军事知识服务平台
机器学习构建富马酸替诺福韦二吡呋酯肾损伤风险预测模型及验证
Development and Validation of a Machine Learning-Based Risk Prediction Model for Tenofovir Disoproxil Fumarate-associated Kidney Injury
摘要 目的 探讨富马酸替诺福韦二吡呋酯(TDF)治疗慢性乙型肝炎(CHB)患者发生药物相关肾损伤的影响因素,并构建风险预测模型.方法 采用多中心回顾性研究,纳入 2020 年 1 月—2024 年 8 月湖北省 5 家三甲医院接受TDF 抗病毒治疗的 CHB 住院患者.应用 Logistic 回归、随机森林、极端梯度提升(XGBoost)、支持向量机、K 近邻、决策树、朴素贝叶斯、神经网络和梯度提升机共 9 种机器学习算法构建预测模型.以精确率、准确率、召回率、ROC 曲线下面积(AUC)等指标评价模型性能.结果 共纳入患者3 595例,男占77.3%,中位年龄 54 岁,TDF 相关肾损伤发生率5.17%(186/3 595).经 Boruta 算法筛选出 10 个核心预测变量:肾小球滤过率、总胆固醇、纤维蛋白原、白蛋白、凝血酶原时间、血清氯、血清钙、总胆红素、尿酸及血清钠.基于训练集(n=2 945)构建的模型显示 XGBoost 性能最优,测试集验证结果显示精确率0.885~0.951,准确率0.939~0.975,召回率0.913~0.957,AUC 0.587~0.728,模型稳定性与临床适用性良好.结论 TDF 治疗与相关肾损伤风险显著相关,其发生受基线肾功能、合并症、联合用药、实验室检查等因素影响.基于XGBoost 算法构建的风险预测模型可用于临床早期识别高危患者,为个体化用药提供参考.
Abstract:
Objective To analyze the influencing factors for drug-induced kidney injury(DIKI)in patients with chron-ic hepatitis B(CHB)receiving tenofovir disoproxil fumarate(TDF)therapy and to construct a risk prediction model.Methods A multicenter retrospective study was conducted,enrolling hospitalized CHB patients who received TDF antiviral therapy at five tertiary hospitals in Hubei Province from January 2020 to August 2024.Nine machine learning algorithms,including Logistic Re-gression,Random Forest,eXtreme Gradient Boosting(XGBoost),Support Vector Machine,K-Nearest Neighbors,Decision Tree,Naive Bayes,Neural Network,and Gradient Boosting Machine,were employed to construct prediction models.Model per-formance was evaluated using accuracy,precision,recall andthe area under the receiver operating characteristic curve(AUC)based on test dataset metrics.Results A total of 3 595 patients were included,among whom 77.3%were male,with a median age of 54 years.The incidence of TDF-related kidney injury was 5.17%(186/3 595).Ten core predictive variables were selected by the Boruta algorithm:estimated glomerular filtration rate(eGFR),total cholesterol,fibrinogen,albumin,prothrombin time,serum chloride,serum calcium,total bilirubin,uric acid,and serum sodium.Models built on the training set(n=2945)showed that XGBoost achieved the best performance.Validation on the test set yielded a precision of 0.885-0.951,an accuracy of 0.939-0.975,a recall of 0.913-0.957,and an AUC of 0.587-0.728,indicating good model stability and clinical applicability.Conclusions TDF therapy is significantly associated with the risk of kidney injury,which is influenced by baseline renal func-tion,comorbidities,concomitant medications,laboratory findings,and other factors.The risk prediction model based on the XG-Boost algorithm can be used for early clinical identification of high-risk patients,providing a reference for individualized medica-tion.
刊 名 医药导报 ISTICPKU
年,卷(期): 2026, 45(10)
分类号: R978.7
机标分类号: R512.620.5 TP391 R969.4
在线出版日期: 2026年11月2日
基金项目: 20230116:武汉市药品不良反应监测中心科研项目,