Depression Detection in Social Media

Using Emotional Intelligence

G. Thiricksha, CSE – [email protected]
N.C. Vandhanaa , CSE – [email protected]

Guided By
Mrs. S. Padmapriya, Professor & Head, CSE
A.V.C College of Engineering Mannampandal, Mayiladuthurai

1. Introduction
Abstract: This project introduces a novel Depression is a prevalent and serious
approach to detecting depression using mental disorder affecting 280 million
multimodal deep learning, integrating people worldwide, which extensively
convolutional neural networks (CNNs) for influences an individual’s quality of life.
image analysis and long short-term This significant public health concern
memory (LSTM) networks for textual impacts an individual’s physical and
understanding. Implemented in MATLAB, mental welfare in various aspects, such as
the system leverages emotional alterations in appetite, diminished
intelligence from social media content to motivation and interest, irregular sleep
capture nuanced indicators of depression. patterns, and in severe instances,
By analyzing both images and text, the contemplation of suicide. Early diagnosis
model aims to provide a comprehensive of depression and treatment can prevent
understanding of users emotional states, patients’ conditions from worsening.
offering a promising avenue for early
intervention and support. Two methods are Major Depressive Disorder (MDD), also
employed: live face-based stress level referred to as depression, is a globally
detection and text-based stress level common mental disorder and causes
detection. This approach addresses the physical and mental health damage to
limitations of traditional detection methods hundreds of thousands. Unlike the usual
by harnessing the rich emotional cues mood swings and transient emotional
present in social media, thereby responses to the challenges of daily life,
contributing to the mitigation of depression will bring pervasive low mood,
depression's escalating burden on lack of confidence, and loss of pleasure or
individuals and society. interest in activities for most of the day.
Moreover, depression could increase the
Keywords: Chatbot, Government risks of diabetes, heart disease, and cancer,
Schemes, NLP, Flask Server, Python andin serious cases, it can lead to suicide.
At present, there are effective techniques. By amalgamating
psychological and pharmacologic convolutional neural networks (CNNs) for
treatments like antidepressants, Dialectical image analysis and long short-term
Behavior Therapy (DBT), and Cognitive memory (LSTM) networks for textual
Behavior Therapy (CBT). However, due to comprehension, the model endeavors to
the lack of medical resources, trained discern subtle indicators of depression
healthcare workers, and social prejudice present in user-generated content.
against mental disorders, people with Executed within the MATLAB
depression are often not properly environment, this system aims to offer a
diagnosed and treated. More specifically, promising avenue for early identification
across all countries at different income and intervention, thus alleviating the
levels, people with depression are often mounting burden of depression on both
underdiagnosed, and even non-depressed individuals and society at large.
people are often misdiagnosed and
2. Literature Review
prescribed antidepressants. Given the high
incidence of depression and the lack of The literature review of this paper is
appropriate treatments for large divided into three subsections as per the
populations, methods that rely solely on detection of emotions concerning different
subjective assessment and diagnosis are no sources. The first sub-section discusses the
longer able to meet current medical needs. studies conducted to detect depression
Therefore, automatic detection of through sentiment analysis of Twitter
depression will be very helpful and tweets. The second sub-section converses
necessary for the diagnosis of this mental about the detection of depression using
disorder. facial expression (image and video
processing). The last sub-section deals with
Some clinical literature has shown that
the use of chatbots, emotional AI, and
depressive states can be expressed from
combined inputs (text, audio, image, and
facial expressions. Compared with the
video) for detecting depression. All these
healthy subjects, the facial expressions of
sources to detect depression are discussed
depressed subjects are usually neutral or
in terms of various machine learning
sad expressions, which are characterized
techniques. In recent years, there has been
by frowning, drooping eyes, and looking
a growing interest in the use of artificial
tired or worried. Several studies have
intelligence (AI) and machine learning
attempted to automatically detect
(ML) to improve mental health care. As
depression through facial information.
Shikha et al. (2023) discussed, AI and ML
Depression represents a significant public can be used to detect and diagnose mental
health challenge globally, with health conditions, develop AI-powered
conventional detection methods often interventions, and improve access to mental
falling short in capturing the intricate health care services. There has been a
emotional nuances embedded within social growing body of research exploring the
media content. To bridge this gap, this detection of depression from social media
study proposes a pioneering approach data, particularly utilizing machine
utilizing multimodal deep learning learning techniques. This section provides
an overview of key studies and machine learning framework that
methodologies in the field, highlighting incorporated sentiment analysis features to
the advancements made in detecting predict depression levels in individuals.
depression through user tweets. Negative Their findings highlighted the importance
comments or expressions of pessimism are of sentiment analysis in capturing
often associated with depressive emotional states and identifying signs of
tendencies. Research studies have explored depression. The system will also be able to
the link between negative language use detect emotions by chatting with a chatbot.
and depression, providing evidence to
support the statement. The study focused on detecting depression
using social media data and machine
