结合SMOTE技术与优化算法的支持向量机在慢性心衰不良结局预测中的应用  

Application of SMOTE Technique and Optimization Algorithm Support Vector Machine in Predicting Adverse Outcome of Chronic Heart Failure

作　　者：李晓桐 程璠 田晶[4,5] 闫晶晶 张岩波 韩清华[2,4] Li Xiaotong;Cheng Fan;Tian Jing(Department of Physiology,School of Basic Medicine,Shanxi Medical University,Taiyuan 030001)

机构地区：[1]山西医科大学基础医学院生理学系,030001 [2]山西医科大学细胞生理学教育部重点实验室 [3]山西医科大学公共卫生学院流行病与卫生统计学教研室 [4]山西医科大学第一医院心内科 [5]重大疾病风险评估山西省重点实验室 [6]山西中医药大学

出　　处：《中国卫生统计》2024年第6期802-806,共5页Chinese Journal of Health Statistics

基　　金：国家自然科学基金项目(82103958);山西省卫生健康委员会资助项目(2021RC03)。

摘　　要：目的应用优化算法的支持向量机(support vector machine,SVM)结合合成少数类过采样技术(synthetic minority over-sampling technique,SMOTE)预测慢性心衰患者不良结局,提高分类模型预测性能。方法顺序入选2014年1月至2017年12月,山西省两所三级甲等医院心内科确诊为慢性心力衰竭的1183例住院患者,收集患者的病历资料。基于原始训练集构建logistic回归(logistic regression,LR)与支持向量机模型,同时结合SMOTE算法构建LR、SVM、遗传算法支持向量机(genetic algorithm support vector machine,GA-SVM)和粒子群支持向量机模型(particle swarm support vector machine,PSO-SVM),通过灵敏度(sensitivity,SEN)、准确度(accuracy,ACC)、特异度(specificity,SPE)、G-means、F-measure、ROC曲线下面积(area under receiver operating characteristic curve,AUC)等指标综合评价各模型的分类性能。结果相较于对原始数据进行直接分类,应用SMOTE技术均衡化数据集后,模型性能明显提高。均衡化训练集构建LR、SVM、GA-SVM和PSO-SVM模型结果表明,GA-SVM和PSO-SVM在SPE、ACC指标低于LR;SEN、G-means、F-measure和AUC均优于LR。GA-SVM和PSO-SVM的综合效果显著高于SVM(SEN、G-means、F-measure指标表现均优于SVM)。结论基于均衡化数据集构建GA-SVM或PSO-SVM模型可提高SVM对于心衰预后的预测性能。Objective Support vector machine(SVM)of optimization algorithm combined with SMOTE technique was used to predict the adverse outcome of patients with chronic heart failure and improve the prediction performance of classification model.Methods From January 2014 to December 2017,1183 inpatients diagnosed with chronic heart failure in the cardiology department of two Third-class hospitals in Shanxi Province were enrolled in this study.The medical records of the patients were collected.Construct logistic regression(LR)and support vector machine model based on the original training set,and combine SMOTE algorithm to construct LR,SVM,genetic algorithm support vector machine(GA-SVM)and particle swarm support vector machine(PSO-SVM).The classification performance of each model was comprehensively evaluated by sensitivity(SEN),accuracy(ACC),specificity(SPE),G-means,F-measure,area under receiver operating characteristic curve(AUC)and other indicators.Results Compared with classifying the original data directly,SMOTE technique was applied to equalize the data set,and the model performance was significantly improved.The results of equalization training set to construct LR,SVM,GA-SVM and PSO-SVM models show that GA-SVM and PSO-SVM are lower than LR in SPE and ACC indicators,and SEN,G-means,F-measure and AUC are better than LR.The comprehensive effect of GA-SVM and PSO-SVM is significantly higher than that of SVM(SEN,G-means and F-measure are better than SVM).Conclusion The GA-SVM or PSO-SVM model based on the equalization dataset can improve the prediction performance of SVM for the prognosis of heart failure.

关 键 词：SMOTE 支持向量机 遗传算法优化 粒子群算法优化 慢性心力衰竭 

分 类 号：R541.4[医药卫生—心血管疾病]