多模态数据融合+动态机器学习构建ICU患者MDRO感染早期预警模型的研究  

Study on early warning model of MDRO infection in ICU patients based on multimodal data fusion and dynamic machine learning

作　　者：左蝶 赵佳 龙晓艳[2] 马学先 刘冰[3] 王小燕 李萍[3,4] ZUO Die;ZHAO Jia;LONG Xiaoyan;MA Xuexian;LIU Bing;WANG Xiaoyan;LI Ping(School of Nursing,Xinjiang Medical University,Urumqi 830091,China;Xiangya Hospital of Central South University,Changsha 410028,China;The Second Affiliated Hospital of Xinjiang Medical University,Urumqi 830063,China;Health Care Research Center for Xinjiang Regional Population,Urumqi 830063,China)

机构地区：[1]新疆医科大学护理学院,新疆乌鲁木齐830091 [2]中南大学湘雅医院,湖南长沙410028 [3]新疆医科大学第二附属医院,新疆乌鲁木齐830063 [4]新疆区域人群疾病与健康照护研究中心,新疆乌鲁木齐830063

出　　处：《海南医科大学学报》2025年第6期421-432,共12页Journal of Hainan Medical University

基　　金：新疆维吾尔自治区自然科学基金(2023D01C121)。

摘　　要：目的:多模态数据融合+动态机器学习构建ICU患者多重耐药菌(multi-drug resistant organism,MDRO)感染风险预测模型,优选出最优预测模型为医院MDRO感染提供有效的评估工具。方法:选取某三甲医院ICU 2018年1月1日~2023年8月30日的1200名患者,按照8∶2的比例随机分为训练集(n=960)和测试集(n=240),基于单因素分析将P<0.05的变量作为构建模型的纳入因素,运用随机森林(RF)、极度梯度提升(XGBoost)、决策树中的分类与回归树(classification and regression trees,CART)和logistic回归分别建立ICU患者MDRO感染风险预测模型,通过准确率、灵敏度、特异度、阳性预测值、阴性预测值、Kappa值、AUC值、决策曲线和校准曲线对4种模型的预测性能进行比较。结果:RF模型在训练集和测试集中表现最佳,其准确率、灵敏度、特异度、阳性预测值、阴性预测值及Kappa值均高于其他模型。AUC值从大到小顺序排列,训练集:RF>XGBoost>CRAT>logistic回归;测试集:RF>CRAT>logistic回归>XGBoost。本研究结果显示肺部感染、脑血管疾病、低蛋白血症及侵入性操作为4种模型的高风险预测因子是MDRO感染筛查及进行临床干预的重要理论依据。结论:基于RF算法建立的风险预测模型对ICU患者MDRO感染风险的预测性能优于其他三个机器算法构建的模型。Objective:To build a risk prediction model of multi-drug resistant organism(MDRO)infection in ICU patients by multi-modal data fusion+dynamic machine learning and to select the optimal prediction model to provide an effective assessment tool for MDRO infection in hospitals.Method:A total of 1200 patients from ICU of a tertiary hospital from between January 1,2018 and August 30,2023 were randomly divided into a training set(n=960)and a test set(n=240)according to the ratio of 8∶2.Based on univariate analysis,variables with P<0.05 were used as the inclusion factors in the model construction.Random forest(RF),eXtreme Gradient Boosting(XGBoost),classification and regression trees(CART)of the decision tree model,and logistic regression were used to establish MDRO infection risk prediction models for ICU patients.Accuracy,sensitivity,specificity,positive predictive value,negative predictive value,Kappa value,AUC value,decision curve,and calibration curve were used to compare the prediction performance of the four models.Results:The RF model performed the best in the training set and test set,and its accuracy,sensitivity,specificity,positive predictive value,negative predictive value,and Kappa value were higher than the other models.AUC values were arranged in order from large to small,in the training set:RF>XGBoost>CRAT>logistic regression and in the test set:RF>CRAT>logistic regression>XGBoost.The results of this study showed that pulmonary infection,cerebrovascular disease,hypoproteinemia,and invasive operation were the risk predictors of the four models,which are important theoretical basis for MDRO infection screening and clinical intervention.Conclusion:The risk prediction model based on RF algorithm is better than the other three machine algorithms in predicting the risk of MDRO infection patients from ICU.

关 键 词：ICU患者 MDRO感染风险 机器学习 预测模型 

分 类 号：R197.323[医药卫生—卫生事业管理]