our proposed model likely contributes significantly to the improved performance
against these prominent benchmarks, we aim to gain insights compared to alternative methods, aligning well with the
into its performance and potential advantages over existing inherent characteristics of EEG data. The utilization of
state-of-the-art approaches. just one LSTM layer to classify the output of SNNr
The model evaluation utilizes a 10-fold cross-validation further substantiates the efficacy of this biologically-inspired
procedure, with 10% of the subjects used as the testing approach.
set in each iteration. It is important to note that the
unsupervised learning step is conducted using the entire IV. C ONCLUSION
dataset in a single step, considering the time-consuming nature This study investigates the potential of depression recogni-
of this process. The aggregate performance is calculated by tion through the utilization of a novel combination of LSTM
averaging the outcomes from all ten evaluations. The results and SNN. For the first time, these models are employed
are showcased and compared in Fig. 11 and Fig. 12. The to model, map, learn, classify, visualize, and comprehend
results clearly indicate that our proposed model surpasses the EEG signals associated with four distinct depression levels,
other three models in classification accuracy. This success namely minimum, mild, moderate, and severe. The proposed
4736 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 31, 2023

model integrates diverse methods and algorithms that facilitate [11] K. Sadatnezhad, R. Boostani, and A. Ghanizadeh, “Proposing an
the exploration and investigation of multiple aspects within adaptive mutation to improve XCSF performance to classify ADHD
and BMD patients,” J. Neural Eng., vol. 7, no. 6, Dec. 2010,
EEG data. This encompasses the spatial mapping of data Art. no. 066006.
onto a three-dimensional SNN structure, unsupervised learning [12] F. C. Morabito et al., “Deep learning representation from elec-
within the SNNr, visualization of connectivity and spiking troencephalography of early-stage Creutzfeldt–Jakob disease and
features for differentiation from rapidly progressive dementia,” Int.
patterns within the trained SNNr to unveil novel insights J. Neural Syst., vol. 27, no. 2, Mar. 2017, Art. no. 1650039, doi:
into the data and underlying brain mechanisms, along with 10.1142/s0129065716500398.
supervised learning within an LSTM network. Comparative [13] S. Kouchaki, R. Boostani, and F. Razavipour, “Fatigue detection using
analysis with other deep learning techniques showcases the the strength of dominant EEG source: A beamforming approach,”
Biomed. Eng., Appl., Basis Commun., vol. 30, no. 3, Jun. 2018,
advantages of employing the SNN approach in modeling time- Art. no. 1850023.
space brain data. This study not only achieves improved [14] F. Razavipour, R. Boostani, S. Kouchaki, and S. Afrasiabi, “Comparative
accuracy in classifying samples from different subject groups application of non-negative decomposition methods in classifying
fatigue and non-fatigue states,” Arabian J. Sci. Eng., vol. 39, no. 10,
but also reveals informative patterns of brain activities, pp. 7049–7058, Oct. 2014.
shedding light on the understanding of different severity of [15] F. Karimzadeh, M. Nami, and R. Boostani, “Sleep microstructure
depression. Our findings unveil significant differences between dynamics and neurocognitive performance in obstructive sleep apnea
different depression levels that hold promise as potential syndrome patients,” J. Integrative Neurosci., vol. 16, no. 2, pp. 127–142,
Feb. 2018.
markers for early prediction and prevention of depression. The [16] R. Boostani, F. Karimzadeh, and M. Nami, “A comparative review on
proposed methodology exhibits wide applicability to diverse sleep stage classification methods in patients and healthy individuals,”
neuroimaging and clinical longitudinal data. Future work will Comput. Methods Programs Biomed., vol. 140, pp. 77–91, Mar. 2017.
focus on refining SNN hyper-parameter optimization, further [17] F. Alimardani and R. Boostani, “DB-FFR: A modified feature selection
algorithm to improve discrimination rate between bipolar mood disorder
enhancing the visualization and analysis of the brain-structured (BMD) and schizophrenic patients,” Iranian J. Sci. Technol., Trans.
