Diabetes question answering method based on knowledge graph and large language model

Abstract

Conventional medical Question Answering (QA) systems often suffer from insufficient domain-specific knowledge grounding, limited semantic understanding, and weak answer traceability. To address these issues, this study proposes a decision-support QA framework that integrates structured retrieval over a diabetes knowledge graph with retrieval-augmented generation (RAG) and large language model (LLM)-based answer synthesis. First, a diabetes knowledge graph is constructed from the DiaKG corpus with explicit entity and relation types. A Qwen2.5–1.5B model fine-tuned with LoRA is used for named entity recognition, and a BERT-CNN hybrid model is designed for intent classification. In the answer generation stage, multi-hop Cypher queries over the knowledge graph are combined with guideline-oriented semantic retrieval. The two evidence sources are then separately presented in the prompt to support traceable answer generation by GLM-4-Flash through LangChain. Experimental results on component-level tasks show that the proposed method achieves disease entity recognition F1 of 81.02%, drug entity recognition F1 of 85.27%, and intent classification accuracy of 84.31%. The current system is intended as a medical information-service and decision-support tool rather than a substitute for professional clinical consultation. Further end-to-end clinical evaluation, ablation analysis, robustness testing, and expert-rated answer assessment are still required before deployment in patient-facing scenarios.