【摘要】 目的 以重症监护病房（intensive care unit,ICU）内缺血性卒中患者为研究对象，分别构建随机生存森林和Cox回归并对两者预测效果进行比较。方法 以是否发生院内死亡及在ICU内接受治疗的时间为结局变量，以人口学特征、实验室指标、合并症等为预测变量，将研究对象按7:3的比例随机划分成训练集和测试集，在训练集提取重要变量并构建模型，在测试集进行模型评价。用受试者工作特征（receiver operation characteristic,ROC）曲线下面积（area under curve,AUC）评价区分度，以Brier score评价校准度，并综合灵敏度（sensitivity,TP）、特异度（specificity,TN）、阳性预测值（positive predictive value,PPV）和阴性预测值（negative predictive value,NPV）对模型比较。结果 共纳入1 505例缺血性卒中患者，其中217例发生院内死亡。500次Bootstrap自助抽样结果表明，在测试集中，随机生存森林和Cox回归AUC_7day分别为0.815（0.771～0.860）和0.756（0.707～0.805）,AUC_14day分别为0.784（0.725～0.843）和0.735（0.68～0.790）,delong检验差异有统计学意义（P<0.05）;Brier score_7day分别为0.055 （0.044～0.065）和0.073 （0.057～0.089）,Brier score_14day分别为0.079 （0.069～0.089）和0.096 （0.078～0.115）;TP_7day、TP_14day分别为0.704 （0.575～0.832）和0.667 （0.503～0.830）、0.675（0.536～0.814）和0.564（0.427～0.701）,TN_7day、TN_14day分别为0.863（0.741～0.985）和0.760（0.595～0.924）、0.873（0.763～0.982）和0.862（0.741～0.982）,PPV_7day、PPV_14day分别为0.843（0.738～0.948）和0.743（0.645～0.840）、0.846（0.746～0.946）和0.810（0.691～0.930）,NPV_7day、NPV_14day分别为0.747（0.680～0.814）和0.699（0.627～0.772）、0.731（0.651～0.812）和0.666（0.607～0.725）。结论 随机生存森林对ICU内缺血性卒中患者死亡风险预测效果优于Cox回归，可为临床决策提供参考。更多还原

【Abstract】 Objective To compare the predictive effects of random survival forest and Cox regression on patients with ischemic stroke in intensive care unit （ICU）.Methods Taking the mortality in hospital and the time of receiving treatment in ICU as outcome variables,demographic characteristics,laboratory indexes and complications as predictive variables,the research subjects were randomly divided into training set and test set according to the proportion of 7:3.Important variables were extracted from the training set and the model was constructed,and the model was evaluated in the test set.The discrimination was evaluated by area under curve （AUC） of receiver operation characteristic （ROC）,and the calibration was evaluated by Brier score.The sensitivity （TP）,specificity （TN）,positive predictive value （PPV） and negative predictive value（NPV） were compared.Results A total of 1 505 patients with ischemic stroke were included,of which 217 patients died in hospital.The results of bootstrap sampling showed that in the test set,the AUC_7dayof random survival forest and Cox regression were 0.815 （0.771-0.860） and 0.756 （0.707-0.805）,AUC_14daywere 0.784 （0.725-0.843） and 0.735 （0.68-0.790）,respectively.The difference of DeLong test was statistically significant.Brier score_7dayscores were 0.055 （0.044-0.065） and 0.073（0.057-0.089）,and Brier score_14daywere 0.079 （0.069-0.089） and 0.096 （0.078-0.115）,respectively.TP_7dayand TP_14daywere0.704 （0.575-0.832） and 0.667 （0.503-0.830）,0.675 （0.536-0.814）,and 0.564 （0.427-0.701）,respectively.TN_7dayand TN_14day were 0.863 （0.741-0.985） and 0.760 （0.595-0.924）,0.873 （0.763-0.982）,and 0.862 （0.741-0.982）,respectively.PPV_7dayand PPV14_daywere 0.843 （0.738-0.948） and 0.743 （0.645-0.840）,0.846 （0.746-0.946） and 0.810 （0.691-0.930）,respectively.NPV7_dayand NPV14_daywere 0.747 （0.680-0.814） and 0.699 （0.627-0.772）,0.731 （0.651-0.812） and 0.666 （0.607-0.725）,respectively.Conclusion Random survival forest is better than Cox regression in predicting the risk of death in patients with ischemic stroke in ICU,which can provide reference for clinical decision-making.更多还原