基于机器学习算法的非计划重返ICU风险预测模型研究

Research on risk prediction model for unplanned return to ICU based on machine learning algorithm

目的:利用机器学习算法构建非计划重返重症监护室(ICU)风险预测模型。方法:选取山西省某三级甲等医院2019年10月12日—2023年5月21日收治的3250例ICU病人为研究对象,基于多种机器学习算法构建非计划重返ICU的风险预测模型,并对模型性能进行比较。基于性能最佳的模型分析各变量的重要性排名。结果:轻量梯度提升机综合效能最佳,其受试者工作特征曲线下面积(AUROC)=0.9968,随后依次为随机森林(AUROC=0.9964)、梯度提升决策树(AUROC=0.9924)、自适应算法(AUROC=0.9530)、Logistic回归(AUROC=0.8145)。基于轻量梯度提升机模型的变量重要性排序前15位分别为钾离子、失血量、格拉斯哥昏迷评分法评分、急性生理学和慢性健康状况评分系统Ⅱ评分、钠离子、C-反应蛋白、饮酒史、体温最小值、ICU入住时长、血肌酐、心率最小值、中性粒细胞计数、舒张压最小值、碳酸氢盐和收缩压最大值。结论:基于机器学习算法构建的非计划重返ICU风险预测模型表现良好,研究者可以借助此类算法建立风险预测模型识别高风险病人,给予其针对性的干预措施,提高医疗保健质量。

Objective:To construct a risk prediction model for unplanned return to intensive care unit(ICU)based on machine learning algorithm.Methods:A total of 3250 ICU patients from a tertiary grade A hospital in Shanxi province from October 12,2019 to May 21,2023 were selected as the research subjects.A risk prediction model for return to ICU was constructed based on multiple machine learning algorithms,and the performance of the models was compared.Analyze the importance ranking of each variable based on the best performing model.Results:The light gradient boosting machine had the best comprehensive performance,with area under the receiver operating characteristic curve(AUROC)of 0.9968,followed by random forest(AUROC=0.9964),gradient boosting decision tree(AUROC=0.9924),adaptive boosting(AUROC=0.9530),and Logistic regression(AUROC=0.8145).The top 15 variables in the importance ranking based on light gradient boosting machine were K+,blood loss,score of Glasgow Coma Scale,score of Acute Physiology and Chronic Health EvaluationⅡ,Na^(+),CRP,alcohol consumption history,minimum body temperature,ICU stay time,blood creatinine,minimum heart rate,neutrophil count,minimum diastolic blood pressure,bicarbonate,and maximum systolic blood pressure.Conclusion:The risk prediction models for unplanned return to ICU based on machine learning algorithm performs well.Researchers can use this type of algorithm to establish the risk prediction model to identify high-risk patients,provide targeted intervention measures,and improve the quality of healthcare.