基于机器学习的急性胰腺炎中医辨证模型构建  

Based on machine learning construction of traditional Chinese medicine syndrome identification model of acute pancreatitis

作　　者：谈贝 郑飞波[2] 张坤 崔云峰[3] Tan Bei;Zheng Feibo;Zhang Kun;Cui Yunfeng(Tianjin University of Traditional Chinese Medicine,Tianjin 301617,China;Tianjin Medical University,Tianjin 300070,China;Department of Hepatobiliary and Pancreatic Surgery,Tianjin Nankai Hospital,Tianjin 300100,China)

机构地区：[1]天津中医药大学,天津301617 [2]天津医科大学,天津300070 [3]天津市南开医院肝胆胰外科,天津300100

出　　处：《中国中西医结合急救杂志》2023年第3期321-324,共4页Chinese Journal of Integrated Traditional and Western Medicine in Intensive and Critical Care

基　　金：天津市医药卫生中医中西医结合科研项目(2021006);天津市中医药重点领域科研项目(2022005)。

摘　　要：目的基于机器学习构建急性胰腺炎(AP)的中医智能辨证模型,并比较不同机器学习算法模型的效能。方法检索中国知网数据库,收集2004年12月至2022年3月公开发表的应用中医药治疗AP的文献资料,建立AP中医辨证信息数据库。运用决策树(DT)、随机森林(RF)、支持向量机(SVM)、人工神经网络(ANN)、K-近邻(KNN)5种机器学习算法构建AP中医辨证模型。采用五折交叉验证法对不同算法模型的效能进行评估;采用RF模型分析各个症状体征对于AP辨证分型的重要性。结果最终纳入符合要求的中医药治疗AP的相关文献260篇。将所有特征中出现频次低于10次的症状体征或证型剔除,最终留取53个症状体征作为特征变量,获得4个AP常见证型,分别为腑实热结证、肝郁气滞证、瘀毒互结证、湿热蕴结证。分别构建不同机器学习算法的AP中医辨证模型,经五折交叉验证显示,基于RF算法的模型效果最佳,其准确率、查准率、查全率和F1分数均在95%以上(分别为96.2%、97.1%、95.6%、96.1%);而DT和KNN模型的各项效能评估结果较差。基于RF模型的特征重要性分析显示,特征重要性数值排名前10位的症状体征依次为身目发黄(0.0768)、脉洪大(0.0597)、苔白(0.0567)、腹满硬痛拒按(0.0535)、舌淡红(0.0531)、脉弦(0.0493)、脉涩(0.0477)、舌质红(0.0459)、舌有瘀斑(0.0430)、苔薄(0.0403)。结论基于RF构建的AP中医辨证模型具有较高的准确率。Objective To establish intelligent traditional Chinese medicine(TCM)syndrome identification models for acute pancreatitis(AP)based on machine learning,and compare the performance of different machine learning algorithm models.Methods The database of China National Knowledge Infrastructure(CNKI)was researched to collect published literatures on the application of TCM for the treatment of AP from December 2004 to March 2022,and a database of TCM identification information of AP was established.Five machine learning methods such as decision tree(DT),random forest(RF),support vector machine(SVM),artificial neural network(ANN),and K-nearest neighbor(KNN)were applied to construct TCM syndrome identification models for AP.Five-fold cross-validation was used to evaluate the effectiveness of different algorithmic models.RF was applied to analyze the importance of each symptom and sign for the TCM syndrome identification of AP.Results A total of 260 papers related to the treatment of AP with TCM that fulfilled the requirements were finally enrolled.The symptoms and signs among all features or TCM syndrome types that occurred less than 10 times were excluded,and finally 53 symptoms and signs were retained as characteristic variables and 4 common TCM syndrome types of AP were obtained,namely,Fu-organ excess and heat retention syndrome,liver Qi stagnation syndrome,intermingled toxin and blood stasis syndrome,and dampness-heat amassment syndrome.TCM syndrome identification models for AP with different machine learning algorithms were constructed.Five-fold cross-validation showed that the model based on the RF algorithm worked best,with accuracy,precision,recall and F1 score all above 95%(96.2%,97.1%,95.6%,96.1%,respectively).However,the DT and KNN models had poorer results for each effectiveness assessment.The feature importance analysis based on the RF model showed that the top 10 signs and symptoms in the ranking of feature importance were yellowing of the skin and eyes(0.0768),flooded pulse(0.0597),white tongue coating(0.05

关 键 词：急性胰腺炎 机器学习 中医辨证模型 随机森林 

分 类 号：R2-031[医药卫生—中西医结合]