五种机器学习模型预测颈动脉粥样硬化患者发生缺血性脑卒中的效能比较  

Comparison of efficacy of five machine learning models for predicting ischemic stroke in patients with carotid atherosclerosis

作　　者：张红珍 杨少玲 赫兰 林文华 顾家红 赵坤 胡静 彭媛媛 ZHANG Hongzhen;YANG Shaoling;HE Lan;LIN Wenhua;GU Jiahong;ZHAO Kun;HU Jing;PENG Yuanyuan(Department of Ultrasound,Fengxian Hospital Affiliated to Anhui University of Technology,Shanghai 201499,China;Department of Ultrasound,the Eighth People's Hospital of Shanghai,Shanghai 200235,China)

机构地区：[1]安徽理工大学附属奉贤医院超声科,上海201499 [2]上海市第八人民医院超声科,上海200235

出　　处：《右江医学》2023年第11期972-979,共8页Chinese Youjiang Medical Journal

基　　金：上海市徐汇区智慧医疗专项研究项目(XHZH202108);上海市科学技术委员会引导项目(18411970000);上海市奉贤区科技发展基金项目(奉科20211838)。

摘　　要：目的比较logistics分类(LR)、高斯朴素贝叶斯分类(GNB)、补充朴素贝叶斯分类(CNB)、支持向量机(SVM)和k近邻分类(KNN)五种机器学习(machine learning,ML)模型预测颈动脉粥样硬化患者发生缺血性脑卒中的效能。方法选取2021年3月1日—11月30日上海市第八人民医院神经内科住院的101例颈动脉粥样硬化患者为研究对象,收集患者基线资料、实验室检查及颈动脉超声检查资料,将数据集按8∶2的比例拆分为训练集和测试集,logistic回归确定模型预测变量,应用5倍重采样技术,测试不同机器学习模型的预测性能,使用准确度、灵敏度、特异度和ROC曲线下面积(AUC)等指标综合比较五种ML模型的预测效能。结果五种预测模型的准确度62.2%~83.0%、灵敏度62.5%~83.6%、特异度77.4%~100.0%、AUC 0.629~0.936,其中GNB模型的准确度83.0%、特异度100.0%、AUC 0.936较高,与其他模型比较差异有统计学意义(P<0.05)。结论五种模型均可用于预测颈动脉粥样硬化患者发生缺血性脑卒中的风险,其中GNB模型预测效能最佳。Objective To compare the efficacy of five machine learning(ML)models,namely,logistic regression classification(LR),Gaussian naive bayesian classification(GNB),complementary naive bayesian classification(CNB),support vector machine(SVM)and k-nearest neighbour classification(KNN),in predicting the occurrence of ischaemic stroke in patients with carotid atherosclerosis.Methods A total of 101 patients with carotid atherosclerosis hospitalized in the Department of Neurology of the Eighth People's Hospital of Shanghai from March 1 to November 30,2021 were selected for the study.Baseline data,laboratory tests and carotid ultrasound examinations were collected from the patients,and the data set was split into a training set and a test set in the ratio of 8∶2.Logistic regression was applied to determine the model prediction variables,and a 5-fold resampling technique was applied to test the prediction performance of different machine learning models,and the prediction efficacy of the five ML models was comprehensively compared by using accuracy,sensitivity,specificity and area under the ROC curve(AUC).Results The accuracy of the five prediction models was 62.2%-83.0%,sensitivity was 62.5%-83.6%,specificity was 77.4%-100.0%and AUC was 0.629-0.936,with the GNB model having higher accuracy(83.0%),specificity(100.0%),and AUC(0.936),and differences were statistically significant compared to the other groups(P<0.05).Conclusion All five models can be used to predict the risk of ischaemic stroke in patients with carotid atherosclerosis,with the GNB model having the best predictive performance.

关 键 词：机器学习 颈动脉粥样硬化 缺血性脑卒中 风险预测 

分 类 号：R543.4[医药卫生—心血管疾病]