中国临床药学杂志

目的 针对中药饮片处方人工点评效率低、辨证逻辑复杂、基层药师专业能力不足等问题，建立基于DeepSeek大语言模型的中药饮片处方智能点评系统，并评估其应用效果。方法 构建融合《中华人民共和国药典》（2020版）、上海市中药饮片炮制规范（2018版）、中医经典文献及临床指南等文献的多维知识库，结合DeepSeek-V3/R1模型搭建检索增强生成架构。设计结构化辨证框架（证型-舌脉-病机-治则）与3级风险规则（高危/中危/低危），实现“药证矛盾”的智能识别。采用德尔菲法建立规则标记的金标准，智能系统点评小组（DeepSeek智能点评系统预审+1名药师终审）与传统人工点评小组（双药师背靠背审核）分别评价50份中药饮片处方，比较2个小组点评中药饮片处方的准确性、效率、安全性和稳健性。结果 智能系统点评小组高危错误检出率和辨证矛盾识别准确率均高于传统人工点评小组（P<0.05）；与传统人工点评小组比较，智能系统点评小组审核效率提升了77.5%；智能系统点评小组的稳健性F1分数达到0.94，高于传统点评小组，差异均有统计学意义（P<0.05）。典型案例表明DeepSeek可精准拦截误用的肝毒性药物（如朱茯苓）和寒热错投药物（如熊胆粉）。结论 基于DeepSeek大语言模型的中药饮片处方智能点评系统通过知识库驱动辨证逻辑与人机协同终审机制，可显著提升中药饮片处方审核的精准性与效率。

AIM To address the inefficiency of manual prescription review, complexity of syndrome differentiation logic, and insufficient expertise of primary care pharmacists in traditional Chinese medicine(TCM), an intelligent prescription review system for Chinese herbal pieces was developed based on the DeepSeek large language model(LLM), and its application efficacy was evaluated. METHODS A multidimensional knowledge base was constructed by integrating the Pharmacopoeia of the People's Republic of China(2020 edition), Shanghai Processing Standards for Chinese Herbal Pieces(2018 edition), classical TCM literature, and clinical guidelines. A retrieval-augmented generation framework was established using the DeepSeek-V3/R1 model. A structured syndrome differentiation framework(syndrome type–tongue/pulse–pathogenesis–treatment principle) and a three-tier risk rule system(high/medium/low risk) were designed to intelligently identify "herb-syndrome incompatibility". Using the Delphi method, a gold standard for rule-based annotation was established. The AI-assisted review group(DeepSeek intelligent system pre-review + final review by one pharmacist) and the traditional manual review group(dual pharmacists conducting back-to-back reviews) independently evaluated 50 Chinese herbal prescriptions. Accuracy, efficiency, safety, and robustness were compared between the2 groups. RESULTS The AI-assisted group demonstrated superior performance than the manual group: a higher rate of high-risk error detection, improved accuracy in syndrome incompatibility recognition(P < 0.05), a 77.5% increase in review efficiency, and a robustness F1-Score of 0.94 with statistically significant differences. Typical cases confirmed the system's precision in intercepting hepatotoxic drug misuse(e.g., Cinnabaris-processed Poria) and mismatched cold-heat syndrome applications(e.g., bear bile powder). CONCLUSION The DeepSeek LLM-based intelligent prescription review system significantly enhances the precision and efficiency of Chinese herbal prescription review through knowledge basedriven syndrome differentiation logic and human-AI collaborative verification.