Early Detection of Student Depression Using Deep

Learning Algorithms
Mr. N. Bala Krishna Midigesi Mounika Pemmana Sowmya
Assistant Professor,Department of Department of Computer Science Department of Computer Science
Computer Science and Engineering, and Engineering, Sree Vidyanikethan and Engineering, Sree Vidyanikethan
Sree Vidyanikethan Engineering Engineering College (Autonomous), Engineering, College (Autonomous),
College (Autonomous) Tirupati, India. Tirupati, India.
Tirupati, India. [email protected] [email protected]
[email protected]

Sedji Vineetha Palukuri Venkata Sai Surya Teja

Department of Computer Science Department of Computer Science
and Engineering, Sree Vidyanikethan and Engineering, Sree Vidyanikethan
Engineering, College (Autonomous), Engineering, College (Autonomous),
Tirupati, India. Tirupati, India.
[email protected] [email protected]

Abstract— The aim of this project is to develop an advanced intervention and support. This delay in the speed of
system that applies deep learning techniques to detect early- detection worsens the condition, resulting in long-term
stage depression in students. Growing concern exists for corrosive effects on both their academic success and their
depression in educational settings and early identification personal development. Early detection is, therefore, key in
facilitates timely intervention and support. The system looks at reducing the life-altering impact of depression and in
a great deal of data from students—responses to standardized
ensuring that depressed students get the necessary support
questionnaires, behavioral factors like mood, sleep, social
interaction, and school performance. For this purpose, the before things worsen.
system applies Convolutional Neural Networks (CNN), Long
Short-Term Memory (LSTM) and Recurrent Neural Networks Although traditionally depression in students has been
(RNN) are deep learning algorithms. to automatically identified through self-reports, clinical interviews, and
recognize complex and subtle patterns in the data that can standardized questionnaires, this study was based on brain
inform about depression. Traditional methods fail to notice imaging technologies rather than relying on any such self-
these patterns which makes this approach more efficient in reporting or interview methods. Although these methods
detecting students at risk. The dataset is trained on a wide and can be very informative, they are based on subjective
complete dataset that integrates fully various emotional and
answers and do not generally include the more nuanced
behavioral factors to offer a sturdy and exact model. We
measure these systems using important factors like how often behavioral clues that occur when depression begins. These
they are right, how many relevant cases they get right, how few approaches are further limited by the existence of trained
they miss, and their combination of both precision and recall professionals and by students’ reluctance to seek help
to make sure they work well. We want to help schools get a tool because help-seeking is stigmatized or students fear they
that can find signs of depression in students early and reliably, will be judged if they ask for help. For this reason, there is
in order to act more quickly and provide timely support. The an increasing demand for improved, efficient, and accurate
system is trying to detect mental health issues in students and methods that can detect depression at an early stage, over
shed off the effect that depression negatively has on student’s the deficiencies of traditional methods.
academic performance and personal development. An
important contribution to the ever-growing need for mental
health monitoring in educational settings for facilitating the The advancements Recent progress in machine learning and
overall well-being and success of students. artificial intelligence enables developers to create early
detection systems that process large data sets and perform
Keywords: CNN, LSTM, RNN, early detection, student analysis of them for patterns that can go unnoticed by a
depression, sleep patterns, social interactions, academic human evaluator. These technologies can examine a number
performance, early intervention, mental health monitoring. of factors in student depression including academic
performance, mood variation, sleep, social interaction, and
I. INTRODUCTION
questionnaire responses. By using these various data
Depression is a massive mental health problem affecting sources, the system can produce an overall and objective
people globally across all kinds of people; students are view of a student’s mental health.
specifically susceptible because of the unique stress they This gap can be addressed through the development of a
face both in educational and social circles. Depression has a deep learning-based system for early detection of student
devastating effect on students, affecting every aspect of their depression that provides an automated, data-driven
emotional, psychological, academic, social, and general approach that can observe subtle depression signs before
well-being. Because depression is so prevalent, many times manifesting as a severe form. However, this system can help
it will go undiagnosed until a late stage, preventing effective

