自动化机器学习在剖宫产术后尿潴留预测模型中的应用  

Application of Automated Machine Learning in Prediction Model of Urinary Retention After Cesarean Section

作　　者：王芳 胡星 朱锦舟 崔欢欢 WANG Fang;HU Xing;ZHU Jin-zhou;CUI Huan-huan(Surgery Room of Suzhou Industrial Park Xinghu Hospital,Suzhou 215000,Jiangsu,China;Department of Anesthesiology,Suzhou Industrial Park Xinghu Hospital,Suzhou 215000,Jiangsu,China;Suzhou Medical College of Soochow University,Suzhou 215000,Jiangsu,China)

机构地区：[1]苏州工业园区星湖医院手术室,江苏苏州215000 [2]苏州工业园区星湖医院麻醉科,江苏苏州215000 [3]苏州大学苏州医学院,江苏苏州215000

出　　处：《医学信息》2023年第5期41-45,共5页Journal of Medical Information

基　　金：苏州市科教兴卫项目(编号:KJXW2019001)。

摘　　要：目的利用自动化机器学习方法,建立剖宫产术后尿潴留预测模型。方法选取我院2018年1月-2022年1月手术室220例行剖宫产住院产妇,根据是否发生术后尿潴留结局分为尿潴留组(38例)和无尿潴留组(182例)。比较两组生育史及术中术后临床资料,利用H_(2)O平台自动化机器学习框架,建立针对术后尿潴留结局的预测模型,通过绘制ROC曲线,计算曲线下面积(AUC)以评价模型的预测能力,并对模型特征进行可视化呈现。结果两组疼痛评分、孕前BMI、产次、剖宫产史、胎儿体重、麻醉时间、手术时间、麻醉方式、尿管拔除时间及焦虑情况比较,差异有统计学意义(P<0.05);最佳模型为梯度提升机模型(GBM),Gini值0.987,R^(2)为0.653,LogLoss为0.168;模型中重要变量包括疼痛评分、焦虑、麻醉时间、产次、麻醉方式、拔尿管时间及孕前BMI;变量SHAP特征图呈现了变量与模型整体预测的相关性,LIME反映在具体案例中变量的角色;GBM模型的ROC下面积为0.909(95%CI:0.880~0.939),准确度0.947,特异度为0.962,敏感度0.856。结论基于GBM算法的剖宫产后尿潴留预测模型显示出良好的区分能力,可作为潜在的产后并发症风险初筛工具。Objective To establish a prediction model of urinary retention after cesarean section by using automatic machine learning method.Methods From January 2018 to January 2022,220 parturients who underwent cesarean section in the operating room of our hospital were selected and divided into urinary retention group(38 cases)and non-urinary retention group(182 cases)according to whether postoperative urinary retention occurred.The fertility history and intraoperative and postoperative clinical data of the two groups were compared.The H_(2)O platform automated machine learning framework was used to establish a prediction model for postoperative urinary retention outcomes.By drawing the ROC curve,the area under the curve(AUC)was calculated to evaluate the predictive ability of the model,and the model features were visualized.Results There were significant differences in pain score,pre-pregnancy BMI,parity,history of cesarean section,fetal weight,anesthesia time,operation time,anesthesia method,catheter removal time and anxiety between the two groups(P<0.05).The best model was the gradient boosting machine(GBM),Gini value was 0.987,R^(2) was 0.653,LogLoss was 0.168;important variables in the model included pain score,anxiety,anesthesia time,parity,anesthesia method,catheter removal time and pre-pregnancy BMI.The variable SHAP feature map had shown the correlation between variables and the overall prediction of the model,and LIME reflects the role of variables in specific cases.The area under the ROC of the GBM model was 0.909(95%CI:0.880-0.939),the accuracy was 0.947,the specificity was 0.962,and the sensitivity was 0.856.Conclusion The prediction model of urinary retention after cesarean section based on GBM algorithm shows good discrimination ability and can be used as a preliminary screening tool for potential risk of postpartum complications..

关 键 词：剖宫产 尿潴留 自动化机器学习 预测模型 SHAP可视化 LIME可视化 

分 类 号：R473[医药卫生—护理学]