基于随机森林和XGBoost算法构建心脏骤停患者自主循环恢复后神经功能预后不良的风险预测模型

Construction of prediction model for poor prognosis of neurological function in cardiac arrest patients after the return of spontaneous circulation based on Random Forest and XGBoost algorithms

目的 利用机器学习算法构建预测心脏骤停(CA)患者自主循环恢复(ROSC)后神经功能预后不良的预测模型,探索结局相关因子。方法 回顾性收集2016年1月至2024年1月沧州市中心医院收治的CA行心肺复苏(CPR)后ROSC的患者481例为研究对象。收集患者临床资料,根据患者转出重症监护病房(ICU)时的格拉斯哥-匹兹堡脑功能表现分级(CPC)评分,将其分为预后良好组(GNO,n=158)和预后不良组(PNO,n=323)。481例患者按7∶3随机分为训练集(n=338)和测试集(n=143),训练集用于构建模型,测试集用评价模型效能。利用极端梯度提升(XGBoost)和随机森林(RF)两种机器学习算法构建患者神经功能预后不良的预测模型,分别得出影响患者神经功能预后的变量,应用SHAP进行XGBoost模型可解释性分析。将XGBoost和RF算法得出的变量取交集,再将交集变量进行多因素Logistic回归分析,得到差异有统计学意义的变量,进而构建决策树模型。在训练集和测试集上利用受试者工作特征(ROC)曲线和曲线下面积(AUC)评估决策树模型的预测性能。结果 通过XGBoost模型得到与神经功能预后不良相关的变量15个,RF模型得到与神经功能预后不良相关的变量14个,两种模型取交集得到11个与神经功能预后不良相关的交集变量[视神经鞘直径(ONSD)变化率、神经元特异性烯醇化酶(NSE)、入ICU第3天ONSD(ONSD day3)、心脏骤停至心肺复苏(CA-CPR)时间、ROSC时间、急性生理学与慢性健康状况评价Ⅱ(APACHEⅡ)评分、血肌酐、白蛋白、住ICU时间、血乳酸及年龄]。将这11个交集变量进行多因素Logistic回归分析,结果显示,PNO组与GNO组ONSD变化率、NSE、ONSD day3、ROSC时间及年龄这5个变量差异有统计学意义(P<0.05)。用这5个重要变量构建决策树模型,得出3个与患者神经功能预后不良最相关的变量(NSE、ROSC时间及ONSD变化率),在训练集上的决策树模型预测CA行CPR后ROSC患者神经功能预后不良的AUC为0.857(95%CI 0.809~0.903,P<0.001),在测试集上的AUC为0.834 (95%CI 0.761~0.906,P<0.001)。结论 基于XGBoost和RF这2种机器学习方法构建的决策树模型能够更准确地评估CA患者ROSC后神经功能的不良预后,且评价指标可能简化为NSE、ROSC时间及ONSD变化率。

Objective To construct a predictive model of poor neurological outcome in cardiac arrest(CA)patients with the return of spontaneous circulation(ROSC)after cardiopulmonary resuscitation(CPR)by using machine learning algorithms,and to explore the factors related to outcome.Methods A total of 481 CA patients with ROSC after CPR admitted to Cangzhou Central Hospital from January 2016 to January 2024 were retrospectively collected as the study objects.Clinical data were collected and the patients were divided into a good neurologic outcome group(GNO,n=158)and a poor neurologic outcome group(PNO,n=323)according to Glasgow-Pittsburgh cerebral performance category(CPC)scores at the time the patient was transferred out of the ICU.The 481 patients were randomly divided into training set(n=338)and test set(n=143)by a ratio of 7∶3.The training set was used to construct the model,and the test set was used to evaluate the model efficacy.Firstly,two machine learning algorithms,eXtreme Gradient Boosting(XGBoost)and Random Forest(RF),were used to construct a prediction model of poor neurological function prognosis of the patients,and the variables affecting the neurological function prognosis of the patients were obtained respectively.The interpretability of XGBoost model was analyzed by Shapley additive explanations(SHAP).Secondly,the intersection of variables obtained by XGBoost and RF algorithms was screened,and the intersection variables were analyzed by multivariate Logistic regression to obtain the variables with significant differences,and then the decision tree model was constructed.Finally,receiver operating characteristic(ROC)curve and the area under curve(AUC)were used to evaluate the predictive performance of the proposed decision tree model on the training set and the test set.Results XGBoost model obtained 15 variables associated with poor neurological function prognosis,Random Forest model obtained 14 variables associated with poor neurological function prognosis.The intersection of the two models obtained 11 intersection variables associated with poor neurological outcomes[change rate of optic nerve sheath diameter(ONSD),neuron-specific enolase(NSE),ONSD day 3,CA-CPR time,ROSC time,acute physiology and chronic health evaluationⅡ(APACHEⅡ),serum creatinine,albumin,length of ICU stay,blood lactic acid,age].Multivariate Logistic regression analysis was performed on these 11 intersection variables,and the results showed that there were statistical significances in 5 variables such as change rate of ONSD,NSE,ONSD day 3,ROSC time and age between PNO group and GNO group(P<0.05).Using these 5 important variables to build a decision tree model,three variables(NSE,ROSC time,change rate of ONSD)correlated with poor neurological function prognosis were obtained.The decision tree model on the training set predicted that the AUC of ROSC patients with poor neurological prognosis after CPR was 0.857(95%CI 0.809-0.903,P<0.001),and the AUC on the test set was 0.834(95%CI 0.761-0.906,P<0.001).Conclusions The decision tree model based on two machine learning algorithms,XGBoost and Random Forest,can more accurately evaluate the poor prognosis of nerve function after ROSC in CA patients,and the evaluation indicators may be simplified to NSE,ROSC time and change rate of ONSD.