A self-correcting Agentic Graph RAG for clinical decision support in hepatology

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Hepatology

Background:

• Clinical decision-making in hepatology faces challenges due to expanding medical knowledge.
• Large Language Models (LLMs) exhibit unreliability and hallucination issues.
• Standard Retrieval-Augmented Generation (RAG) struggles with complex medical knowledge structures.

Purpose of the Study:

• To develop an advanced framework for reliable clinical decision support in hepatology.
• To overcome the limitations of existing LLMs and RAG paradigms in processing complex medical data.
• To enhance the accuracy and trustworthiness of AI-driven medical information systems.

Main Methods:

• Proposed an Agentic Graph RAG framework utilizing a clinically-verified hepatology knowledge graph.
• Implemented a state-driven agentic system with a self-correcting "retrieve-evaluate-refine" loop.
• Agents dynamically generated, validated, and optimized graph search strategies for context construction.

Main Results:

• The framework significantly outperformed baseline models (GPT-4, standard RAG, Graph RAG).
• Achieved superior scores in faithfulness (0.94), context recall (0.92), and answer relevancy (0.91).
• Demonstrated effective mitigation of LLM hallucinations.

Conclusions:

• The agentic approach provides accurate and interpretable answers, mitigating LLM hallucinations.
• This framework shows potential as a next-generation intelligent clinical decision support tool for hepatology.
• The system offers a robust solution for enhancing AI reliability in medical applications.