MGPRAG: Enhancing Medical Large Language Models via Precision Retrieval-Augmented Generation
Y Shen, Dahong Xu, Xi Li, Lei Fu, Hong Liu · IEEE Access · 2025
Large language models(LLMs) have demonstrated strong performance in general tasks, but remain insufficiently trusted in complex clinical question answering (CQA). This is largely due to concerns about the accuracy of the generated content. Retrieval-Augmented Generation (RAG), which supplements LLMs with external knowledge, has emerged as a promising solution. However, existing RAG approaches often suffer from imprecise knowledge retrieval, leading to redundant information and reducing response quality. In this paper, we propose MGPRAG (Medical Graph Precision RAG), a precision retrieval-augmented generation framework based on medical knowledge graphs. Specifically, it incorporates a structured task parsing module that decomposes complex CQA tasks into clinical manifestations, core medical entities, and sub-questions. Additionally, it adopts a Dynamic-Frozen Collaborative Encoder to improve semantic alignment between the extracted core medical entities and corresponding entities in the medical knowledge graph. Finally, it implements a Medical Value-Driven retrieval algorithm, which leverages a medical encoder and a disease relevance filtering strategy to achieve precise knowledge retrieval. Extensive experiments conducted on CQA benchmarks show that MGPRAG significantly outperforms strong baselines in both retrieval precision and generation quality. Moreover, the proposed Dynamic-Frozen Collaborative Encoder exhibits superior performance on both public datasets and self-constructed Medical Entity Semantic Matching (MESM) dataset.