Quality-Controllable automatic construction method of Chinese knowledge graph for medical decision-making applications

Medical Knowledge Graphs (KGs) store complex medical knowledge in a structured manner, increasingly becoming the foundation of medical artificial intelligence. They provide interpretable evidence for disease diagnosis and treatment, and enhance the accuracy and interpretability of medical information in large language models (LLMs), thus mitigating the hallucination issues. However, existing medical KGs lack diverse knowledge types, sufficient coverage, fine granularity, and high quality, resulting in low utilization rates. To address these issues, this paper, under the guidance of medical professionals, proposes guidelines and automated methods for constructing a Chinese medical KG, drawing from existing experience in building KGs and the requirements of medical decision systems. The construction principles include (1) universality and personalization, (2) comprehensiveness and granularity, (3) knowledge quality control. Furthermore, the automated construction method integrates a chain of thought-based knowledge mining approach and an axiom logic-based quality control module, which improves the scalability of mining and the quality of the knowledge. Based on these, a Chinese medical KG named WiMedKG has been developed. It meets the established construction guidelines by: (1) including both commonsense and experiential medical knowledge, (2) comprehensively covering 111 departments with content ranging from clinical practice to preventive medicine and rehabilitation treatments. The granularity of the knowledge is detailed, featuring 29 entity types, 128 refined relationship types, and 40 attribute types, comprising a total of 367,108 entities, 3,176,389 relational triples, and 1,021,966 attribute triples. (3) The knowledge has been validated and completed, receiving an evaluation score of 90.66% from medical professionals, which demonstrates the reliability of the quality-controlled automatic KG construction method. Finally, we constructed medical LLM WiMedLLM enhanced by WiMedKG. Experimental results on the medical test dataset show an average performance improvement of 1.51% after KG enhancement, demonstrating the necessity of KG construction and the effectiveness of the automatic construction method. The data and system resources can be found on our page: https://github.com/lx-hit/WiMedKG.