心脏骤停后院内死亡机器学习预测模型的建立  

Establishment of the prediction model for inhospital mortality of post-cardiac arrest patients based on machine learning

作　　者：林清婷 张楠[1] 姜辉 朱华栋[1] Lin Qingting;Zhang Nan;Jiang Hui;Zhu Huadong(Emergency Department,State Key Laboratory of Complex Severe and Rare Diseases,Peking Union Medical College Hospital,the Chinese Academy of Medical Science&Peking Union Medical College,Beijing 100730,China)

机构地区：[1]中国医学科学院北京协和医学院、北京协和医院急诊科、疑难重症及罕见病全国重点实验室,北京100730

出　　处：《中国急救医学》2024年第1期63-68,共6页Chinese Journal of Critical Care Medicine

基　　金：中央高水平医院临床科研项目(2022-PUMCH-B-110)。

摘　　要：目的探索影响心脏骤停患者预后的相关因素,并通过机器学习建立一个准确、快速的预后预测模型。方法对美国重症监护医学信息数据库(MIMIC)中1772例18岁以上心脏骤停患者的数据进行回顾性分析,通过三种机器学习算法建立预测模型,包括逻辑回归(logistic regression,LR)、极致梯度提升(extreme gradient boosting,XGBoost)和支持向量机(support vector machine,SVM)算法,用于预测患者心脏骤停后院内病死率。计算受试者工作特征曲线下面积(area under the curve,AUC)、准确度、精确度、召回率和F1分数,以评估所建立模型的预测性能。结果XGBoost算法的表现优于另外两种算法。XGBoost算法建立的预测模型准确度、召回率、精确度和F1分数分别为0.762、0.812、0.765和0.788。XGBoost模型的AUC大于LR和SVM模型(0.847 vs.0.834和0.820)。XGBoost模型中最重要的前10个特征是入院24 h内乳酸、格拉斯哥昏迷评分(GCS)量表、尿素氮、血糖、血氧饱和度、白细胞和心率的最小值,入院24 h内体温和肌酸激酶同工酶(CK-MB),以及体质量的最大值。结论与LR和SVM算法相比,XGBoost算法建立的心脏骤停患者预后预测模型有更准确的预测效果。Objective To explore the relevant factors affecting the prognosis of patients with cardiac arrest and to establish an accurate and fast prognostic prediction model by machine learning.Methods Data from 1772 cardiac arrest patients over 18 years old from the medical information mart for intensive care(MIMIC)database were retrospectively analyzed and used to develop three machine learning models,including support vector machine(SVM),logistic regression(LR),and extreme gradient boosting(XGBoost)models,for predicting inhospital mortality.The areas under the receiver operating characteristic(AUC)curve,accuracy,precision value,recall value and F1 score were calculated to evaluate these models.Results In our study,the XGBoost algorithm outperformed the other algorithms.The accuracy,recall value,precision value and F1 score of the XGBoost algorithm were 0.762,0.812,0.765,and 0.788,respectively.In addition,the AUC of the XGBoost model was larger than that of the LR and SVM models(0.847 vs.0.834 and 0.820,respectively).The top 10 most important features of the XGBoost algorithm were minimum values of lactate,Glasgow coma scale(GCS),blood urea nitrogen,blood glucose,white blood cell,oxygen saturation and heart rate within 24 h after admission,and maximum values of temperature,creatine kinase-MB(CK-MB)and weight within 24 h after admission.Conclusions Compared with LR and SVM algorithms,the prediction model of cardiac arrest patients established by XGBoost algorithm in this study has more accurate prediction effect.

关 键 词：心脏骤停 院内病死率 机器学习 预后 逻辑回归 极致梯度提升 支持向量机 

分 类 号：R541.78[医药卫生—心血管疾病]