MedSumGraph: enhancing GraphRAG for medical QA with summarization and optimized prompts.

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing
• Knowledge Representation

Background:

• Large language models (LLMs) show promise in medical AI but struggle with specialized terminology, hallucinations, and interpretability.
• Existing methods like retrieval-augmented generation (RAG) and GraphRAG improve knowledge integration but increase complexity and reliance on external resources.

Purpose of the Study:

• To introduce MedSumGraph, a novel system enhancing GraphRAG for medical question answering.
• To improve LLM understanding of domain-specific knowledge and enhance response reliability and interpretability.

Main Methods:

• Developed MedSumGraph by integrating structured medical knowledge summaries and optimized prompt designs into a GraphRAG framework.
• Enabled LLMs to interpret domain knowledge and embed factual evidence and graph-based reasoning directly into the generation process.

Main Results:

• MedSumGraph achieved competitive performance on medical QA benchmarks, including MedQA (USMLE), outperforming closed-source LLMs and domain-specific models.
• Demonstrated effective generalization to open-domain QA tasks, improving reasoning and truthfulness evaluation.

Conclusions:

• Structured summarization and graph-based reasoning significantly enhance the trustworthiness and versatility of LLM-driven medical AI.
• MedSumGraph offers a promising approach to overcome LLM limitations in complex medical applications.