决策树模型对糖尿病合并急性心肌梗死行急诊PCI患者术后不良事件的预测价值  

Predictive value of decision tree model to post-operative adverse endpoint events in patients with acute myocardial infarction complicated by diabetes after emergency PCI

作　　者：刘志煜 孔亚伟 张扬辉 张亚豪 王喆[4] 路新源 郑颖颖 沈德良[1,2] 张金盈 Liu Zhiyu;Kong Yawei;Zhang Yanghui;Zhang Yahao;Wang Zhe;Lu Xinyuan;Zheng Yingying;Shen Deliang;Zhang Jinying(Department of Cardiovascular Medicine,First Affiliated Hospital of Zhengzhou University,Zhengzhou 450052,China;不详)

机构地区：[1]郑州大学第一附属医院心血管内科,郑州450052 [2]河南省心脏损伤修复重点实验室,郑州450052 [3]首都医科大学附属北京儿童医院慢病管理中心,国家儿童医学中心,北京100045 [4]中国医学科学院,北京协和医学院,中日友好医院心脏科,北京100730 [5]复旦大学公共卫生学院卫生部卫生技术评估重点实验室,上海200032

出　　处：《中国循证心血管医学杂志》2022年第11期1334-1340,共7页Chinese Journal of Evidence-Based Cardiovascular Medicine

基　　金：国家自然科学基金(81570274;81870328)。

摘　　要：目的应用决策树卡方自动交互检测(CHAID)算法和二分类Logistic回归分析法分别构建罹患糖尿病(DM)的急性心肌梗死(AMI)患者行急诊冠状动脉介入治疗(PCI)术后院外2年内的不良终点事件的风险预测模型,并对模型的预测结果进行对比分析。方法回顾性纳入2016年1月至2017年1月于郑州大学第一附属医院心脏重症科(CCU)的DM-AMI行急诊PCI术后患者(信息采集来自CORFCHD-ZZ研究),并对其院外2年内的不良终点事件进行随访,应用CHAID法和二分类Logistic回归分析分别建立风险预测模型,通过受试者工作特征曲线(ROC)的曲线下面积(AUC)对两种模型的预测效果进行对比评价。结果纳入分析患者525例,其中2年发生不良终点事件203例(38.7%);CHAID法和Logistic回归分析法均显示“年龄≥60岁”、“BNP≥350ng/L”、“CRP>8.3mmol/L”、“不使用降糖药物”和“糖尿病病程大于5年”是不良终点的重要危险因素,并且在决策树中“年龄”是首要影响因素;决策树模型风险预测的正确率为69.1%,模型拟合效果较好;Logistic回归模型Hosmer-Lemeshow拟合优度检验显示模型拟合较好(χ^(2)=11.976,P>0.05)。决策树模型AUC为0.765(95%CI:0.727~0.801),Logistic回归模型AUC为0.784(95%CI:0.746~0.818),两模型预测价值均为中等,其存在的差异无统计学意义;决策树模型灵敏度高于Logistic回归预测模型,二者分别为77.6%和73.9%。结论决策树分析结果能更为直观、形象地反映DM-AMI行PCI手术人群的术后生存风险特征,“年龄≥60岁”、“BNP≥350ng/L”、“CRP>8.3mmol/L”、“不使用降糖药物”和“糖尿病病程大于5年”参与不良终点事件的发生,并且年龄是首要危险因素,决策树有利于临床医师对高危人群进行风险预测和制定随访方案。Objective To establish the risk predictive models of adverse endpoint events respectively by using decision tree Chi-square automatic interactive detection(CHAID)algorithm and binary Logistic regression analysis in patients with acute myocardial infarction(AMI)complicated by diabetes(DM-AMI)within 2y after emergency PCI,and to compare and analyze the predictive results of the models.Methods DM-AMI patients undergone emergency PCI(data collected from CORFCHD-ZZ study)were retrospectively chosen from Cardiac Care Unit in the First Affiliated Hospital of Zhengzhou University from Jan.2016 to Jan.2017,and followed up on adverse endpoint events within 2y after discharged.The risk predictive models were established respectively by using CHAID and binary Logistic regression analysis.The predictive efficacy of 2 models was compared and reviewed by using the area under curve(AUC)of receiver operating characteristic(ROC)curve.Results There were totally 525 patients included in the analysis,and adverse endpoint events occurred in 203 cases(38.7%)within 2 y after PCI.The results of CHAID and binary Logistic regression analysis all showed that age≥60,BNP≥350ng/L,CRP>8.3 mmol/L,without the usage of antidiabetics,and diabetes course>5 y were important risk factors of adverse endpoint events,and age was a primary influence factor.The correct rate of risk prediction was 69.1% in the CHAID model,and the model fitting effect was good.Hosmer-lemeshow test in the Logistic regression model showed that the model fitting was good(χ^(2)=11.976,P>0.05).AUC was 0.765(95%CI:0.727~0.801)in the CHAID model,and AUC was 0.784(95%CI:0.746~0.818)in the Logistic regression model.The predictive value of both models was moderate and the difference had no statistical significance.The sensitivity of the CHAID model was higher than that of the Logistic regression model(77.6%vs.73.9%).Conclusion The results of CHAID can reflect more vividly postoperative survival risk characteristics in DM-AMI patients undergone PCI.The factors of age≥60,BNP≥35

关 键 词：急性心肌梗死 糖尿病 不良终点事件 决策树模型 LOGISTIC回归 

分 类 号：R542.22[医药卫生—心血管疾病]