In a study conducted, the researchers learning employed various text
analyzed social media data and found a classification algorithms, including
significant correlation between the Support Vector Machines (SVMs) and
language used in tweets and the prevalence random forests, to classify tweets as
of depression symptoms. They identified depressive or non-depressive. As
that individuals with higher levels of explained by Kim (2017), SVMs work by
depression were more likely to express finding a hyperplane in the data that
negative sentiments in their tweets. separates the two classes (depressed vs.
not depressed) with the maximum margin.
Another study investigated the association The study achieved promising results in
between language markers and depression terms of classification accuracy,
on social media platforms. They found that demonstrating the potential of machine-
individuals with depressive symptoms learning approaches for depression
tended to use more negative language, detection. While the study demonstrated
indicating a correlation between negative effective depression detection from social
expression and depression. media data, it primarily focused on
Gkotsis et al.employed informed deep- traditional machine-learning algorithms.
learning techniques to characterize mental Incorporating more advanced deep
health conditions in social media. They learning models such as recurrent neural
utilized a large-scale dataset of Twitter networks or transformers could potentially
posts and applied deep learning algorithms improve the performance and capture
to detect mental health conditions, complex patterns within the tweet data.
including depression. Their approach Overall, these studies collectively
showcased the potential of leveraging deep demonstrate the potential of using machine
learning models to gain insights from user- learning techniques for detecting
generated content and improve mental depression from user tweets. By analyzing
health monitoring. linguistic patterns, social interactions, and
Moreover, the study by Resnik et al. contextual information, researchers have
explored the role of sentiment analysis and made strides in developing computational
linguistic markers in detecting depression models capable of identifying individuals
from Twitter data. They developed a at risk of depression.
 3. Methodology optimization techniques (e.g., Adam,
Data Collection: RMSprop) to train the model
Gather a diverse dataset of social media posts efficiently.Regularize the model to prevent
(e.g., Twitter, Reddit) with a mix of depressed overfitting, using techniques such as dropout
and non-depressed individuals. Ensure the and batch normalization.Tune
dataset is labeled for depression status. Extract hyperparameters through cross-validation or
additional contextual information such as user grid search.
demographics, posting frequency, time of Evaluation:
posting, etc., to enrich the dataset. Evaluate the model's performance using
Preprocessing: metrics such as accuracy, precision, recall, F1-
Clean the text data by removing noise (HTML score, and area under the ROC curve
tags, special characters, URLs) and (AUC).Conduct thorough error analysis to
standardizing the text (lowercasing, stemming, understand the model's weaknesses and areas
lemmatization). Tokenize the text into words for improvement. Validate the model's
orsubwords to represent them generalizability on the test set and assess its
numerically. Perform embedding (e.g., robustness to unseen data.
Word2Vec, GloVe) to convert text into dense Deployment and Monitoring:
vectors that capture semantic meaning. Deploy the trained model in a suitable
environment, such as a web application or
Feature Engineering: API, to enable real-time depression
Utilize emotional intelligence metrics to detection in social media posts. Implement
extract emotional features from the text, such monitoring mechanisms to track model
as sentiment scores, emotion distribution (e.g., performance over time and ensure its efficacy
joy, sadness, anger), and linguistic style in real-world scenarios. Continuously update
indicators (e.g., pronoun usage). Combine and refine the model based on feedback and
textual embeddings with emotional features to new data to
create a rich representation for each post. enhance its accuracy and reliability.
Model Architecture:
4. Proposed System
Design a hybrid neural network architecture
combining CNN and LSTM layers to capture The proposed system presents a novel
both local and global dependencies in the text framework that integrates live face-based
data. Use CNNs for feature extraction to stress level detection and text-based stress
identify patterns and relevant features from the level detection using multimodal deep
textual embeddings. Employ LSTM networks learning techniques. The proposed system
to model the sequential dependencies in the for depression detection in social media
text data and capture long-term dependencies. leverages a sophisticated fusion of
Concatenate or merge the outputs from CNN emotional intelligence, Convolutional
and LSTM layers to combine their respective Neural Networks (CNNs), and Long Short-
strengths. Term Memory (LSTM) networks. By
Training: harnessing the rich emotional context
Split the dataset into training, validation, and embedded within social media posts, the
test sets to evaluate model performance. Train system aims to discern patterns indicative
the model using the training set, optimizing it of depression with high accuracy. Through
for depression detection. Utilize appropriate
loss functions (e.g., binary cross-entropy) and
an innovative hybrid architecture, LSTM 5. Architecture
layers capture the nuanced sequential
dependencies in text data, while CNN
layers extract local features, ensuring a
holistic understanding of users' emotional
expressions. Integrating emotional
intelligence metrics further enriches the
model's comprehension, enabling it to
discern subtle shifts in sentiment and
emotion distribution. By deploying this
system, we envision a proactive approach
to mental health support, providing timely
interventions and resources for individuals
in need within online communities, while
upholding privacy and ethical standards. 6. Results