SNN during both the learning process and post-learning phase Electr. Eng., vol. 42, no. 3, pp. 251–260, Sep. 2018.
to deepen our comprehension of brain processes related to [18] D. Jarchi, R. Boostani, M. Taheri, and S. Sanei, “Seizure source
localization using a hybrid second order blind identification and
depression. Furthermore, the exploration of larger and well- extended rival penalized competitive learning algorithm,” Biomed. Signal
balanced datasets, as well as task-specific EEG signals, will Process. Control, vol. 4, no. 2, pp. 108–117, Apr. 2009.
be undertaken to delve into the impact of varying tasks on the [19] M. Cukic, M. Stokic, S. Simic, and D. Pokrajac, “The successful
analysis. This approach seeks to comprehensively investigate discrimination of depression from EEG could be attributed to proper
feature extraction and not to a particular classification method,” Cognit.
the effects of different tasks on the model’s performance and Neurodynamics, vol. 14, no. 4, pp. 443–455, Aug. 2020.
insights. [20] S. D. Puthankattil and P. K. Joseph, “Classification of EEG signals
in normal and depression conditions by ANN using RWE and
signal entropy,” J. Mech. Med. Biol., vol. 12, no. 4, Sep. 2012,
R EFERENCES Art. no. 1240019.
[21] U. R. Acharya, V. K. Sudarshan, H. Adeli, J. Santhosh, J. E. W. Koh, and
[1] Depression and Other Common Mental Disorders: Global Health A. Adeli, “Computer-aided diagnosis of depression using EEG signals,”
Estimates, World Health Organization, Geneva, Switzerland, 2017. Eur. Neurol., vol. 73, nos. 5–6, pp. 329–336, 2015.
[Online]. Available: https://apps.who.int/iris/handle/10665/254610 [22] U. R. Acharya, S. L. Oh, Y. Hagiwara, J. H. Tan, H. Adeli, and
[2] Y. Takahashi, “Depression and suicide,” Jpn. Med. Assoc. J., vol. 44, D. P. Subha, “Automated EEG-based screening of depression using deep
no. 8, pp. 359–363, 2001. convolutional neural network,” Comput. Methods Programs Biomed.,
[3] A. T. Beck, R. A. Steer, and G. Brown, Manual For The Beck Depression vol. 161, pp. 103–113, Jul. 2018.
Inventory-II. San Antonio, TX, USA: Psychological Corporation, 1996, [23] M. Xia, Y. Zhang, Y. Wu, and X. Wang, “An end-to-end
doi: 10.1037/t00742-000. deep learning model for EEG-based major depressive disorder
[4] K. L. Smarr and A. L. Keefer, “Measures of depression and classification,” IEEE Access, vol. 11, pp. 41337–41347, 2023, doi:
depressive symptoms: Beck depression Inventory-II (BDI-II), center for 10.1109/ACCESS.2023.3270426.
epidemiologic studies depression scale (CES-D), geriatric depression [24] S. Hashempour, R. Boostani, M. Mohammadi, and S. Sanei, “Continuous
scale (GDS), hospital anxiety and depression scale (HADS), and patient scoring of depression from EEG signals via a hybrid of convolutional
health Questionnaire-9 (PHQ-9),” Arthritis Care Res., vol. 63, no. 11, neural networks,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 30,
pp. 454–466, Nov. 2011. pp. 176–183, 2022, doi: 10.1109/TNSRE.2022.3143162.
[5] R. P. N. Rao, Brain-Computer Interfacing: An Introduction. [25] A. Seal, R. Bajpai, J. Agnihotri, A. Yazidi, E. Herrera-Viedma, and
U.K.: Cambridge Univ. Press, 2011, ch. 3.1.12, p. 26, doi: O. Krejcar, “DeprNet: A deep convolution neural network framework for
10.1017/CBO9781139032803. detecting depression using EEG,” IEEE Trans. Instrum. Meas., vol. 70,
[6] M. Wang and J. Y. Chen, “A GMM-IG framework for selecting genes pp. 1–13, 2021, doi: 10.1109/TIM.2021.3053999.
