A Self-Correcting Agentic Graph RAG for Clinical Decision Support in Hepatology

Yalan HuWenjie XuanQingqing ZhouZhi LiYa LiJili HuFang Fang^*

• First Affiliated Hospital of Anhui University of Traditional Chinese Medicine, Hefei, China

Clinical decision-making in hepatology is challenged by the rapid growth of medical knowledge and the unreliability of Large Language Models (LLMs), which are prone to hallucination. Standard Retrieval-Augmented Generation (RAG) often fails to leverage complex medical knowledge structures. To address this, we propose an Agentic Graph RAG framework built upon a clinically-verified hepatology knowledge graph. Our state-driven agentic system employs a self-correcting "retrieve-evaluate-refine" loop, where agents dynamically generate, semantically validate, assess, and iteratively optimize graph search strategies to build a comprehensive and accurate context. This final context is then used by an LLM to generate a reliable response. Evaluated on a custom dataset of clinical questions, our framework significantly outperformed baseline models (GPT-4, standard RAG, Graph RAG) across all metrics, achieving superior faithfulness (0.94), context recall (0.92), and answer relevancy (0.91). This agentic approach mitigates LLM hallucinations and provides accurate, interpretable answers, demonstrating its potential as a next-generation intelligent clinical decision support tool.