Mental Health Monitoring from Social Media Using Emotion-Aware NLP Models

Abstract: Mental health disorders are on the rise globally, and early detection through non-invasive means has become a critical public health need. Social media platforms serve as real-time expressions of public emotions and personal states, offering a rich source for mental health assessment. This study explores the use of emotion-aware natural language processing (NLP) models, including RoBERTa, BERTweet, and DistilBERT, for detecting signs of depression, anxiety, and emotional distress from Twitter and Reddit posts. Using fine-tuned models on annotated mental health datasets, we evaluate performance across emotion classification, risk scoring, and early warning detection. Regression and predictive analyses show strong correlations (R² > 0.85) between linguistic features (e.g., affective valence, negation frequency) and mental health indicators. Interpretability tools such as SHAP and LIME validate the decision-making process of the models, ensuring transparency and reliability. Keywords: Mental Health, Social Media, Emotion Detection, NLP, BERTweet, RoBERTa, SHAP, LIME, Depression Detection, Anxiety Classification