基于XGBoost构建老年慢性肾脏病患者万古霉素血药浓度预测模型OA
Development of an XGBoost-based prediction model for vancomycin plasma concentrations in elderly patients with chronic kidney disease
目的 基于机器学习技术构建老年慢性肾脏病患者万古霉素稳态谷浓度的预测模型,并利用SHapley可加性解释(SHAP)方法阐明模型中特征的重要性.方法 回顾性分析2013年11月至2023年5月在医院接受万古霉素治疗并进行治疗药物监测(TDM)的革兰阳性菌感染的老年慢性肾脏病患者.采用XGBoost算法构建万古霉素稳态谷浓度预测模型,并利用SHAP值量化各特征对模型预测的贡献度(特征重要性)及其作用方向(归因分析),评估模型的应用价值.结果 最终模型的预测性能指标:模型拟合度(R2)为0.530,平均绝对误差(MAE)为4.248 mg·L-1,均方误差(MSE)为28.822(mg·L-1)2.模型预测的绝对准确度和相对准确度分别为67.27%和60.91%.SHAP分析结果显示,万古霉素单次给药剂量和肌酐水平是影响谷浓度的最重要因素.其他影响因素包括给药后采血时间、丙氨酸氨基转移酶、血小板计数、合并用药情况、尿素氮、胱抑素C、总蛋白、血红蛋白和身体质量指数.结论 基于机器学习技术的万古霉素稳态谷浓度预测模型及SHAP分析,可为老年慢性肾脏病患者万古霉素个体化用药提供高效、经济的辅助工具,有助于临床医师优化给药方案,提升治疗效益与用药安全性.
AIM To construct a predictive model for vancomycin steady-state trough concentrations in elderly patients with chronic kidney disease using machine learning techniques,and elucidate the importance of features in the model using SHapley Additive exPlanations(SHAP).METHODS A retrospective analysis was conducted on elderly patients with chronic kidney disease and Gram-positive bacterial infection who received vancomycin therapy and underwent therapeutic drug monitoring(TDM)from November 2013 to May 2023.An XGBoost-based predictive model for vancomycin steady-state trough concentrations was developed.SHAP values were subsequently employed to quantify feature importance(contribution magnitude to predictions)and directional impact(attribution analysis)of each variable,followed by evaluation of the model's clinical applicability.RESULTS The final model demonstrated the following performance metrics:model fit(R2)=0.530,mean absolute error(MAE)=4.248 mg·L-1,and mean squared error(MSE)=28.822(mg·L-1)2.The absolute and relative prediction accuracies were 67.27%and 60.91%,respectively.SHAP analysis identified vancomycin single-dose and creatinine level as the most critical factors influencing trough concentrations.Other significant contributors included time of blood collection after drug administration,alanine aminotransferase,platelet count,concomitant medications,urea nitrogen,cystatin C,total protein,hemoglobin,and body mass index.CONCLUSION The machine learning-based predictive model for vancomycin steady-state trough concentration,along with SHAP analysis,provides an efficient and cost-effective adjunctive tool for individualised vancomycin dosing in elderly patients with chronic kidney disease.It assists clinicians in optimising dosing regimens,thereby improving therapeutic outcomes and medication safety.
张馨;谢林利;马欢;廖燕;熊丽蓉;马攀
陆军军医大学第一附属医院药学部,重庆 400038陆军军医大学第一附属医院药学部,重庆 400038陆军军医大学第一附属医院药学部,重庆 400038陆军军医大学第一附属医院药学部,重庆 400038陆军军医大学第一附属医院药学部,重庆 400038陆军军医大学第一附属医院药学部,重庆 400038
医药卫生
万古霉素血药浓度XGBoost模型SHapley可加性解释老年慢性肾脏病患者
vancomycinplasma concentrationXGBoost modelSHapley Additive exPlanationselderly patients with chronic kidney disease
《中国临床药学杂志》 2026 (8)
684-691,8
国家卫生健康委医院管理研究所项目(编号NIHAYSZX2556)重庆市临床药学重点专科建设项目(渝卫办发[2021]52号)
评论