DEPRESSION DETECTION USING DEEP LEARNING MODELS BASED ON MULTIMODAL SOCIAL MEDIA CONTENT

Abstract

The rapid digital transformation and extensive adoption of social media platforms have revolutionized the advancements in mental health monitoring and depression detection. Social media platforms like Twitter, Facebook, Instagram, Reddit, etc., has firmly established itself as an indispensable part of life for the majority of the population nowadays. The constant presence of users on these platforms provides a rich user generated content that can be leveraged to monitor the mental health in comparison to traditional clinical settings. Utilizing advancements in artificial intelligence, natural language processing and computer vision, researchers and clinical mental health experts can detect early signs of depression by analysing text, audio, video, image and emoticons content generated by users on social media platforms. This thesis presents a novel deep learning-based frameworks for depression detection using the user generated English multimodal social media content. Initially a dataset was created for the same as one of the main constraints was the non-availability of the multimodal dataset for depression detection. Additionally, a framework was presented that is a combination of bidirectional encoder representations from transformers and convolutional neural networks. This model was designed to detect the depressive posts using the created dataset, also a new model is presented to detect the severity of the post using publicly available dataset. The research further introduces a deep learning-based framework for depression detection for hindi and hinglish (code-mixed) dataset. To overcome the challenge of regional diversity a hindi and hinglish language-based dataset was created from openly available social media platforms. This model was a combination of bidirectional encoder representations from transformers and particle swarm optimization based optimized convolutional neural network. A comprehensive experimental evaluation is conducted to assess the performance and scalability of the proposed frameworks. The Comparative analyses with existing methodologies are carried out to demonstrate improvements in detection accuracy, efficiency, and overall system robustness. It provides a practical foundation for mental health monitoring and secure healthcare applications in real-world deployments.