SentinelGPT: A Large Language Model for Detecting Suicide and Depression Tendencies among Adolescents Based on the Micro-blog’s “Tree Hole” and GraphRAG

Authors:
  • SWShaojie Wang
    Shaojie Wang
  • ZHZhisheng Huang
    Zhisheng Huang
  • HSHaoru Su
    Haoru Su
  • CGChaohui Guo
    Chaohui Guo
Open Access
Publisher website
Google Scholar
Abstract
In recent years, suicide rates among Chinese adolescents have risen markedly. Moreover, the onset of suicidal behaviour has shifted toward increasingly younger age groups. As a result, expressions of suicidal ideation on social media have become an important public health concern. Although the deep integration of natural language processing and large language models offers new opportunities for adolescent mental-health detection, significant technical challenges remain in terms of model generalizability and detection accuracy. To address these challenges, we propose SentinelGPT, a large language model designed to detect suicidal and depressive tendencies in adolescents. First, we construct a Chinese social-media corpus for suicide and depression detection by performing multi-dimensional cleaning, screening, and preprocessing on original data collected from the Micro-blog “Tree Hole”. Second, we adopt a collaborative GraphRAG architecture that combines a knowledge graph with a vector database to enhance the model’s ability to integrate information across multiple dialogue turns and capture semantic relationships in complex contexts. Third, we fine-tune a dialogue-enhanced LLaMA 3.1 model using LoRA and deploy it privately via the Ollama framework to improve parsing of fragmented expressions, metaphorical language, and cross-platform contextual variations commonly found in online text. To further verify model effectiveness, we conduct comparative experiments against four representative baselines DeepSeek V3, Qwen 2.5, Fine-tuned Qwen 2.5 and ELECTRA using the same dataset and evaluation metrics including precision, recall, and F1-score. SentinelGPT achieves the highest overall F1-scores, outperforming all baselines and demonstrating superior robustness and generalization ability. The results demonstrate that SentinelGPT effectively detects suicidal and depressive tendencies in the complex context of Micro-blog comments, maintains semantic consistency across multi-turn conversations, and offers a reliable technical solution for adolescent mental-health monitoring and early warning.

This publication has 17 references indexed in Scilit: