SENTINEL-LLM: Suicide Ensemble-Based Text Intelligence and Natural Language Evaluation Through Large Language Models

Detecting suicidal ideation in social media posts is a critical challenge in mental health intervention. This study introduces a novel suicide-focused Frequent-Rare (FR) Dictionary, specifically designed to enhance the sensitivity of Large Language Models (LLMs) to suicide-related linguistic patterns. Unlike conventional fine-tuning approaches, our methodology integrates domain-specific lexicons, leveraging TF-IDF weighting and sentiment-based adjustments to optimize LLMs for suicide risk detection. We fine-tune Qwen2.5, LLaMA3, and GPT-NEO with this enriched vocabulary, achieving remarkable performance improvements over their baseline versions. For instance, Qwen2.5's accuracy leaps from 30.42% to 92.93%, while GPT-NEO improves from 63.10% to 93.52%, demonstrating the profound impact of targeted linguistic augmentation. This advancement not only significantly boosts detection accuracy but also enhances model recall and precision, ensuring more reliable identification of high-risk cases. Our findings establish domain-aware finetuning as a powerful strategy for suicide risk assessment, highlighting the necessity of integrating specialized lexicons into mental health AI models.