A Hybrid LLM–Knowledge Graph Framework for Accurate Biomedical Question Answering

Biomedical question answering requires accurate and interpretable systems; however, existing approaches often face challenges such as language model hallucinations and limited reasoning when relying solely on standalone knowledge graphs. To address these limitations, this study proposes a hybrid framework that integrates the LLaMA-3B language model with a Neo4j-based drug–disease–symptom knowledge graph. The system translates natural language questions into executable Cypher queries, operates on an iBKH-derived graph comprising over 65,000 entities and 3 million relationships, and returns answers with supporting evidence through a transparent interface. Experiments conducted on 60 biomedical questions across three levels of difficulty demonstrate the robustness of the approach: 96% exact match for simple queries, 95% for medium queries, and 86.7% for complex queries. Overall, the system achieves Precision@5 of 96.1%, Recall@5 of 89.0%, F1@5 of 91.0%, Hits@k of 96.1%, and an MRR of 94.4%, while maintaining an average response time of only 6.07 seconds. These results indicate that the system retrieves nearly all relevant answers, ranks them correctly, and delivers them with latency low enough for interactive use. Moreover, unlike cloud-based APIs such as ChatGPT, which require internet connectivity and external data transmission, the proposed framework operates fully offline, ensuring privacy, reproducibility, and compliance with biomedical data governance. Overall, this pipeline provides an accurate, efficient, and privacy-preserving solution for biomedical question answering, making it a practical alternative to cloud-dependent approaches in sensitive healthcare contexts.