A Novel Multi-Modal Framework for Sentiment Driven Depression Intensity Assessment

Abstract

Introduction: Depression is a mental illness that can cause a low mood, loss of interest in doing things, and sudden behavioral changes. Major depression may lead to suicide. It is the life threatening problem to be addressed immediately. According to the World Health Organization (WHO), about 280 million people worldwide have depression, including 5% of the world’s adults and 5.7% of adults above age 60. Once a person diagnosed with depression treatment continue by attending therapy sessions. All these traditional approaches will work on those who already diagnosed with the depression. Through the social media platforms vast data is generating, from the content taken from theses platforms by applying Large Language Models, which works on contest based rather than key based would give better results. So a preventive mechanism of finding depression severity from the text is a novel idea for earlier prediction.

Objectives: The main objective of the paper is to identify the depression intensity from the social media posts by applying the LLMs to find the behavioral and emotional patterns of posts. Identification of negative sentiments from the posts by using LLMs. PHQ-9 Questionnaire is used to find the t The LLM based models identifies the negative sentiments based on polarity. From these posts the seed terms were identified based PHQ-9 questionnaire. The next objective is to identify the intensity based the terms used in the posts.

Methods: This study proposes a multi-modal framework for depression intensity using NLP techniques. The data is extracted from twitter posted based on depression related hashtags. Preprocessing and feature extraction is done. Once data is ready after cleaning the transformer models like BERT, ALBERT techniques are applied fro depression severity estimation. Model is evaluated using deep learning models (BiLSTM, GRU) with accuracy, F1-score and loss estimation was evaluated.

Results: The bLSTM model achieved 92% accuracy, outperforming traditional methods. Transformer-based models (BERT, ALBERT) demonstrated improved classification performance.

Conclusions: The multi modal approach proved to be more accurate for earlier prediction purpose. The Same approach can be applied on various questionnaire based approaches for model evaluation. This approach can be used for self-assessment purpose.