MINDMETRICS -Mental Health Text Classification Using BERT with Sentiment Fusion and Explainable AI

Abstract

Mental health disorders such as depression, anxiety, bipolar disorder, post-traumatic stress disorder (PTSD), suicidal ideation, and Stress have become a significant global concern, affecting millions of individuals. With the increasing expression of mental health issues on social media platforms, automated text classification has emerged as a promising approach for early detection and intervention. This study presents a novel frame-work for mental health text classification that leverages a fine-tuned Bidirectional Encoder Representations from Transformers (BERT) model combined with sentiment fusion and explainable Artificial Intelligence techniques. The proposed system integrates contextual embeddings generated by BERT with sentiment fea-tures derived from TextBlob to enhance classification perfor-mance. Additionally, gradient-based attribution methods and SHAP (SHapley Additive exPlanations) are employed to improve model interpretability. The model classifies input text into seven categories: Normal, Anxiety, Depression, Suicidal, Bipolar, PTSD, and Stress, and further evaluates the severity of conditions across three levels: Initial, Moderate, and Severe. A user-friendly inter-face is developed using Gradio to facilitate real-time interaction and analysis. Experimental evaluation on a Reddit-based mental health dataset demonstrates that the proposed approach achieves an accuracy of 91.7 and a macro-average F1-score of 0.893, indicating its effectiveness compared to existing methods. The integration of classification, sentiment analysis, severity detection, and explainability makes the proposed system a comprehensive solution for mental health monitoring.