The result of implementing the proposed

system for depression detection in social
media using LSTM and CNN in emotional
intelligence is a robust and effective tool
for identifying individuals at risk of
depression within online communities.
Through rigorous training and validation,
the model demonstrates high accuracy in
discerning patterns indicative of
depression, leveraging both the semantic
meaning of the text and the emotional
Leveraging CNNs for image analysis and
nuances expressed within it. Real-time
LSTMs for textual understanding, the
deployment of the system enables timely
system aims to capture nuanced indicators
interventions and support for individuals in
of depression within user-generated
distress, facilitating access to mental health
content. By employing advanced machine
resources and professional assistance when
learning algorithms, this approach offers a
needed most. Ethical considerations are
comprehensive analysis of emotional
carefully addressed to ensure user privacy
states, facilitating early identification and
and transparency, while continuous
intervention for individuals at risk of
improvement mechanisms allow for
depression. The system operates within the
iterative refinement of the model over
MATLAB environment, providing a user-
time. Overall, the result is a proactive
friendly interface for seamless
online space, that enhances awareness,
implementation and deployment. Through
intervention, and support for individuals
its multimodal approach, the proposed
experiencing depression.
system addresses the limitations of
existing detection methods, offering
enhanced accuracy and efficiency in
detecting depression from social media
data.
 7. Discussion cure for it. In the future, these approaches
Depression is becoming an epidemic can be integrated into a vast system to
disease that affects people of all social clinically categorize patients suffering
classes, cultures, and countries frequently. from depression based on discovering their
Due to the inherent nature of solitude, emotional profiles. To conclude,
finding those who seek help because of depression, mood, and emotion can be
mental conditions but are unable to detected through text, images, videos,
express their desire is difficult, and often speech, gestures, etc. through different AI
goes overlooked even by those who are and ML techniques.
depressed. Since textual sentiment analysis
is a non-invasive technique that can be 8. Future Enhancements

constantly monitored and controlled, it can In future enhancements for depression