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

determine which students are at risk and administer performing transformer models DistilBERT, Squeeze
appropriate interventions in time to improve their mental BERT, DeBERTA, and RoBERTa. Our study showed that
health and academic outcomes. The system is able to take the RoBERTa model was the most accurate at spotting
data from many sources to represent a student's overall well- depression from social media messages, reaching 98.1%
being which helps the educators and counselors to success. Our analysis establishes a valuable foundation for
contribute to the student based on the student's needs. tracking depression transmission patterns across social
platforms using this data.
Early detection of depression in students is of paramount
importance. Studies show that people with depression who The limitations of previous models [5] for depression
get early intervention fare significantly better than people detection including their incomplete use of data distribution
who do not and are less likely to experience long-term or ordinal relationships between facial expressions and
psychological and emotional distress. In addition, handling depression levels led Carneiro de Melo and their team [6] to
mental health problems during the early stages can improve set up a new deep learning system to solve this challenge.
students’ general academic achievements and social They introduce a new expectation loss function in their
integration, as well as promote an allusive environment of model which optimizes the underlying data distribution to
school support and productivity. Therefore, this project's predict depression levels better, even with unknown
intention is to support the field of mental health monitoring labeling. We prove the approach utilized has better
in education by creating a tool not only to already identify performance on AVEC2013 and AVEC2014 datasets on
Depression (early), but also to promote the success and predicting depression levels from facial expressions, than
well-being of students in School and in their overall lives. existing methods.
II. LITERATURE SURVEY Convolutional Neural Networks (CNNs) utilized with
OpenCV and Haar Cascade Classifier for facial expression
Gaikwad et al. [1 The authors introduced a deep learning
recognition and proposed as a system for depression
system designed to assess how depressed students feel, with
detection by Nazira and Das [7]. The system uses Haar
a focus on early detection and intervention. The study shows
Cascade for face detection and CNN to analyze facial
their finding about the rapid spikes of depression among
expressions which are indicative of psychological states.
students, largely due to academic and personal problems. In
For this purpose a new Depressed and I put together the
this thesis, the authors recommend to use of statistical
Non-Depressed (DND) dataset, making it up of 5000
methods like Pearson correlation, We use K Nearest
images. Results show that the system reached 81 percent
Neighbors plus Artificial Neural Networks and Decision
accuracy, 87 percent precision, and 88 percent recall,
Tree to forecast depression levels more precisely. The
proving that facial expression analysis is effective for the
research is intended to present strategies for the effective
diagnosis of depression.
management of student mental health, including possible
solutions for educational institutions to help foster student The effect of the COVID-19 crisis on mental health was
well-being and academic success. studied by Sofia et al. [8], where they focused in their study
on depression. Their research underscores the message that
Moreover, Shah et al. [2] presented We use a combined deep
we need to catch depression early and intervene before it
learning model to spot depression signs early from written
gets worse and people put themselves at risk of more serious
posts shared on social media websites. Depression affects
health problems like heart disease, stroke or suicide. The
300 million people across the entire planet which is why the
work shows how machine learning models can be used to
study highlights social networks as a sensor of early
detect depression during the pandemic, providing a tool for
detection. To predict depression, the authors trained We
timely intervention for the exacerbated mental health crisis
tested our BiLSTM model while using several word
caused by the pandemic.
embedding methods in combination with metadata. features
and produced promising results. We test our approach on the The problem of depression has been discussed by Nikhil
Reddit dataset of the CLEF eRisk 2017 challenge and show Chauhan et al. [9], with a special emphasis put on the case
that social media analysis can be leveraged to prevent of students suffering stress due to academic, extracurricular,
diseases and suicides associated with depression. and social factors. According to their research, they suggest
a system where a questionnaire on human stress levels is
Inspired, Amanat et al. [3] launched a new deep learning
measured by utilizing machine learning. The system based
technique to spot depression early on from textual data that
on the responses it receives provides them with suggestions
exploited the predictive power of social media and internet
if they should consult a psychologist or not and serves them
users' emotional responses. With Long Short Term Memory
a self-assessment tool useful for those who may hardly seek
(LSTM) and Recurrent Neural Network (RNN), the authors
professional help.
built a model that has a depression prediction accuracy of
99.0%, superior to traditional frequency-based models. This Our study systematically evaluates existing machine
study shows how semantic analysis and written content can learning and deep learning techniques for detecting
detect depression and that this may be a potential model for depression in social media content was conducted by
timely intervention in the world of mental health, and Wadzani Aduwamai Gadzama et al. [10]. To this end, the
potentially, to save lives from suicide. review analyzed 60 articles, 28 of which were on machine
learning techniques The survey examines 37 research
Bokolo and Liu [4] examine Our research using machine
studies on traditional classifiers including decision trees, K-
learning and transformer methods to find depression signs
nearest neighbors, naive Bayes, random forests, and support
in tweets people post on social media. The study tested
vector machines along with 32 investigations on deep
different prediction methods on its collection of 632,000
learning methods like convolutional neural networks, long
tweets including basic techniques like logistic regression
short term memory and recurrent neural networks. I applied
and Naive Bayes plus random forest then evaluated high-
these models to datasets from platforms like Twitter, pressure and study satisfaction can improve the model's
Facebook, and Reddit, and also from Kaggle clinical patient capacity to identify student depression. The accuracy and
records. Recent research about research shows that deep efficacy of the machine learning models are eventually
learning helps detect depression earlier by using social increased by these preprocessing procedures, which
media data to recognize symptoms. guarantee that the data is clear, consistent, and prepared for
analysis.
III. DATASET
We pulled this useful educational data from Kaggle to 3. Model Selection:
include significant features that help us identify depression
in students. Demographic and behavioral factors, such as We developed deep learning models for early depression
Age, Gender, and Family History form the foundational detection using the Student Depression Dataset. The data
understanding of the students. Academic Pressure, Study was preprocessed with normalization, encoding, and feature
Satisfaction and Study Hours provide key insight into the engineering, then split into 80% training and 20% testing.
student’s academic activities, while Financial Stress is We implemented CNN, LSTM, and RNN models to identify
relevant to external stresses affecting Academic Activities. patterns and relationships in the data effectively.
Physical well-being has indicators such as sleep Duration
and Dietary Habits which also have an influence on mental The overall methodology is illustrated in Figure 1 which
health. The ability to identify severe depression symptoms presents a structured flowchart of our process. Our approach
is ensured by the presence of Suicidal Thoughts. Depression aims to provide an early and accurate detection mechanism,
is the target variable: Is the student suffering from enabling timely intervention and mental health support for
depression. When these features are taken in unison, they students at risk of depression.
can give a comprehensive look at what causes depression
and this can help deep learning models to catch complicated
patterns and help define at-risk students accurately.
IV. METHODOLOGY
Our project method combines six stages from data
collection to system deployment. Our system monitors
student actions and information to spot depression signs.
Our system uses advanced deep learning methods CNN,
LSTM, and RNN to discover the meaningful signs of
depression. Our method uses state-of-the-art tools to find
issues early and take action.
1. Data Collection:
Our study uses mental health datasets from Kaggle that
show multiple factors influencing student health outcomes.
Figure 1: Flow chart for depression detection
This research uses distinct features such as age and
academic stress through to familial patterns in history and
A). Convolutional Neural Networks (CNN):
diet to target the main focus of depression. The data includes
essential information about students' actions and feelings This project uses Convolutional Neural Networks
which cannot be seen directly but help find early depression (CNN) to find patterns in students' behavioral and academic
data in order to identify depression in those students. In
indications.
order to identify significant characteristics that might point
2. Data Preprocessing: to depression, CNNs apply filters to input data, such as
Data preprocessing is essential to this project because it study hours, sleep patterns, and social interactions.
guarantees the quality of the data used in machine learning
models. Imputation techniques, which fill numerical By using these filters, the model is better able to
features with the mean and categorical features with the identify minute patterns in the data that may be connected
mode, are crucial steps in handling missing values. to mental health problems. Due to their ability to
Normalization and standardization are used to make sure comprehend both local and global relationships in the data,
that characteristics like age, study hours, sleep duration, and CNNs are especially well-suited for this task. This enables
financial stress are on a comparable scale. This guarantees them to identify complex patterns and generate precise
that the model's learning process is not dominated by predictions regarding student depression.
features with higher values. To make sure the model learns The equations used in this project are essential to the
from enough data and is tested on unseen examples, the efficient operation of the CNN. The convolution operation
dataset is further split into three sections: 80% for training, in the first equation enables the network to extract features
10% for validation, and 10% for testing. and apply filters to the input data. It is shown as:

