Revealing Hidden Pain A Comparative Analysis of Traditional Versus New Deep Learning Approaches for Detecting Depression on Social Media

Depression is a pervasive mental health disorder often undetected due to social stigma and limited access to professional care. This study presents a comprehensive hybrid deep learning framework for early depression detection from Reddit textual data. Leveraging a cleaned dataset of 7,731 posts, the pipeline integrates traditional machine learning (Random Forest, SVM) with novel deep architectures: HSAN (Hybrid Self-Attention Network), EViT-CMF (Enhanced Vision Transformer with Cross-Modal Fusion), and UAME (Uncertainty-Aware Monte-Carlo Ensemble).

Textual features are extracted using TF-IDF (200 dims) and Word2Vec embeddings (200 dims), fused into 400-dimensional vectors. EViT-CMF further transforms embeddings into 32×32 multi-channel images for cross-modal learning. The UAME ensemble achieves superior performance with 93.93% accuracy and 0.9697 AUC, outperforming baselines. SHAP interpretability and uncertainty quantification enhance clinical trustworthiness. A user-friendly Flask-based web application with MySQL backend enables real-time predictions, supporting multiple models and user authentication. This end-to-end system demonstrates the potential of hybrid AI for scalable, accessible mental health screening.