as expression panel biomarkers,” Artif. Intell. Med., vol. 48, nos. 2–3, [26] S. Soni, A. Seal, A. Yazidi, and O. Krejcar, “Graphical representation
pp. 75–82, Feb. 2010, doi: 10.1016/j.artmed.2009.07.006. learning-based approach for automatic classification of electroencephalo-
[7] A. Kolahi, M. Hoviattalab, T. Rezaeian, M. Alizadeh, M. Bostan, and gram signals in depression,” Comput. Biol. Med., vol. 145, Jun. 2022,
H. Mokhtarzadeh, “Design of a marker-based human motion tracking Art. no. 105420, doi: 10.1016/j.compbiomed.2022.105420.
system,” Biomed. Signal Process. Control, vol. 2, no. 1, pp. 59–67, [27] S. Soni, A. Seal, S. K. Mohanty, and K. Sakurai, “Electroencephalog-
Jan. 2007, doi: 10.1016/j.bspc.2007.02.001. raphy signals-based sparse networks integration using a fuzzy ensemble
[8] M. Sabeti, R. Boostani, and T. Zoughi, “Using genetic programming to technique for depression detection,” Biomed. Signal Process. Control,
select the informative EEG-based features to distinguish schizophrenic vol. 85, Aug. 2023, Art. no. 104873, doi: 10.1016/j.bspc.2023.104873.
patients,” Neural Netw. World, vol. 22, no. 1, pp. 3–20, 2012. [28] A. Seal et al., “Benchmarks for machine learning in depression discrim-
[9] B. Kazimipour, R. Boostani, A. Borhani-Haghighi, S. Almatarneh, and ination using electroencephalography signals,” Int. J. Speech Technol.,
M. Aljaidi, “EEG-based discrimination between patients with MCI vol. 53, no. 10, pp. 12666–12683, May 2023, doi: 10.1007/s10489-022-
and Alzheimer,” in Proc. Int. Eng. Conf. Electr., Energy, Artif. Intell. 04159-y.
(EICEEAI), Nov. 2022, pp. 1–5. [29] A. L. Hodgkin, A. F. Huxley, and B. Katz, “Measurement of
[10] K. Sadatnezhad, R. Boostani, and A. Ghanizadeh, “Classification of current-voltage relations in the membrane of the giant axon of
BMD and ADHD patients using their EEG signals,” Exp. Syst. Appl., Loligo,” J. Physiol., vol. 116, no. 4, pp. 424–448, Apr. 1952, doi:
vol. 38, no. 3, pp. 1956–1963, Mar. 2011. 10.1113/jphysiol.1952.sp004716.
SAM et al.: DEPRESSION IDENTIFICATION USING EEG SIGNALS VIA A HYBRID OF LSTM AND SNNs 4737

[30] J. E. Hall, Guyton and Hall Textbook of Medical Physiology, 13th ed. [50] E. Tu et al., “NeuCube(ST) for spatio-temporal data predictive modelling
Amsterdam, The Netherlands: Elsevier, 2015. with a case study on ecological data,” in Proc. Int. Joint Conf. Neural
[31] W. Maass, “Networks of spiking neurons: The third generation of neural Netw. (IJCNN), Jul. 2014, pp. 638–645.
network models,” Neural Netw., vol. 10, no. 9, pp. 1659–1671, 1997, [51] M. G. Doborjeh, E. Capecci, and N. Kasabov, “Classification and
doi: 10.1016/S0893-6080(97)00011-7. segmentation of fMRI spatio-temporal brain data with a NeuCube
[32] T. Masquelier, R. Guyonneau, and S. J. Thorpe, “Competitive STDP- evolving spiking neural network model,” in Proc. IEEE Symp.
based spike pattern learning,” Neural Comput., vol. 21, no. 5, Evolving Auto. Learn. Syst. (EALS), Orlando, FL, USA, Dec. 2014,
pp. 1259–1276, May 2009, doi: 10.1162/neco.2008.06-08-804. pp. 73–80.
[33] S. Thorpe and J. Gautrais, “Rank order coding,” in Computational [52] N. Kasabov, “Brain-, gene-, and quantum inspired computational intelli-
Neuroscience. Berlin, Germany: Springer, 1998, pp. 113–118. gence: Challenges and opportunities,” in Challenges for Computational
[34] A. Tavanaei, M. Ghodrati, S. R. Kheradpisheh, T. Masquelier, and Intelligence, vol. 63, W. Duch and J. Mandziuk, Eds. Berlin, Germany:
A. Maida, “Deep learning in spiking neural networks,” Neural Netw., Springer, 2007, pp. 81–97, doi: 10.1007/978-3-540-71984-7_9.
vol. 111, pp. 47–63, Mar. 2019. [53] I. J. Goodfellow et al., “Challenges in representation learning: A report
[35] R. Mohanty, A. Priyadarshini, V. S. Desai, and G. Sirisha, “Appli- on three machine learning contests,” Neural Netw., vol. 64, pp. 59–63,
cations of spiking neural network to predict software reliability,” in Apr. 2015, doi: 10.1016/j.neunet.2014.09.005.
Intelligent Engineering Informatics, vol. 695. Singapore: Springer, 2018, [54] B. Petro, N. Kasabov, and R. M. Kiss, “Selection and optimization of
pp. 149–157. temporal spike encoding methods for spiking neural networks,” IEEE
[36] Z. Doborjeh et al., “Spiking neural network modelling approach reveals Trans. Neural Netw. Learn. Syst., vol. 31, no. 2, pp. 358–370, Feb. 2020,
how mindfulness training rewires the brain,” Sci. Rep., vol. 9, no. 1, doi: 10.1109/TNNLS.2019.2906158.
p. 6367, Apr. 2019. [55] S. M. Bohte, “The evidence for neural information processing with
[37] N. Kasabov et al., “Evolving spatio-temporal data machines based on the precise spike-times: A survey,” Natural Comput., vol. 3, no. 2,
NeuCube neuromorphic framework: Design methodology and selected pp. 195–206, 2004, doi: 10.1023/B:NACO.0000027755.02868.60.
applications,” Neural Netw., vol. 78, pp. 1–14, Jun. 2016. [56] M. Ghesmoune, M. Lebbah, and H. Azzag, “A new growing neural gas
[38] D. Shah, G. Y. Wang, M. Doborjeh, Z. Doborjeh, and N. Kasabov, for clustering data streams,” Neural Netw., vol. 78, pp. 36–50, Jun. 2016,
“Deep learning of EEG data in the NeuCube brain-inspired spiking doi: 10.1016/j.neunet.2016.02.003.
neural network architecture for a better understanding of depression,” in [57] N. Nuntalid, K. Dhoble, and N. Kasabov, “EEG classification with
Neural Information Processing (Lecture Notes in Computer Science), BSA spike encoding algorithm and evolving probabilistic spiking neural
vol. 11955. Berlin, Germany: Springer, Dec. 2019, pp. 62–76, doi: network,” in Proc. Int. Conf. Neural Inf. Process. (ICONIP). Berlin,
10.1007/978-3-030-36718-3_17. Germany: Springer, 2011, pp. 451–460.
[39] J. F. Cavanagh, A. Napolitano, C. Wu, and A. Mueen, “The [58] T. Delbruck. (2007). JAER Open Source Project. [Online]. Available:
patient repository for EEG data + computational tools (PRED+CT),” http://jaer.wiki.sourceforge.net
Frontiers Neuroinform., vol. 11, p. 67, Nov. 2017. [Online]. Available: [59] P. Lichtsteiner and T. Delbruck, “A 64×64 AER logarithmic temporal
http://predict.cs.unm.edu/downloads.php derivative silicon retina,” Res. Microelectron. Electron., vol. 2,
[40] J. F. Cavanagh, A. W. Bismark, M. J. Frank, and J. J. B. Allen, “Multiple pp. 202–205, Jul. 2005.
dissociations between comorbid depression and anxiety on reward and [60] C. Tan, M. Sarlija, and N. Kasabov, “Spiking neural networks:
punishment processing: Evidence from computationally informed EEG,” Background, recent development and the NeuCube architecture,” Neural
Comput. Psychiatry, vol. 3, pp. 1–17, Jan. 2019. Process. Lett., vol. 52, no. 2, pp. 1675–1701, Oct. 2020.