One-hot encoding is used to process categorical features, 𝑠(ⅈ, 𝑗) = (𝐼 ∗ 𝑘)(ⅈ, 𝑗) = ∑𝑚 ∑𝑛 𝐼(𝑚, 𝑛) . 𝐾( ⅈ − 𝑚 , 𝑗 − 𝑛)

such as gender, and convert them into binary vectors that (1)
machine learning algorithms can use. Feature engineering is Where S(i,j) is the output feature map, I(m,n)is the input
also used to develop new features that capture more intricate data, and K(i−m,j−n) is the filter being applied. This
relationships in the data. For example, combining academic operation scans the data to identify key features. After
convolution, the activation function (like ReLU) is used to emotional state over time. The output layer then predicts the
introduce non-linearity and help the model learn more likelihood of depression based on these learned patterns.
complex relationships as represented in equation 2.
4. Model Training:
𝐴(𝑥) = max(0, 𝑥) (2) We separate the dataset into training, validation,
Finally, the pooling operation (in equation 3) is applied and test sets to make sure our model learns efficiently. This
to reduce the size of the feature maps, making the model improves the model's generalization by preventing it from
more efficient while preserving important features.
learning noise. We measure prediction errors during training
using entropy as the loss function, and then we apply
𝑃(ⅈ, 𝑗) = max({ 𝑆(ⅈ + 𝑘, 𝑗 + 𝑙) |0 ≤ 𝑘, 𝑙 < 𝑓}) (3)
optimization techniques like Adam or SGD to minimize
In this above equation, P(i,j) represents the output these errors. In order to ensure that our model gets better
value after applying max pooling at position (i,j). The term over time, this step is essential.
S(i+k,j+l) refers to the values in the input feature map within The model repeatedly runs through the entire
the pooling window. The max function selects the highest dataset during the several stages (epochs) of the training
value from this window, helping retain the most important process. By modifying the model's internal parameters
features. The variables k and l indicate the positions within through backpropagation, each epoch aids in its
the window, while f defines the size of the pooling window, improvement. Binary cross-entropy is perfect for detecting
determining how many values are considered at once. depression because it allows us to compare expected and
actual values. After every iteration, the Adam optimizer
B). Long Short-Term Memory (LSTM): modifies the model's parameters to minimize error and
guarantee consistent gains.
LSTM networks are a type of RNN that handles long-
term dependencies by solving the vanishing gradient 5. Model Evaluation:
problem. They use memory cells with three gates: the input The number of correct predictions the model made when
gate stores new information, the forget gate removes identifying student depression is known as accuracy. A high
unnecessary data, and the output gate decides what to pass accuracy indicates that the model is generally functioning
forward. This helps LSTMs retain important details over well, but if there is an imbalance between depressed and
long sequences, making them useful for speech recognition, non-depressed cases, it might not always be dependable.
text analysis, and behavior tracking.
𝐴𝑐𝑡𝑢𝑎𝑙 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠 𝑇𝑃 + 𝑇𝑁
Our working on temporal data, for example, changes in 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = = 𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁 (5)
𝑇𝑜𝑡𝑎𝑙 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠
Academic Pressure, Study Satisfaction, and Suicidal
Thoughts over time, led us to use LSTMs. This is especially Where TP is the true positives for depressed students, TN is
powerful for detecting slow-developing depression patterns the true negatives for non-depressed students, FP is the false
like a student’s emotional state deteriorating over weeks or
positives for non-depressed students and FN is the false
months.
negatives for depressed students.
C). Recurrent Neural Networks (RNN): Precision focuses on how many of the students predicted as
depressed are actually depressed. This is important because
The ability of recurrent neural networks (RNNs) to false positives can cause unnecessary stress and
process sequential data makes them perfect for applications interventions.
like time series and behavioral data analysis where the 𝑇𝑃
sequence of inputs is crucial. In contrast to conventional 𝑃𝑟𝑒𝑐ⅈ𝑠ⅈ𝑜𝑛 = 𝑇𝑃+𝐹𝑃 (6)
feedforward neural networks, RNNs are able to retain a
memory of past inputs because of feedback loops. Recall looks at how well the model identifies all the actual
When examining patterns that change over time, RNNs' positive cases (depressed students). It calculates the
ability to learn temporal dependencies in data is crucial, and proportion of actual depressed students that the model
this special structure makes it possible. To capture the correctly identifies. A high recall means the model is good
impact of previous inputs on current data, the RNN updates at catching most of the true positive cases, though it might
(equation 4) its hidden state with each new input. also misidentify some non-depressed students as depressed
(false positives).
ℎ𝑡 = tanh(𝑤ℎ ℎ𝑡−1 + 𝑤𝑥 𝑥𝑡 + 𝑏ℎ ) (4)
The F1-Score combines precision and recall into one metric.
where, ht is the current hidden state, influenced by the
It’s the harmonic mean of the two, balancing the trade-off
previous hidden state ht-1 (capturing past information) and
between false positives and false negatives. A high F1-Score
the current input Xt. The weights Wh and Wx determine how
means the model is both accurate in its positive predictions
much past and present information contribute, while bh is a
and thorough in identifying true positives.
biased term, and tanh adds non-linearity.
The model is designed to process sequential data, with
an input layer containing nodes for features like Suicidal V. RESULTS AND DISCUSSIONS
Thoughts and Family History. Hidden layers apply This project reveals that deep learning-based methods
weighted transformations to capture evolving patterns in the can detect depression in students from behavioral and
demographic features. Our research used three various
models: CNN, LSTM, and RNN to test the system's
performance. Our evaluation of these models relied on four
performance measurements: accuracy rating and four other
metrics(table 1). Our experiments showed that the LSTM
model achieved the highest performance by reaching 99%
accuracy. Due to the capability of LSTM to learn long-term
dependencies in sequential data, for example, the temporally
varying fluctuations in mood, sleep patterns, academic
pressure, etc.
Table 1: Comparison of Model Performance Metrics

Model Accuracy Precision Recall F1-

Score
99 97 96 96
LSTM Fig. 3 Confusion matrix of the LSTM model
97 95 96 95 showing actual vs. predicted classifications.
CNN
96 94 94 94 2). CNN Accuracy Curve and Confusion Matrix:
RNN
The accuracy plot in Figure 4 shows the training and
validation accuracy of the CNN model over multiple
The following images represent the accuracy curves and
epochs. Both curves demonstrate a steady increase,
confusion matrices for each of the models: indicating effective learning. The model achieves high
1). LSTM Accuracy Curve and Confusion Matrix: accuracy with minimal overfitting.
The LSTM model experiences a sharp rise in training
accuracy within the first few epochs, reaching close to 99%.
Figure 2 shows that validation accuracy follows a similar
trajectory, stabilizing alongside training accuracy, which
indicates minimal overfitting. The high accuracy achieved
suggests that the model has effectively learned and
generalized well to unseen data.

Fig. 4 Training and validation accuracy of the

CNN model over multiple epochs.

The confusion matrix in Figure 5 illustrates the CNN

model's classification performance. It correctly identified
153 true negatives and 136 true positives, with only 11
misclassifications (7 false positives and 4 false negatives).
Fig. 2 Training and validation This demonstrates that the model performs well with high
accuracy of the LSTM model over multiple epochs. accuracy and minimal errors.

The confusion matrix (Fig. 3) shows that our LSTM model

accurately predicted 159 true negatives and 137 true
positives, making only 4 misclassifications (1 false positive
and 3 false negatives). This indicates that the model
performs well with high accuracy and minimal errors.
 Fig. 5 Confusion matrix of the CNN
model showing actual vs. predicted classifications Fig. 5 Confusion matrix of the RNN
model showing actual vs. predicted classifications
2) RNN Accuracy Curve and Confusion Matrix:
Our tests revealed that deep learning systems especially
LSTM models excel at recognizing patterns of depression
The accuracy of an RNN model over 20 epochs is displayed early in students. These models allow subtle changes
in Figure 6, where both training and validation accuracy rise detectable in depression to go unnoticed by traditional
quickly. The validation accuracy stays marginally higher at methods by analyzing a range of behavioral and
roughly 0.97, while the training accuracy levels out at
demographic factors. Deep learning provides a promising
around 0.95. This suggests that there is little to no
remedy for educational institutions to intervene and offer
overfitting and the model learns effectively.
help at the right time to reduce school depression's negative
impact on students' academic performance and well-being
points of view.

