基于机器学习的ICU多发伤患者发生谵妄预测模型的构建与评估

Construction and evaluation of machine learning‑based delirium prediction models for ICU patients with multiple trauma

目的构建基于机器学习的ICU多发伤患者发生谵妄的预测模型,并评估其预测效能。方法采用回顾性病例对照研究分析2019年7月至2022年6月郑州大学第一附属医院收治的417例ICU多发伤患者的临床资料,其中男305例,女112例;年龄18~88岁[(47.8±15.7)岁]。急性生理学与慢性健康状况评估Ⅱ(APACHEⅡ)评分为0~50分[(9.80±0.29)分]。将患者按7∶3随机分为训练集(291例)和测试集(126例)。收集患者的人口学信息、既往史、治疗情况和实验室检查结果等资料。采用Lasso回归分析在训练集中筛选与谵妄发生有显著关联的变量,并将其纳入机器学习模型。分别采用随机森林、梯度提升树、极限梯度提升、逻辑回归、支持向量机、k最近邻等6种机器学习方法构建ICU多发伤患者谵妄预测模型。使用测试集数据计算准确率、灵敏度、精确率、F1分数、受试者工作特征(ROC)曲线下面积(AUC)以评估模型预测效能。结果随机森林、梯度提升树、极限梯度提升、逻辑回归、支持向量机、k最近邻等6个预测模型在测试集中的准确率分别为0.70、0.68、0.69、0.73、0.70、0.60;灵敏度分别为0.74、0.80、0.81、0.86、0.85、0.69;精确率分别为0.72、0.69、0.70、0.73、0.71、0.65;F1分数分别为0.73、0.74、0.75、0.79、0.78、0.67;AUC值分别为0.72、0.73、0.72、0.80、0.74、0.64。其中逻辑回归模型区分度最佳。结论成功构建出预测ICU多发伤患者发生谵妄的模型,其中逻辑回归模型具有较好的预测效能,可为多发伤患者临床护理工作提供一个早期预测和防治谵妄发生的有效工具。

Objective To construct machine learning⁃based delirium prediction models for ICU patients with multiple trauma and evaluate their prediction efficiency.Methods A retrospective case⁃control study was conducted to analyze the clinical data of 417 ICU multiple trauma patients admitted to the First Affiliated Hospital of Zhengzhou University from July 2019 to June 2022,including 305 males and 112 females,aged 18⁃88 years[(47.8±15.7)years].The score of acute physiology and chronic health status assessment II(APACHE II)was 0⁃50 points[(9.80±0.29)points].The patients were randomly divided into training set(n=291)and test set(n=126)with a ratio of 7∶3.The demographic data,past history,treatment and laboratory results of the patients were collected.Lasso regression analysis was applied to screen variables that were significantly correlated to the incidence of delirium in the training set and the variables were then included into the machine learning models.Six machine learning methods including the random forest,gradient boosting tree,extreme gradient boosting,logistic regression,support vector machine and K nearest neighbor were used to construct the delirium prediction models for ICU multiple trauma patients.The accuracy,sensitivity,precision,F1 fraction and area under the curve(AUC)of the receiver′s operating characteristics(ROC)curve were calculated by using the data in the test set to evaluate the prediction efficiency of the models.Results With regards to the six prediction models,namely random forests,gradient boosting tree,extreme gradient boosting,logistic regression,support vector machine and K nearest neighbor prediction models,the accuracy in the test set was 0.70,0.68,0.69,0.73,0.70 and 0.60 respectively;the sensitivity was 0.74,0.80,0.81,0.86,0.85 and 0.69 respectively;the precision was 0.72,0.69,0.70,0.73,0.71 and 0.65 respectively;the F1 fraction was 0.73,0.74,0.75,0.79,0.78 and 0.67 respectively;the AUC was 0.72,0.73,0.72,0.80,0.74 and 0.64 respectively.Among them,the logistic regression model had the best discriminability.Conclusion Delirium prediction models for ICU patients with multiple trauma have been successfully constructed,among which the logistic regression model has the best prediction efficiency and can serve as an effective tool for early prediction and prevention of delirium in the clinical care of patients with multiple trauma.