Scholarlink

Purpose <p>Artificial intelligence (AI), especially large language models (LLMs), faces a clinical trust gap due to black-box reasoning and factual hallucinations–risks amplified in otorhinolaryngology (ENT), where diagnostic pathways are complex and symptoms overlap. We aim to build an explainable and accurate ENT diagnostic assistant grounded in a dual-knowledge framework.</p> Methods <p>We construct a rule-centric <i>Core Knowledge Graph</i> (KG) distilled from peer-reviewed guidelines and handbooks, and a <i>Contextual Instance Library</i> (CIL) of real-world case narratives. At inference, the system performs dual retrieval–logical KG reasoning for citable rules and dense similarity search for experiential corroboration–then applies schema-constrained prompting to synthesize both streams into auditable reports. Offline, KG/CIL are automated with LLM-based NER/RE, discourse-aware chunking, and vector indexing. Evaluation uses a 251-item ENT question set and curated vertigo cases, compared against base LLMs and standard RAG baselines; explanation quality is assessed with LLM-as-a-judge across multiple generator/judge pairs.</p> Results <p>The dual system achieves higher diagnostic accuracy than base LLMs and standard RAG baselines and produces consistently superior, citation-backed explanations. Reports explicitly link conclusions to rules and corroborating instances.</p> Conclusions <p>Integrating rule knowledge (KG) with experiential evidence (CIL) simultaneously improves diagnostic accuracy and interpretability, offering a pragmatic path from black-box LLMs to trustworthy, auditable clinical decision support in ENT.</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

Explainable and accurate ENT diagnosis via a dual knowledge LLM

  • 期刊论文

  • 发表日期：

    2026年4月6日

查看全文
  • Jing Gong,
  • Hao Yuan,
  • Fuchun Li

摘要

Purpose

Artificial intelligence (AI), especially large language models (LLMs), faces a clinical trust gap due to black-box reasoning and factual hallucinations–risks amplified in otorhinolaryngology (ENT), where diagnostic pathways are complex and symptoms overlap. We aim to build an explainable and accurate ENT diagnostic assistant grounded in a dual-knowledge framework.

Methods

We construct a rule-centric Core Knowledge Graph (KG) distilled from peer-reviewed guidelines and handbooks, and a Contextual Instance Library (CIL) of real-world case narratives. At inference, the system performs dual retrieval–logical KG reasoning for citable rules and dense similarity search for experiential corroboration–then applies schema-constrained prompting to synthesize both streams into auditable reports. Offline, KG/CIL are automated with LLM-based NER/RE, discourse-aware chunking, and vector indexing. Evaluation uses a 251-item ENT question set and curated vertigo cases, compared against base LLMs and standard RAG baselines; explanation quality is assessed with LLM-as-a-judge across multiple generator/judge pairs.

Results

The dual system achieves higher diagnostic accuracy than base LLMs and standard RAG baselines and produces consistently superior, citation-backed explanations. Reports explicitly link conclusions to rules and corroborating instances.

Conclusions

Integrating rule knowledge (KG) with experiential evidence (CIL) simultaneously improves diagnostic accuracy and interpretability, offering a pragmatic path from black-box LLMs to trustworthy, auditable clinical decision support in ENT.