[41] M. Tohen, C. Bowden, A. A. Nierenberg, and J. Geddes, Clinical [61] J. L. Lancaster et al., “Automated Talairach atlas labels for functional
Trial Design Challenges in Mood Disorders. Cambridge, MA, USA: brain mapping,” Hum. Brain Mapping, vol. 10, no. 3, pp. 120–131, 2000.
Academic Press, 2015. [62] J. L. Lancaster et al., “Automated labeling of the human brain:
[42] G. H. Klem, H. O. Lüeders, H. H. Jasper, and C. Elger, “The ten-twenty A preliminary report on the development and evaluation of a
electrode system of the international federation,” Electroencephalogr. forward-transform method,” Hum. Brain Mapping, vol. 5, no. 4,
Clin. Neurophysiol., vol. 52, no. 3, pp. 3–6, 1999. pp. 238–242, 1997.
[43] A. Zajac and P. Paszkiel, “Using brain-computer interface technology [63] L. Koessler et al., “Automated cortical projection of EEG sensors:
as a controller in video games,” Informatyka, Automatyka, Pomiary w Anatomical correlation via the international 10–10 system,” NeuroImage,
Gospodarce i Ochronie Srodowiska, vol. 10, no. 3, pp. 26–31, 2020. vol. 46, no. 1, pp. 64–72, 2009, doi: 10.1016/j.neuroimage.2009.02.006.
[44] X. Ding, X. Yue, R. Zheng, C. Bi, D. Li, and G. Yao, “Classifying major [64] S. Song, K. D. Miller, and L. F. Abbott, “Competitive Hebbian learning
depression patients and healthy controls using EEG, eye tracking and through spike-timing-dependent synaptic plasticity,” Nature Neurosci.,
galvanic skin response data,” J. Affect. Disorders, vol. 251, pp. 156–161, vol. 3, no. 9, pp. 919–926, Sep. 2000.
May 2019. [65] P. Dayan and L. F. Abbott, “Neurons and neural circuits,” in Theoretical
[45] A. Gramfort, “MEG and EEG data analysis with MNE- Neuroscience: Computational and Mathematical Modeling of Neural
Python,” Frontiers Neurosci., vol. 7, pp. 1–13, Jan. 2013, doi: Systems. Cambridge, MA, USA: MIT Press, ch. 5, 2001, pp. 153–166.
10.3389/fnins.2013.00267. [66] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
[46] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, Comput., vol. 9, no. 8, pp. 1735–1780, Nov. 1997.
no. 7553, pp. 436–444, 2015. [67] N. Kasabov, Time-Space, Spiking Neural Networks and Brain-
[47] W. Wang et al., “Learning of spatiotemporal patterns in a spiking neural Inspired Artificial Intelligence. Heidelberg, Germany: Springer,
network with resistive switching synapses,” Sci. Adv., vol. 4, no. 9, 2019.
Sep. 2018, Art. no. eaat4752. [68] L. Schmaal et al., “Subcortical brain alterations in major depressive
[48] A. Taherkhani, A. Belatreche, Y. Li, G. Cosma, L. P. Maguire, and disorder: Findings from the ENIGMA major depressive disorder working
T. M. McGinnity, “A review of learning in biologically plausible spiking group,” Mol. Psychiatry, vol. 21, no. 6, pp. 806–812, Jun. 2016, doi:
neural networks,” Neural Netw., vol. 122, pp. 253–272, Feb. 2020. 10.1038/mp.2015.69.
[49] C. Teeter et al., “Generalized leaky integrate-and-fire models classify [69] B. Ay et al., “Automated depression detection using deep representation
multiple neuron types,” Nature Commun., vol. 9, no. 1, p. 709, and sequence learning with EEG signals,” J. Med. Syst., vol. 43, no. 7,
Feb. 2018. pp. 1–12, Jul. 2019.