DEEP LEARNING MODELS FOR EARLY DETECTION OF DEPRESSION FROM SOCIAL MEDIA TEXT

Abstract

Depression is a major global health concern. Recent studies suggest that language used on social media may reveal early signs of psychological distress before individuals seek formal clinical support. The objective of this study was to develop and evaluate transformer-based natural language processing models for detecting depressive risk signals in long-form Reddit posts using the ERISK2025 dataset. A traditional TF-IDF representation combined with Logistic regression served as the baseline and was compared with BERT-base and the emotion-aware Emotion-RoBERTa model under identical experimental conditions. Model performance was assessed using accuracy, risk-class recall, macro-F1, AUC, and confusion-matrix-based error distribution. Emotion-RoBERTa achieved the highest overall accuracy (0.76) and the strongest sensitivity to high-risk users with a risk-class recall of 0.91 and the lowest false-negative rate among the tested models. BERT-base showed moderate performance. The TF-IDF baseline was weakest at identifying high-risk cases. These findings indicate that emotion-aware language representations can improve early detection of depression risk
in social media text and may identify vulnerable users more reliably than traditional text-representation methods. The results also indicate that model evaluation in this setting should rely on clinically meaningful metrics rather than accuracy alone. Future work should consider larger datasets together with temporal patterns in user posts and ethical constraints related to real-world mental health screening.