Sentiment-enhanced multi-class Transformer model for efficient suicide detection in Reddit texts

This study presents a transformer-based framework for suicide risk stratification using Reddit data, addressing limitations in existing binary classification approaches and emotional nuance modeling. Leveraging 70,000 posts from mental health subreddits as weakly supervised risk indicators, we integrate VADER sentiment tokens directly into model inputs and implement hybrid text normalization to preserve platform-specific semantics. Six transformer architectures were evaluated, with BERT, RoBERTa, and ELECTRA achieving 94.27% accuracy in multi-class risk categorization. Key innovations include explicit emotion-content associations through prepended sentiment labels ([VERY NEG] to [VERY POS]) and efficient processing of noisy user-generated content. DistilBERT demonstrated optimal efficiency-accuracy balance with 94.16% accuracy score. The framework enables granular detection of emotional escalation patterns, offering computational feasibility for real-world deployment while maintaining clinical relevance through alignment with suicidality continua. Error analysis highlights persistent challenges in distinguishing semantically adjacent risk categories, informing future directions for context-aware mental health monitoring systems.