help diagnose the disease. This is a big detection in social media using emotional
support in the combat against depression intelligence, several avenues promise to
since it allows us to distinguish times of advance the field significantly. Integrating
happiness and melancholy without visiting additional modalities like images, videos,
a psychologist, allowing us to fight back and audio into analysis would provide a
quickly when needed. Depression is more comprehensive understanding of
detected, analyzed, and prevented through users' emotional expressions. This
Twitter sentiment analysis (text multimodal approach could enrich the
processing), facial expressions (image and system's context comprehension, improving
video processing), and the use of chatbots, detection accuracy. Moreover, enhancing
emotional AI, and combined inputs (text, the system's ability to interpret contextual
audio, image, video). Various artificial nuances such as sarcasm and irony would
intelligence and machine learning refine its understanding of complex
techniques like Naïve-Bayes, LSTM– linguistic constructs.
RNN, Logistic Regression, Linear Support Personalized intervention strategies based
Vector, PCA, KNN Classification, etc. are on individual emotional profiles and
being used to recognize emotions and preferences could be developed, tailoring
hence to detect depression. Efficiency and support resources and recommendations to
performance of various algorithms like unique needs. Real-time monitoring
SVM, and Multinomial NaïveBayes are capabilities could offer immediate
analyzed to see which works better to feedback and crisis intervention resources
detect emotions and hence depression to distressed users. Longitudinal studies
through tweets. Interactive tracking mental health trajectories over
technology-driven AI-based solutions are time could provide insights into
also discussed. For example, a Chatbot intervention effectiveness and factors
detects depression and responds with a influencing outcomes. Cross-platform
joke or song to lighten the mood of the
integration would extend analysis across
user. Such kind of emotional AI and ML-
various platforms, offering a more holistic
based solutions can prove to be beneficial
view of individuals' online behavior.
in detecting, analyzing, and preventing
Interdisciplinary collaboration between
depression and also provide a
 researchers, mental health professionals, and Beck Depression Inventory-II (BDI-II), center
technology developers would address for epidemiologic studies depression scale
complex challenges. Empowering users with (CES-D), geriatric depression scale (GDS),
education and tools to manage mental health Hospital Anxiety and depression scale
online, along with implementing measures to
(HADS), and Patient Health Questionnaire-9
mitigate biases and ensure fairness, would
foster a responsible and ethical approach to (PHQ-9),” Arthritis Care Res., vol. 63, no.
depression detection in social media 11, pp. 454–
466, Nov. 2011.

References
[7] R. P. N. Rao, Brain-Computer Interfacing:
[1] L. Zhou, Z. Liu, Z. Shangguan, X. Yuan, Y.
Li, and B. Hu, "TAMFN: Time-Aware An Introduction. U.K.: Cambridge Univ.
Attention Multimodal Fusion Network for Press, 2011, 3.1.12, p. 26, doi:
Depression Detection," in IEEE Transactions on 10.1017/CBO9781139032803.
Neural Systems and Rehabilitation Engineering, [8] M. Wang and J. Y. Chen, “A GMM-IG
vol. 31, pp. 669-679, 2023, doi: framework for selecting genes as expression
10.1109/TNSRE.2022.3224135.
panel biomarkers,” Artif. Intell. Med., vol. 48,
[2] Z. Shangguan, Z. Liu, G. Li, Q. Chen, Z. nos. 2–3, pp. 75–82, Feb. 2010, doi:
Ding, and B. Hu, "Dual-Stream Multiple 10.1016/j.artmed.2009.07.006.
Instance Learning for Depression Detection
With Facial Expression Videos," in IEEE [9] A. Kolahi, M. Hoviattalab, T. Rezaeian, M.
Transactions on Neural Systems and Alizadeh, M. Bostan, and H. Mokhtarzadeh,
Rehabilitation Engineering, vol. 31, pp. 554-
“Design of a marker-based human motion
563, 2023, doi: 10.1109/TNSRE.2022.3204757.
tracking system,” Biomed. Signal Process.
[3] A. Whiston, E. R. Igou, D. G. Fortune, Control, vol. 2, no. 1, pp. 59–67, Jan. 2007,
Analog Devices Team, and M. Semkovska, doi: 10.1016/j.bspc.2007.02.001.
"Examining Stress and Residual Symptoms in
Remitted and Partially Remitted Depression
Using a Wearable Electrodermal Activity
Device: A Pilot Study," in IEEE Journal of
Translational Engineering in Health and
Medicine, vol. 11, pp. 96-106, 2023, doi:
10.1109/JTEHM.2022.3228483.

[4] W. C. de Melo, E. Granger, and M. B.

López, "MDN: A Deep Maximization-
Differentiation Network for Spatio-Temporal
Depression Detection," in IEEE Transactions on
Affective Computing, vol. 14, no. 1, pp. 578-
590, 1 Jan.-March 2023, doi:
10.1109/TAFFC.2021.3072579.

[5] N. K. Iyortsuun, S.-H. Kim, H.-J. Yang, S.-

W. Kim, and M. Jhon, "Additive Cross-Modal
Attention Network (ACMA) for Depression
Detection Based on Audio and Textual
Features," in IEEE Access, vol. 12, pp. 20479-
20489, 2024, doi:
10.1109/ACCESS.2024.3362233.
[6] K. L. Smarr and A. L. Keefer, “Measures
of depression and depressive symptoms: