【摘要】 目的 基于机器学习算法构建ICU患者中心静脉导管（CVC）相关性血流感染（CLABSI）风险预测模型，为早期识别CLABSI高危患者提供依据。方法 选取美国重症医学信息数据库MIMIC-IV中4 581例留置CVC的患者为研究对象，回顾性收集患者的临床资料。使用5种机器学习算法：逻辑回归（LR）、随机森林（RF）、极端梯度提升（XGBoost）、轻量梯度提升机（Light GBM）、自适应算法（AdaBoost）构建ICU患者CLABSI风险预测模型，通过F₁值、精确率、召回率、布莱尔分数、和受试者工作特征（ROC）曲线下面积（AUC）、校准曲线、决策曲线评估模型的预测效能。运用SHapley Additive exPlanations（SHAP）方法对表现最佳的模型进行可解释性分析，并基于该模型建立ICU患者CLABSI评分表。结果 共纳入4 581例留置CVC的ICU患者，CLABSI的发生率为2.82%。5种预测模型：LR、RF、XGBoost、Light GBM、AdaBoost在验证集的AUC值分别为0.877、0.824、0.901、0.856、0.790，其中XGBoost模型的预测性能最佳。SHAP方法分析显示抗生素种类≥3种、高SOFA评分、高APSⅢ评分、使用免疫抑制剂、低白蛋白水平及年龄≥60岁是ICU患者发生CLABSI的危险因素，颈内静脉置管和锁骨下静脉置管为保护因素。基于XGBoost模型构建的ICU患者CLABSI评分表总分为0～268分，最佳截断值为172分，此时评分表AUC为0.796。结论 基于XGBoost构建的ICU患者CLABSI风险预测模型和评分表具有良好的预测效能和临床适用性。更多还原

【Abstract】 Objective Construction of a risk prediction model for central line-associated bloodstream infections(CLABSI) in ICU patients based on machine learning algorithms to provide a basis for early identification of high-risk CLABSI patients. Methods A total of 4 581 ICU patients with central venous catheters(CVC) from the MIMIC-IV database in the United States were selected as the research subjects, and their clinical data were retrospectively collected.Five machine learning algorithms, namely logistic regression(LR), random forest(RF), extreme gradient boosting(XGBoost), light gradient boosting machine(Light GBM), and adaptive boosting(AdaBoost), were used to construct CLABSI risk prediction models for ICU patients. The predictive performance of these models was evaluated using metrics including F1 score, precision, recall, brier score, area under the receiver operating characteristic(ROC) curve(AUC),calibration curve, and decision curve analysis. The SHapley Additive Explanation(SHAP) method was used to perform interpretability analysis on the optimal model, based on which a CLABSI scoring table for ICU patients was developed.Results A total of 4 581 ICU patients with CVC were included, and the incidence of CLABSI was 2.82%. The AUC values of the five prediction models, including LR, RF, XGBoost, Light GBM, and AdaBoost, in the test set were 0.877, 0.824,0.901, 0.856, and 0.790, respectively. Among them, XGBoost had the best predictive performance. The SHAP method analysis revealed that the following factors are risk factors for the occurrence of CLABSI in ICU patients: the use of ≥3 types of antibiotics, high SOFA scores, high APS Ⅲ scores, the administration of immunosuppressants, low albumin levels, and age ≥60 years. In contrast, internal jugular vein catheterization and subclavian vein catheterization were identified as protective factors. The total score of the CLABSI scoring table for ICU patients constructed based on the XGBoost model ranged from 0 to 268 points. When the optimal cut-off value was 172 points, the AUC was 0.796.Conclusion The risk prediction model and scoring table for CLABSI in ICU patients constructed based on XGBoost demonstrate good predictive efficacy and clinical applicability.更多还原