MedKA: A knowledge graph-augmented approach to improve factuality in medical Large Language Models.
Journal:
Journal of biomedical informatics
Published Date:
Jul 8, 2025
Abstract
Large language models (LLMs) have demonstrated remarkable potential in medical applications. However, they still face critical challenges such as hallucinations, knowledge inconsistency, and insufficient integration of domain-specific medical expertise. To address these limitations, we introduce MedKA, a novel knowledge graph-augmented approach for fine-tuning and evaluating medical LLMs. Our approach systematically transforms structured knowledge from a medical knowledge graph into a high-quality QA corpus, cMKGQA, by clustering multiple fields around clinically meaningful scenarios (e.g., diagnosis, treatment planning). This grouping strategy enables comprehensive and use-case-specific data construction and supports one-stage training of the LLM, ensuring better alignment with structured medical knowledge. This transformation process ensures the comprehensive integration of domain-specific knowledge while maintaining factual consistency. To evaluate the factuality of LLM-generated response, we further propose the Knowledge Graph-based Auxiliary Evaluation Metrics (KG-AEMs)-a novel benchmarking framework that compares LLM outputs with fine-grained, attribute-level ground truth from knowledge graph. Experimental results demonstrate that MedKA achieves state-of-the-art performance, significantly outperforming existing models, including LLaMA-3.1-8B-Chinese-Chat, HuatuoGPT2-7B, and Apollo2-7B. On the cMKGQA dataset, MedKA achieves 44.63 BLEU-1 and 17.62 BLEU-4 scores, with particularly strong performance in areas such as medication recommendations and diagnostic tests as measured by KG-AEMs. Our approach highlights the potential of integrating knowledge graphs into LLM fine-tuning to improve the accuracy and reliability of medical AI systems. It advances factual accuracy in medical dialogue systems and provides a comprehensive framework for evaluating the integration of medical knowledge into LLMs. This work is publicly available on Github: https://github.com/Yai017/MedKA.
