脓毒症相关持续性急性肾损伤的危险因素及预测模型

Risk factors and predictive models for sepsis-associated persistent acute kidney injury

目的本文旨在利用大型国际数据库,确定脓毒症相关持续性急性肾损伤(sepsis associated AKI,SA-AKI)的危险因素,构建预测模型森林图,并验证模型准确率。方法从2019年4月15日公开发布的大型国际数据库eICU数据库中识别脓毒症相关急性肾损伤(SA-AKI)患者(886例),将SA-AKI患者分为短暂性SA-AKI组(668例)和持续性SA-AKI组(218例)。使用单因素和多因素logistic回归分析,寻找持续性SA-AKI独立危险因素,构建预测模型森林图,并通过分类模型评价指标评估模型。回顾性收集湖北某三甲医院195例数据作为外部验证组进行模型外部验证。结果单因素logistic回归分析表明:年龄、尿素氮、胆红素、SOFA评分、CRRT、KDIGO 2期、高血压病、心血管病、慢性肾病、肝病、升压剂、OASIS评分、Apache IV评分、纳入时肌酐水平、入院24 h肌酐变化、血红蛋白、血小板、ICU住院天数是持续性SA-AKI独立危险因素;多因素logistic回归向前LR分析表明:CRRT、升压剂、Apache IV评分、胆红素、纳入时肌酐、ICU内天数、24 h肌酐水平变化是持续性SA-AKI的独立预测因素,预测模型灵敏度78.9%,特异度78.7%,AUC曲线下面积0.858。以此构建SA-AKI模型整体预测准确率为83.4%,召回率83.4%,精确率82.6%,F1分数82.1%。将验证集中数据带入最终模型中,得出ROC曲线下面积0.865,灵敏度为89.1%,特异度为78.6%。结论确定了持续性SA-AKI危险因素并成功构建持续性SA-AKI预测模型。

Objective This paper aims to use a large international database to determine the risk factors of persistent sepsis associated acute kidney injury(SA-AKI),construct a prediction model forest map,and verify the accuracy of the model.Methods Patients with SA-AKI were identified from the eICU database published on April 15,2019 and divided into transient SA-AKI and persistent SA-AKI groups.Independent risk factors for persistent SA-AKI were i⁃dentified using univariate and multifactor logistic regression analyses,a forest plot of the prediction model was construc⁃ted,and the model was validated using classification model evaluation indicators.Data from 195 cases from The hospital in Hubei were retrospectively collected as an external validation group for external model validation.Results One-way logistic regression analysis showed that persistent SA-AKI was significantly different for age,urea nitrogen,bilirubin,SO⁃FA score,CRRT,KDIGO stage 2,hypertension,cardiovascular disease,chronic kidney disease,liver disease,boost⁃ers,OASIS score,Apache IV score,creatinine level at inclusion,24h creatinine change at admission,hemoglobin,platelets,length of stay ICU.The multifactorial logistic regression forward LR analysis showed that CRRT,boosters,A⁃pache IV score,bilirubin,creatinine at inclusion,days in ICU,and 24h creatinine level change were independent pre⁃dictors of persistent SA-AKI,with a model sensitivity of 78.9%,specificity of 78.7%,and area under the AUC curve of 0.858.The overall prediction accuracy of this SA-AKI model was 83.4%,recall 83.4%,precision 82.6%and F1 score 82.1%.Taking the validation set data into the final model yielded the area under the ROC curve of 0.865,with sensitivity of 89.1%and specificity of 78.6%.Conclusion Risk factors for persistent SA-AKI were identified and a predictive model for persistent SA-AKI was successfully constructed.