文章摘要
基于机器学习构建全髋置换术患者术后谵妄风险的预测模型
Prediction model for the risk of postoperative delirium in total hip arthroplasty patients based on machine learning
  
DOI：10.12089/jca.2025.10.003
中文关键词: 全髋置换术  术后谵妄  机器学习  Boruta算法  Shapley加性解释
英文关键词: Total hip arthroplasty  Postoperative delirium  Machine learning  Boruta's algorithm  Shapley additive explanations
基金项目:安徽医科大学校科研基金（2023xkj083）；蚌埠医科大学科技项目自然科学类（2023byzd214）
作者单位E-mail
王小锋 236000,安徽省阜阳市,阜阳市人民医院麻醉科  
蔡宁 236000,安徽省阜阳市,阜阳市人民医院麻醉科  
王建彭 236000,安徽省阜阳市,阜阳市人民医院麻醉科  
杨芳芳 236000,安徽省阜阳市,阜阳市人民医院麻醉科  
王秋锋 236000,安徽省阜阳市,阜阳市人民医院麻醉科 qiufeng1253@163.com 
摘要点击次数: 1684
全文下载次数: 460
中文摘要:
      
目的：基于Shapley加性解释（SHAP）和机器学习（ML）模型构建预测全髋置换术（THA）患者术后谵妄（POD）风险预测模型并开发在线应用程序。
方法：选择2023年1月至2024年11月行THA的患者277例，男76例，女201例，年龄≥65岁，BMI 18.5～30.0 kg/m2，ASA Ⅰ—Ⅲ级。通过Boruta算法筛选THA患者发生POD的风险因素。基于风险因素构建、训练和比较6种ML模型，使用受试者工作特征（ROC）曲线和校准曲线评估6种ML模型，筛选最佳预测性能模型。使用SHAP值对ML模型进行解释和可视化，并使用Shiny等R包开发预测THA患者POD风险的在线应用程序。
结果：有65例（23.8%）患者发生POD。Boruta算法筛选出麻醉时间、C-反应蛋白（CRP）浓度、年龄、顺式阿曲库铵用量、七氟醚用量、术中失血量、手术时间以及白蛋白（Alb）浓度是THA患者发生POD的风险因素。6种ML算法中，ROC曲线和校准曲线证实极端梯度提升（XGBoost）模型预测POD风险性能最高。基于SHAP值附加解释和可视化XGBoost模型能以极高准确度预测POD风险。在线应用程序网址https://mldynamic.shinyapps.io/PD-web/。
结论：麻醉时间、CRP浓度、年龄、顺式阿曲库铵用量、七氟醚用量、术中失血量、手术时间以及Alb浓度是THA患者发生POD的风险因素。基于SHAP值解释的XGBoost模型有极高的预测性能，基于此开发的在线应用程序能帮助使用者快捷计算THA患者POD风险，优化治疗方案。
英文摘要:
      
Objective: To develop of a machine learning (ML) model based on Shapley additive explanations (SHAP) for predicting postoperative delirium (POD) risk in total hip arthroplasty (THA) patients and create an online application.
Methods: A total of 277 patients, 76 males and 201 females, aged ≥ 65 years, BMI 18.5-30.0 kg/m2, ASA physical status Ⅰ-Ⅲ, who underwent total hip arthroplasty (THA) between January 2023 and November 2024 were enrolled in this study. Risk factors for POD in THA patients were identified through Boruta algorithm-based feature selection. Construct, train and compare 6 ML models based on risk factors. The 6 ML models were evaluated using receiver operating characteristic (ROC) curves and calibration curves to screen the best predictive performance models. The ML models were interpreted and visualized using SHAP values, and R packages such as Shiny were used to develop a online applications for predicting POD risk in THA patients.
Results: Among the study cohort, 65 patients (23.8%) developed POD. The Boruta algorithm identified anesthesia duration, C-reactive protein (CRP), age, cisatracurium, sevoflurane, intraoperative blood loss, operative time, and albumin (Alb) as risk factors for POD in THA patients. Among the 6 ML algorithms, the ROC curves and calibration curves confirmed that the extreme gradient boost (XGBoost) model had the highest performance in predicting POD risk. Based on the SHAP value additional interpretation and visualization XGBoost model can predict POD risk with very high accuracy. The online applications interface is shown in https://mldynamic. shinyapps. io/PD-web/.
Conclusion: Anesthesia duration, CRP, age, cisatracurium, sevoflurane, intraoperative blood loss, surgical duration, and Alb are risk factors for POD in THA patients. The XGBoost model based on the interpretation of SHAP values has an extremely high predictive performance, and the online applications developed based on this can help users to quickly calculate the risk of POD in THA patients and optimize the treatment plan.
查看全文   查看/发表评论  下载PDF阅读器
关闭