基于迁移学习构建李素卿治疗反复呼吸道感染智能处方模型  

Study on Constructing an Intelligent Prescription Model for LI Suqing's Treatment of Recurrent Respiratory Tract Infections Based on Pre-training Transfer Learning

作　　者：骆长永 刘智 于河[3] 闫雨柔 王俊宏[4] 何冰[4] 谷晓红[3] LUO Changyong;LIU Zhi;YU He;YAN Yurou;WANG Junhong;HE Bing;GU Xiaohong(Dongfang Hospital,Beijing University of Chinese Medicine,Beijing 100078,China;Nanjing University of Chinese Medicine,Nanjing 210023,Jiangsu,China;Beijing University of Chinese Medicine,Beijing 100029,China;Dongzhimen Hospital,Beijing University of Chinese Medicine,Beijing 100700,China)

机构地区：[1]北京中医药大学东方医院,北京100078 [2]南京中医药大学,江苏南京210023 [3]北京中医药大学,北京100029 [4]北京中医药大学东直门医院,北京100700

出　　处：《中华中医药学刊》2025年第4期14-17,共4页Chinese Archives of Traditional Chinese Medicine

基　　金：国家自然科学基金项目(82205317);中央高水平中医医院临床科研基金项目(K2023C14);国家重点研发计划项目(2018YFC1704100,2018YFC1704101)。

摘　　要：目的通过迁移学习方法构建李素卿教授治疗反复呼吸道感染智能处方模型。方法首先使用李素卿教授的大量医案数据作为源域知识数据,基于长短期记忆网络(Long Short-Term Memory,LSTM)的端到端(Sequence to Sequence,seq2seq)算法,融合了症状和药物关注的两种注意机制,对模型进行预训练,再以李素卿教授典型医案作为目标域知识数据进行迁移学习训练。结果图灵测试结果显示各类参与评估的中医医生的误差率均超过30%,其中长期跟诊李素卿教授的医生误差率最低。进一步的定性与定量评价显示,采用迁移学习方法的模型平均分为6.14,与李素卿教授原始处方模型得分6.50非常接近,而单一的LSTM模型评分较低。结论预训练迁移学习方法在模拟名老中医的处方用药思路上表现出色,为中医人工智能辅助决策系统提供了有效的解决方案,展示了深度学习在传统医学领域的巨大应用潜力。Objective This study aimed to construct a deep learning model for the treatment of recurrent respiratory tract infections by using pre-training transfer learning methods.Method Firstly,a large number of medical case data from Professor LI Suqing's tutor was used as the source domain knowledge data.Based on the Sequence to Sequence(seq2seq)Algorithm of Long Short-Term Memory Network(LSTM),It combined the two attention mechanisms of symptom and drug attention to pre-train the model.Then,Professor LI Suqing's typical medical cases were used as target domain knowledge data for transfer learning training.Result The Turing Test results showed that the error rate of participating traditional Chinese medicine doctors exceeded 30%,with the lowest error rate among those who had long-term follow-ups with Professor LI Suqing.Further qualitative and quantitative evaluation showed that the average score of the model using transfer learning method was 6.14,which was very close to the score of Professor LI Suqing's original prescription model of 6.50,while the score of the simple LSTM model was lower.Conclusion Pre-training transfer learning method performed well in simulating the prescription patterns of famous traditional Chinese medicine doctors,providing an effective solution for the artificial intelligence-assisted decision-making system of traditional Chinese medicine,and demonstrating the great potential of deep learning in the field of traditional Chinese medicine.

关 键 词：名医传承 迁移学习 李素卿 

分 类 号：R256.1[医药卫生—中医内科学]