基于机器学习的ICU老年患者呼吸机相关肺炎风险预测模型的构建及评价  

Establishment of risk prediction model for ventilator-associated pneumonia in elderly ICU patients based on machine learning and its performance

作　　者：甘文思 黄一睿 王笑青 陶真 杨西帆 李文渊 GAN Wen-si;HUANG Yi-rui;WANG Xiao-qing;TAO Zhen;YANG Xi-fan;LI Wen-yuan(Wenzhou Hospital of Integrated Traditional Chinese and Western Medicine,Wenzhou,Zhejiang 325000,China;不详)

机构地区：[1]温州市中西医结合医院感染管理科,浙江温州325000 [2]浙江大学公共卫生学院大数据健康科学系,浙江大学医学院附属第二医院临床大数据与统计中心,浙江杭州310058

出　　处：《中华医院感染学杂志》2025年第2期290-296,共7页Chinese Journal of Nosocomiology

基　　金：国家青年自然科学基金资助项目(82203984);健康浙江百万人群队列基金资助项目(K-20230085);浙江省预防智能医学重点实验室基金资助项目(2020E10004);浙江省领军型创新团队引进基金资助项目(2019R01007);浙江省重点研发计划基金资助项目(2020C03002);温州市科技局基金资助项目(Y2020253)。

摘　　要：目的构建重症监护室(ICU)老年患者呼吸机相关肺炎(VAP)的机器学习模型,为临床决策提供依据。方法本研究回顾性收集温州市某三甲医院2016年1月—2022年2月收治的377名ICU老年机械通气患者临床资料数据,并将数据集按3:1随机划分为训练集和测试集。使用递归特征消除法对训练集的自变量进行筛选并构建了多种机器学习模型,包括logistic回归、支持向量机(SVM)、决策树(DT)、极端梯度提升(XGBoost)和随机森林(RF);采用敏感度、阳性预测值、特异度、阴性预测值、F1分数、准确度和曲线下面积(AUC)等指标评估模型性能,确定较优机器学习模型,通过校准曲线评价最优模型的校准度,采用沙普利加性解释(SHAP)方法对较优模型进行解释,开发较优机器学习模型的在线计算工具。结果ICU老年患者VAP发生率16.98%。logistic回归、DT、RF、SVM和XGBoost五个模型在训练集中的AUC值为0.97、0.96、0.99、0.97、0.99;测试集中AUC值为0.84、0.78、0.90、0.88、0.90。XGBoost、RF模型被选为较优的机器学习模型。通过SHAP方法确定了机械通气时间、白蛋白水平、ICU住院时间、长期联合使用抗菌药物为重要的预测因子。基于RF、XGBoost模型开发了在线计算工具。结论本研究建立了基于机器学习算法的ICU老年患者VAP风险预测模型,并发现XGBoost和RF模型在总体性能上表现较优。为便于应用,开发了在线计算工具。较优模型及在线工具有助于医务人员及时准确地识别高危患者,为临床决策提供重要依据。OBJECTIVE To establish machine learning models for ventilator-associated pneumonia(VAP)in elderly patients of intensive care unit(ICU)so as to provide bases for clinical decisions.METHODS The clinical data were retrospectively collected from 377 elderly ICU patients who were treated with mechanical ventilation in a three-A hospital of Wenzhou from Jan.2016 to Feb.2022 The data set were randomly divided into the training set and the testing set in a 3:1 ratio.The independent variables were screened out from the training set by recursive feature elimination,and various machine learning models were established,including Logistic regression,support vector machine(SVM),decision tree(DT),extreme gradient boosting(XGBoost)and random forest(RF).The performance of the models was evaluated to choose the best-performing machine learning model based on such indexes as sensitivity,positive predictive value,specificity,negative predictive value,F1 score,accuracy and area under the curve(AUC).The reliability of the optimal model was assessed through calibration curves,and the optimal model was interpreted by Shapley Additive Explanations(SHAP).The online computational tools for the optimal machine learning model were developed.RESULTS The incidence of VAP was 16.98%among the elderly ICU patients.The AUC of Logistic regression was 0.97 for the training set,DT 0.96,RF 0.99,SVM 0.97,XGBoost 0.99;the AUC of the Logistic regression was 0.84 for the testing set,DT 0.78,RF 0.90,SVM 0.88,XGBoost 0.90.XGBoost and RF were chosen as the better-performing machine learning models.SHAP analysis revealed that mechanical ventilation duration,albumin level,length of ICU stay and long-term combined use of antibiotics were the major predictive factors.The online computational tools were developed based on the RF and XGBoost models.CONCLUSIONS The risk prediction models for VAP of the elderly ICU patients are established based on the machine learning algorithms.the XGBoost and RF models show better overall performance.The online computational tool

关 键 词：机器学习 预测模型 重症监护室 老年患者 呼吸机相关肺炎 

分 类 号：R197.323[医药卫生—卫生事业管理]