Scholarlink

Background <p>Depression imposes substantial individual and societal burdens, yet more than 80% of cases remain undetected in non‑psychiatric settings, and approximately 15% of individuals with severe depression die by suicide. Existing screening methods are limited by subjective self‑report, fragmented care pathways, and stigma‑related underutilization.</p> Methods <p>We developed and locally deployed a Retrieval‑Augmented Generation (RAG) framework that integrates large language models (LLMs) with an evidence‑based clinical knowledge base for depression and suicidality. Three leading Chinese LLMs (DeepSeek, Qwen, Baichuan) were evaluated under base and RAG‑augmented configurations. The knowledge base comprised clinical practice guidelines, standard textbooks, and peer‑reviewed literature. Model performance was assessed on multi‑source datasets. Outcomes included classification of depression and suicide risk, consistency of diagnostic outputs, symptom‑summary quality, and clinician‑rated empathy and dialogue quality.</p> Results <p>RAG augmentation consistently improved diagnostic performance across datasets. For depression detection, Qwen3 + RAG achieved an F1‑score of 0.9107, and for suicide risk stratification, RAG enhanced F1‑scores by approximately 0.15 compared with non‑RAG baselines. The best‑performing configuration demonstrated high reproducibility for both depression diagnosis and suicide risk classification (overall κ = 0.93; 95% CI: 0.91–0.95). Symptom summarization was substantially improved with RAG, with relative gains of 44.1–207.2% in BLEU‑2 and 37.0–94.0% in ROUGE‑L. Clinician‑blinded ratings indicated clinician-evaluated high-level empathic expression in RAG‑enhanced models (mean scores 4.23–4.98/5), with Qwen3 + RAG exhibiting particularly strong emotional concordance (95% CI: 4.80–4.92). Dialogue quality was rated as high across domains, including fluency (mean 4.98/5) and perceived doctor‑likeness (mean 4.94/5).</p> Conclusions <p>A domain‑specific, RAG‑enhanced LLM framework can deliver clinically reliable, reproducible, and empathetic depression screening and suicide risk stratification in both real‑world and simulated settings. Local on‑premises deployment supports privacy and data‑governance requirements, while the scalable architecture and structured outputs position this approach as a promising tool for augmenting mental health triage and longitudinal monitoring in resource‑constrained environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Retrieval‑augmented large language models for depression screening and suicide risk stratification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月31日

查看全文
  • Wenbo Xie,
  • Xulai Song,
  • Zerui Lu,
  • Gaoqiang Fei,
  • Yujia Zhou,
  • Siqi Ding,
  • Fengyi Zuo,
  • Bangyu Wu,
  • Tianhao Gu,
  • Lin Xu,
  • Xiaomeng De,
  • Bin Zhu

摘要

Background

Depression imposes substantial individual and societal burdens, yet more than 80% of cases remain undetected in non‑psychiatric settings, and approximately 15% of individuals with severe depression die by suicide. Existing screening methods are limited by subjective self‑report, fragmented care pathways, and stigma‑related underutilization.

Methods

We developed and locally deployed a Retrieval‑Augmented Generation (RAG) framework that integrates large language models (LLMs) with an evidence‑based clinical knowledge base for depression and suicidality. Three leading Chinese LLMs (DeepSeek, Qwen, Baichuan) were evaluated under base and RAG‑augmented configurations. The knowledge base comprised clinical practice guidelines, standard textbooks, and peer‑reviewed literature. Model performance was assessed on multi‑source datasets. Outcomes included classification of depression and suicide risk, consistency of diagnostic outputs, symptom‑summary quality, and clinician‑rated empathy and dialogue quality.

Results

RAG augmentation consistently improved diagnostic performance across datasets. For depression detection, Qwen3 + RAG achieved an F1‑score of 0.9107, and for suicide risk stratification, RAG enhanced F1‑scores by approximately 0.15 compared with non‑RAG baselines. The best‑performing configuration demonstrated high reproducibility for both depression diagnosis and suicide risk classification (overall κ = 0.93; 95% CI: 0.91–0.95). Symptom summarization was substantially improved with RAG, with relative gains of 44.1–207.2% in BLEU‑2 and 37.0–94.0% in ROUGE‑L. Clinician‑blinded ratings indicated clinician-evaluated high-level empathic expression in RAG‑enhanced models (mean scores 4.23–4.98/5), with Qwen3 + RAG exhibiting particularly strong emotional concordance (95% CI: 4.80–4.92). Dialogue quality was rated as high across domains, including fluency (mean 4.98/5) and perceived doctor‑likeness (mean 4.94/5).

Conclusions

A domain‑specific, RAG‑enhanced LLM framework can deliver clinically reliable, reproducible, and empathetic depression screening and suicide risk stratification in both real‑world and simulated settings. Local on‑premises deployment supports privacy and data‑governance requirements, while the scalable architecture and structured outputs position this approach as a promising tool for augmenting mental health triage and longitudinal monitoring in resource‑constrained environments.