Construction and Application Research of a Large Model for Knowledge Fusion and Reasoning in Traditional Chinese Medicine Classics
TONG Lin
CHAI Zhongyan
LI Bing
ZENG Ziling
ZHU Wenhui
ZHANG Lei
ZHENG Danping
LIU Sihong
HUANG Yu
ZHANG Huamin
Abstract：Objective To construct a large model for knowledge reasoning and decision-making of traditional Chinese medicine(TCM)classics,and realize intelligent mining of classic knowledge and clinical auxiliary decision-making.Methods Based on the Qwen2.5-32B pre-trained version of Tongyi Qianwen as the model base,the"Linglan Midian·Zhongyan"large language model was established by applying the meta-terminology engine technology for TCM classics,injecting ancient-modern multi-source parallel knowledge graphs,combining multi-level incremental and self-reflective reasoning chain-of-thought technology,and through four-stage training optimization including continuous pre-training,supervised fine-tuning,reinforcement learning,and knowledge graph enhancement.Results Through zero-shot testing on the self-built TCM knowledge evaluation benchmark,the accuracy of the model increased from 71.32％ of the base model to 79.16％ after three-stage optimization of continuous pre-training,supervised fine-tuning,and reinforcement learning.15 TCM experts with associate senior titles or above conducted blind reviews on 200 complex medical cases,with an average consistency κ=0.79,indicating that"Linglan Midian·Zhongyan"has reached a usable level in understanding TCM ancient books and supporting clinical decision-making.Conclusion The large model effectively improves the efficiency of knowledge understanding of TCM classics and the application capability of clinical auxiliary decision-making,providing new technologies and tools for intelligent interpretation of classics and their clinical transformation and application.
Keywords：TCM classicsLarge language modelKnowledge graphClinical decision support
Publication Date：2025-12-28
Online Publishing Date：2026-01-07(First online date of this platform, not the publication date of the document)
Pages：6( 2130-2135 )