BERT-Based Mental Health Monitoring System: Early Detection of Depression Indicators through Social Media Text Analysis

Abstract

This paper introduces a novel BERT-based mental health monitoring architecture for automated depression detection from social media text, addressing SDG 3: Good Health and Well-being. This paper implements a comprehensive comparative analysis of four different transformer models: BERT-base-uncased, RoBERTa-base, MentalBERT, and DistilBERT rather than the traditional classifier-based approaches. The proposed model is fine-tuned using transfer learning on more than 50,000 labeled social media posts from Reddit mental health communities and Twitter datasets. This work utilizes sophisticated preprocessing methods such as WordPiece tokenization, class balancing using SMOTE, and stratified data splitting. The proposed model design includes bidirectional self-attention with multi-head attention layers, followed by custom classification heads with dual dropout regularization and dense layer with ReLU activation for four-level depression severity prediction. The interpretability of the model is guaranteed using LIME (Local Interpretable Model – agnostic Explanation) and attention visualization to detect psycholinguistic features including first-person pronoun frequency, absolutist language patterns and temporal expression shifts. The training process uses the AdamW optimizer with learning rate 2e-5, batch size 16, and early stopping mechanism. The novel BERT-based mental health monitoring architecture aims to achieve over 85% accuracy with high recall rates critical for healthcare applications, identification of domain-specific linguistic depression indicators it also incorporates ethical AI practices and enables scalable and privacy-conscious early mental health intervention and population-level depression surveillance.