基于机器学习的他克莫司剂量/体质量调整谷浓度个体化预测模型  

Individualized prediction model of tacrolimus dose/weight-adjusted trough concentration based on machine learning approach

作　　者：蒋卉 唐亮 汪鑫 姜凡[1] 王德光[3] 蓝晓锋[1] 解翔[1] Jiang Hui;Tang Liang;Wang Xin;Jiang Fan;Wang Deguang;Lan Xiaofeng;Xie Xiang(Dept of Ultrasound,The Second Hospital of Anhui Medical Universty,Hefei 230601;Dept of Urology,The Second Hospital of Anhui Medical Universty,Hefei 230601;Dept of Nephrology,The Second Hospital of Anhui Medical Universty,Hefei 230601)

机构地区：[1]安徽医科大学第二附属医院超声医学科,合肥230601 [2]安徽医科大学第二附属医院泌尿外科,合肥230601 [3]安徽医科大学第二附属医院肾脏内科,合肥230601

出　　处：《安徽医科大学学报》2025年第2期344-350,共7页Acta Universitatis Medicinalis Anhui

基　　金：安徽省自然科学基金面上项目(编号:2008085MH244)。

摘　　要：目的利用机器学习(ML)算法建立他克莫司剂量/体质量调整谷浓度(TAC C0/D)预测模型。方法收集来自72例接受同种异体肾脏移植手术患者的264份TAC血药浓度监测数据。分析人口统计学资料、临床特征、联合用药及超声特征参数对TAC C0/D的影响。选择在TAC C0/D的单因素分析中P<0.05的特征纳入随机森林(RF)算法进行重要性分析,筛选重要特征,并利用偏依赖图对重要特征进行解释。使用RF、支持向量回归(SVR)、极端梯度提升(XGBoost)、决策树(DT)和人工神经网络(ANN)5种ML算法建立TAC C0/D预测模型,并对性能最佳的RF模型进行超参数调优。结果保留了重要性评分>5的10个特征变量并纳入ML模型,分别是:谷氨酰胺转氨酶、红细胞计数、尿素氮、体质量、血肌酐、肾段动脉阻力指数、肾主动脉阻力指数、血细胞比容、肾盂杨氏模量值和移植术后时间。偏依赖图显示筛选出的10个重要变量均与TAC C0/D呈正相关。调优后的RF模型性能优于其他模型,其R 2为0.81,RMSE为43.93,MAE为29.97。结论ML模型预测TAC C0/D的表现良好,创新性地利用偏依赖图对特征进行了解释。调优后的RF模型性能最佳,为肾移植患者TAC个体化调整用药提供了新的手段。Objective To utilize machine learning(ML)algorithms to develop accurate and effective prediction models for TAC dose/weight-adjusted trough concentration(C0/D).Methods Data were collected on 264 TAC blood concentration monitoring data from 72 patients undergoing kidney transplantation.The effects of population statistical data,clinical features,combined medication,and ultrasound feature parameters on TAC C0/D were analyzed.Features with a significance level less than 0.05 in the univariate analysis of TAC C0/D were selected for inclusion in the random forest(RF)algorithm to identify significant features.These features were interpreted using partial dependency plots.Five ML algorithms,including RF,support vector regression(SVR),extreme gradient boosting(XGBoost),decision trees(DT)and artificial neural networks(ANN),were employed to establish the TAC C0/D prediction model.Hyper-parameter tuning was performed on the RF model that performed the best.Results Ten characteristic variables with importance scores>5 were retained and included in the ML model:transglutaminase,red blood cell count,blood urea nitrogen,weight,serum creatinine,renal segmental arterial resistance index,renal aortic resistance index,hematocrit,renal pelvic Young′s modulus value,and time after transplantation.The partial dependence plots showed that all 10 important variables screened were positively correlated with TAC C0/D.The tuned RF model outperformed the other models with a R 2 of 0.81,a RMSE of 43.93,and a MAE of 29.97.Conclusion The ML models demonstrate good performance in predicting TAC C0/D and provide innovative interpretations using partial dependence plot.The optimized RF model shows optimal performance and offers a novel tool for individualized medication adjustment for TAC in renal transplant patients.

关 键 词：异体肾脏移植 他克莫司 多普勒超声 偏依赖图 机器学习 个体化药物治疗 

分 类 号：R617.5[医药卫生—外科学]