Scholarlink

In high-stakes domains like medicine, AI systems must do more than generate fluent responses—they must ensure that clinical reasoning is safe, auditable, and grounded in verified knowledge. This paper presents a comparative analysis of reasoning outputs produced by large language models and by curated, ontology-grounded AI systems. We use KaiMed—a hybrid multi-agent platform built on GPT-4o-mini and structured medical knowledge—as a representative of the latter approach. Unlike single-model pipelines, KaiMed decomposes reasoning into modular, role-specific agents specialized in diagnosis, treatment planning, literature validation, and safety filtering. Each agent operates over a semantically structured knowledge base, comprising a proprietary clinical ontology and two curated layers: a graph of clinical trials and a semantic retriever built on peer-reviewed literature. This architecture enables traceable, constraint-aware, and clinically aligned reasoning. Our hypothesis is that in medicine, the source, structure, and semantic integrity of knowledge are not technical details—they are prerequisites for reliability and trust. We evaluate both systems on complex prompts in two domains: Inflammatory Bowel Disease (IBD) and Chronic Rhinosinusitis with Nasal Polyps (CRSwNP). Results show that while GPT-4o-mini generates plausible responses, it lacks epistemic grounding and fails to surface non-obvious or investigational options. KaiMed, by contrast, consistently produces evidence-aligned, phenotype-specific recommendations—demonstrating that in Clinical AI, the key differentiator is not model size, but the structure and orchestration of knowledge.

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

Structured Clinical Reasoning in AI: Comparing LLMs and Curated, Ontology-Grounded Multi-Agent Systems

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Francesco Barbato,
  • Filippo Sorichetti,
  • Emanuele Ielo,
  • Lorenzo Megliola

摘要

In high-stakes domains like medicine, AI systems must do more than generate fluent responses—they must ensure that clinical reasoning is safe, auditable, and grounded in verified knowledge. This paper presents a comparative analysis of reasoning outputs produced by large language models and by curated, ontology-grounded AI systems. We use KaiMed—a hybrid multi-agent platform built on GPT-4o-mini and structured medical knowledge—as a representative of the latter approach. Unlike single-model pipelines, KaiMed decomposes reasoning into modular, role-specific agents specialized in diagnosis, treatment planning, literature validation, and safety filtering. Each agent operates over a semantically structured knowledge base, comprising a proprietary clinical ontology and two curated layers: a graph of clinical trials and a semantic retriever built on peer-reviewed literature. This architecture enables traceable, constraint-aware, and clinically aligned reasoning. Our hypothesis is that in medicine, the source, structure, and semantic integrity of knowledge are not technical details—they are prerequisites for reliability and trust. We evaluate both systems on complex prompts in two domains: Inflammatory Bowel Disease (IBD) and Chronic Rhinosinusitis with Nasal Polyps (CRSwNP). Results show that while GPT-4o-mini generates plausible responses, it lacks epistemic grounding and fails to surface non-obvious or investigational options. KaiMed, by contrast, consistently produces evidence-aligned, phenotype-specific recommendations—demonstrating that in Clinical AI, the key differentiator is not model size, but the structure and orchestration of knowledge.