结肠镜肠道准备失败风险预测模型及Web应用程序的构建与验证  

Construction and validation of risk prediction model and web application for colonoscopy bowel preparation failure

作　　者：王甘红 陈健[2] 夏开建[3] 汤洪 奚美娟[1] 周燕婷 WANG Ganhong;CHEN Jian;XIA Kaijian;TANG Hong;XI Meijuan;ZHOU Yanting(Department of Gastroenterology,Changshu Hospital of Traditional Chinese Medicine,Changshu 215500,Jiangsu Province,China;Department of Gastroenterology,Changshu First People's Hospital;Laboratory of Medical AI and Big Data,Changshu First People's Hospital;Changshu First People's Hospital,Department of Gastroenterology,Xinzhuang People's Hospital)

机构地区：[1]常熟市中医院(新区医院)消化内科,江苏常熟215500 [2]常熟市第一人民医院消化内科 [3]常熟市第一人民医院人工智能与大数据实验室 [4]常熟市辛庄人民医院消化内科

出　　处：《中国数字医学》2025年第3期11-20,共10页China Digital Medicine

基　　金：常熟市医药卫生科技计划项目(CSWS202316);常熟市科技发展计划项目(CS202019)。

摘　　要：目的:利用机器学习算法,开发一款预测结肠镜检查前肠道准备失败风险的模型及应用程序。方法:回顾性收集拟行结肠镜检查的患者数据,纳入21个潜在预测变量,构建和内部验证传统的逻辑回归(LR)模型、机器学习(ML)模型。以受试者工作特征(ROC)曲线、校准曲线、敏感度、特异度和准确率等指标评估模型性能,并通过特征重要性图、LIME图、力图等进行可视化解释。然后,利用Python环境中的Dash库和性能最佳的模型构建Web应用程序,并在前瞻性收集的外部测试集上进行验证。结果:共纳入429例患者,其中141例(32.87%)存在肠道准备失败(BBPS评分≤5分)。基于XGBoost算法的ML模型,敏感度为0.864、特异度为0.930、准确率为0.911;在验证集中AUC值达0.910,在测试集中为0.820,性能优于传统Lasso回归模型。ML模型和逻辑回归模型共同识别的肠道准备失败变量包括便秘病史、服完泻药至检查间隔时间、术前积极运动、家属陪同、钙通道阻滞剂服用史、糖尿病、抗抑郁药服用史、年龄。结论:基于XGBoost机器学习算法构建的Web应用程序,在早期预测结肠镜肠道准备失败风险方面具有明显的临床实用性。Objective To develop a model and application for predicting the risk of bowel preparation failure before colonoscopy by using machine learning algorithm.Methods Traditional logistic regression(LR)and machine learning(ML)models were constructed and internally validated by retrospectively collecting data from patients scheduled for colonoscopy,including 21 potential predictors.The model performance was evaluated by receiver operating characteristic(ROC)curve,calibration curve,sensitivity,specificity and accuracy,and visualized by feature importance plots,LIME plots,and force plots.Then the Web application was developed using the Dash library and best-performing models in Python,and validated on a prospectively collected set of external tests.Results A total of 429 patients were included in the study,of which 141(32.87%)experienced bowel preparation failure(BBPS score≤5).The sensitivity of ML model based on XGBoost algorithm was 0.864,the specificity was 0.930,and the accuracy was 0.911.The model achieved an AUC value of 0.910 in the validation set and 0.820 in the test set,outperforming the traditional Lasso logistic regression model.The variables identified as predictors of bowel preparation failure by both the ML model and the LR model included history of constipation,interval between laxatives and examination,preoperative physical activity,family companion,history of calcium channel blocker use,diabetes,history of antidepressant use,and age.Conclusion The web application built on the XGBoost machine learning algorithm demonstrates significant clinical utility in early prediction of colonoscopy bowel preparation failure risk.

关 键 词：肠道准备 结肠镜 机器学习 预测模型 应用程序 

分 类 号：R574.6[医药卫生—消化系统] TP181[医药卫生—内科学] R319[医药卫生—临床医学]