KoSEL: Knowledge subgraph enhanced large language model for medical question answering

Zefan Zeng, Qing Cheng, Xingchen Hu, Yan Zhuang, Xinwang Liu, Kunlun He, Zhong Liu

Abstract: The integration of medical knowledge graphs (KGs) and large language models (LLMs) for medical question answering (Q&A) has attracted considerable interest in recent studies. However, current approaches that combine KGs and LLMs tend to either integrate KGs directly into the fine-tuning process of LLMs or use entire KGs as a contextual prompt base for LLMs to reason, raising concerns regarding potential data leakage and reasoning confusion. In this study, we propose KoSEL (Knowledge Subgraph Enhanced Large Language Model), a novel medical Q&A framework based on KG-enhanced LLMs. KoSEL comprises two modules: Knowledge Retrieval (KR) and Reasoning and Answering (RA). The KR module is LLM-independent and employs an entity-linking algorithm and a subgraph construction and fusion strategy to retrieve question-relevant knowledge. The RA module conveys prompts to the LLM for information extraction, knowledge fusion, reasoning, and answer generation. KoSEL, which is designed as a plug-and-play framework, effectively fuses structural and textual knowledge while ensuring efficiency and privacy. The construction of a precise and refined subgraph reduces knowledge noise and the number of input graph tokens, thus mitigating hallucination issues. Extensive experiments demonstrated that KoSEL outperformed advanced methods in terms of knowledge retrieval efficiency (20.27% reduction in retrieval time), knowledge utilization (15.16% increase in utilization rate), and data protection (113.50% reduction in data leakage rate), resulting in higher-quality answers for medical Q&A tasks (1.50% improvement in answer score).

External IDs:dblp:journals/kbs/ZengCHZLHL25