大语言模型在中医古籍智能解读与临床辅助 诊疗中的应用

The Application of LLM in the Intelligent Interpretation of Classical Chinese Medical Texts and Clinical Diagnostic Support

摘要: 中医药经典古籍承载数千年理论精髓与临床经验，是传承创新核心载体，但文辞深奥、知识零散、非标准化术语，导致古籍传承低效、经典与临床脱节；通用大语言模型(LLM)直接应用于中医领域，易出现知识幻觉、辨证偏差、结论无经典依据等问题，难以满足临床科研需求。针对上述痛点，本文构建融合检索增强生成(RAG)与中医知识图谱的LLM优化框架，以四大经典古籍、名老中医医案、国标术语及方剂药材数据为数据集，经文本预处理、知识融合、小样本微调、RAG约束等流程，实现古籍深度解读、临床辨证、方剂推荐、配伍校验全流程功能。实验选取200条古籍片段、150例临床医案构建测试集，从五大维度开展对比实验，结果显示，优化框架古籍术语识别准确率92.3%，临床辨证一致率88.7%，方剂合规率94.2%，幻觉率降至4.7%，远优于通用模型与单一知识图谱模型。该框架实现中医经典与临床诊疗闭环衔接，解决LLM中医应用可信性难题，为古籍活化、基层诊疗规范化、名医经验传承提供可行方案，助力中医药数字化智能化转型。

Abstract: Classical texts of Traditional Chinese Medicine (TCM) encapsulate thousands of years of theoretical essence and clinical experience, serving as the core vehicle for the transmission and innovation of this discipline. However, their archaic language, fragmented knowledge and non-standardised terminology have led to inefficient transmission of these texts and a disconnect between classical texts and clinical practice. The direct application of general-purpose large language models (LLMs) to the field of TCM is prone to issues such as knowledge illusions, diagnostic biases and conclusions lacking classical support, making it difficult to meet the demands of clinical research. To address these challenges, this paper constructs an optimised LLM framework that integrates Retrieval-Augmented Generation (RAG) with a TCM knowledge graph. Utilising a dataset comprising four major classical texts, case records from renowned senior TCM practitioners, national standard terminology, and data on formulae and medicinal materials, the framework undergoes text pre-processing, knowledge fusion, few-shot fine-tuning, and RAG constraints to achieve end-to-end capabilities in classical text interpretation, clinical pattern differentiation, formula recommendation, and ingredient compatibility verification. For the experiments, a test set comprising 200 classical text excerpts and 150 clinical case records was constructed. Comparative experiments were conducted across five dimensions. The results demonstrated that the optimised framework achieved a 92.3% accuracy rate in classical text terminology recognition, an 88.7% consistency rate in clinical pattern differentiation, and a 94.2% compliance rate for formulae, whilst reducing the hallucination rate to 4.7%. These results significantly outperform those of general-purpose models and models relying on a single knowledge graph. This framework achieves a closed-loop integration between classical TCM texts and clinical diagnosis and treatment, addressing the credibility challenges associated with the application of Large Language Models (LLMs) in TCM. It provides a viable solution for the revitalisation of classical texts, the standardisation of primary-level diagnosis and treatment, and the transmission of renowned physicians’ expertise, thereby supporting the digital and intelligent transformation of Traditional Chinese Medicine.