基于决策树的重症儿童急性肾损伤预测模型的建立及其预测价值分析  

Development of a prediction model based on decision tree for acute kidney injury in critically ill children and its predictive value

作　　者：李会文 陈娇[2] 胡俊龙 许婧 柏振江[2] 李晓忠[1] 李艳红[1,3] Li Huiwen;Chen Jiao;Hu Junlong;Xu Jing;Bai Zhenjiang;Li Xiaozhong;Li Yanhong(Department of Nephrology and Immunology,Children′s Hospital of Soochow University,Suzhou 215000,China;Pediatric Intensive Care Unit,Children′s Hospital of Soochow University,Suzhou 215000,China;Institute of Pediatric Research,Soochow University,Suzhou 215000,China)

机构地区：[1]苏州大学附属儿童医院肾脏免疫科,215000 [2]苏州大学附属儿童医院重症医学科,215000 [3]苏州大学儿科临床研究院,215000

出　　处：《中国小儿急救医学》2025年第2期128-134,共7页Chinese Pediatric Emergency Medicine

基　　金：国家自然科学基金项目(81971432);江苏省科技项目重点研发计划(BE2020660)。

摘　　要：目的基于最小绝对收缩和选择算子(LASSO)回归、分类和回归树(CART)算法筛选重症儿童急性肾损伤(AKI)相关的临床因素并构建预测AKI的决策树模型,探索决策树模型的临床有效性。方法采用前瞻性队列研究,预测模型推导队列纳入350例2020年9月至2021年1月苏州大学附属儿童医院PICU收治的重症患儿;外部验证队列纳入866例2021年2月至2022年2月苏州大学附属儿童医院PICU收治的重症患儿。于电子病历系统收集患儿的临床资料,包括人口统计学指标、实验室指标及疾病严重程度评分。通过LASSO回归筛选与AKI显著相关的变量,并进一步基于CART算法构建决策树模型,应用受试者工作特征(ROC)曲线、校准曲线及临床决策曲线评价预测模型的预测效能。结果推导队列的350例患儿中,AKI患儿107例(30.6%);外部验证队列的866例患儿中,AKI患儿165例(19.1%)。LASSO回归筛选出16个与AKI显著相关的变量。决策树模型最终筛选出4个与AKI更为相关的变量,包括血肌酐较基线血肌酐的变化倍数、尿量、第三代儿童死亡风险评分及C-反应蛋白。训练队列、内部验证队列和外部验证队列中决策树模型的ROC曲线下面积分别为0.92、0.88、0.86,模型的校准曲线接近理想的对角线,临床决策曲线显示临床适用性良好。结论血肌酐较基线血肌酐的变化倍数、尿量、第三代儿童死亡风险评分及C-反应蛋白4个指标构建的决策树模型对于重症儿童AKI的早期预测效果良好。Objective To establish and validate a prediction model based on least absolute shrinkage and selection operator(LASSO)regression and classification and regression tree(CART)algorithm for acute kidney injury(AKI)in PICU.Methods The prospective derivation cohort consisted of 350 critically ill children admitted to the PICU of Children′s Hospital of Soochow University from September 2020 to January 2021.The external data set consisting of 866 critically ill children admitted to the PICU of Children′s Hospital of Soochow University from February 2021 to February 2022 was employed for the external validation.Clinical data was obtained from the electronic medical record system,including demographic characteristics,laboratory data and the pediatric risk of mortalityⅢ(PRISMⅢ)score.The variables associated with AKI were identified using LASSO regression.Subsequently,a decision tree prediction model was built using the CART algorithm.The predictive value of decision tree prediction model was evaluated using the receiver operating characteristic(ROC)curve,calibration curve,and decision curve analysis.Results Among the 350 children in the derivation cohort,107(30.6%)developed AKI during the PICU stay;and of 866 children in the external validation cohort,165(19.1%)developed AKI during the PICU stay.The LASSO regression screened 16 candidate variables for further analysis,and the decision tree model ultimately identified 4 variables more closely associated with AKI,including fold change in serum creatinine from baseline,urine volume,PRISM Ⅲ,and C-reactive protein.The decision tree model exhibited high accuracy with AUC of 0.92,0.88,and 0.86 in the training,internal validation,and external validation cohorts,respectively. The model demonstrated good calibration and clinical applicability based on the calibration curve and decision curve analysis. Conclusion The decision tree model based on the 4 identified clinical indicators,including fold change in serum creatinine from baseline,urine volume,PRISM Ⅲ,and C-reactive

关 键 词：重症儿童 急性肾损伤 机器学习 预测模型 决策树 

分 类 号：R69[医药卫生—泌尿科学]