VI. CONCLUSION
As a result, this project demonstrates the potential of deep
learning methods in detecting students’ depression at an
early stage. Analyzing a variety of behavioral and
demographic factors such as academic pressure, sleep
patterns, financial stress, and so forth, the system was able
to predict with high accuracy students who may get
depressed. Among deep learning models Among deep
learning models, Long Short-term memory (LSTM) got the
highest performance results with 99% accuracy. It
performed better than convolutional neural networks (CNN)
and recurrent neural networks (RNN) which could detect
depression also. The results indicate that those models can
detect complex patterns and subtle behavioral changes that
Fig. 6 Training and validation accuracy of the may imply depression but where conventional means may
fail. The significance of early intervention, and how
CNN model over multiple epochs. technology enables the delivery of mental health support in
educational institutions is highlighted by this project. The
The provided RNN model's confusion matrix, which
proposed system stands to be scalable, non-invasive, and
displays both accurate and inaccurate predictions, is shown
potent in its efficacy for depression detection, ultimately
in Figure 7. With few misclassifications, the model
relieving students from the burden of dealing with mental
accurately classifies 159 instances of class 0 and 134
health issues. Timely insights from this system can help
instances of class 1. This suggests that the RNN model has
improve the support and resources available to students who
very few errors and a high accuracy.
require the help. This approach proves successful and can
be built out into future mental health monitoring
improvements, ultimately leading to better overall student
well-being and success.
REFERENCES
[1] S. Gaikwad, S. Dangat, S. Hande, D. Kawade, and R. M.
Shelake, "Student depression estimation through deep
learning," Int. Res. J. of Modern Engineering and [8] N. Chauhan, S. Swati, and D. Rani, "Depression
Technology, vol. 6, no. 3, pp. 15-20, Mar. 2024. [Online]. detection using ML," Int. J. Adv. Eng. Manag., vol. 5, no. 5,
Available: www.irjmets.com. [Accessed: Jan. 5, 2025]. pp. 621-626, May 2023.
[2] F. M. Shah, F. Ahmed, S. K. Saha Joy, S. Ahmed, S. [9] W. A. Gadzama, D. Gabi, M. S. Argungu, and H. U.
Sadek, and R. Shil, "Early depression detection from social Suru, "The use of machine learning and deep learning
network using deep learning techniques," IEEE models in detecting depression on social media: A
Transactions on Engineering in Medicine and Biology systematic literature review," Review Article, vol.
Society, vol. 2020, pp. 1-6, Nov. 2020, doi: 45100125, July-August 2024.
10.1109/TENSYMP50017.2020.9231008. [10] S. Hegde et al., "Depression Detection Using Machine
[3] A. Amanat, M. Rizwan, A. R. Javed, and M. Abdelhaq, Learning," NCRTCA – 2023, Volume 11, Issue 06, Paper
"Deep learning for depression detection from textual data," ID: NCRTCA-PID-027, 2023.
Electronics, vol. 11, no. 5, p. 676, May 2022, doi: [11] Z. Elaskar et al., "Depression Detection System,"
10.3390/electronics11050676. IJNRD, Volume 9, Issue 5, May 2024, ISSN: 2456-
[4] B. G. Bokolo and Q. Liu, "Deep learning-based [12] Ch. L. Susanna et al., "Depression Detection based on
depression detection from social media: Comparative Speech format with Deep learning and Advanced Voice
evaluation of ML and transformer techniques," Electronics, processing Techniques," IJRTI, Volume 9, Issue 3, 2024,
vol. 12, no. 21, p. 4396, Nov. 2023, doi: ISSN: 2456-3315.
10.3390/electronics12214396. [13] Dr. B. Gopinathan et al., "Depression Detection from
[5] W. C. de Melo, E. Granger, and A. Hadid, "Depression Facial Behaviour through Deep Learning Using MTCNN
detection based on deep distribution learning," in Proc. Algorithm," JETIR, Volume 10, Issue 3, March 2023, ISSN:
IEEE Int. Conf. Image Processing (ICIP), 2019, pp. 3067- 2349-5162.
3071, doi: 10.1109/ICIP.2019.8803467. [14] Peter J. Yu et al., "A Machine Learning Method for
[6] F. A. Nazira and S. R. Das, "Depression detection using Detecting Depression Among College Students,"
convolutional neural networks," in Proc. 2021 IEEE Int. International Journal of Computer Applications, Volume
Conf. Signal Processing, Information, Communication & 185, Issue 24, July 2023, pp. 44-51, DOI:
Systems (SPICSCON), 2021, pp. 1-5, doi: 10.5120/ijca2023923003.
10.1109/SPICSCON54707.2021.9885517. [15] Hyewon Yoo et al., "Depression Detection Model
[7] S. Sofia, A. Malik, M. Shabaz, and E. Asenso, "Machine Using Multimodal Deep Learning," Preprints, May 9, 2023,
learning based model for detecting depression during doi:10.20944/preprints202305.0663.v1.
Covid-19 crisis," Sci. Afr., vol. 20, e01716, May 2023, doi:
10.1016/j.sciaf.2023.e01716.