群体药动学模型引导的机器学习用于预测抽动障碍儿童阿立哌唑清除率

Population pharmacokinetic model-guided machine learning for predicting the clearance of aripiprazole in children with tic disorders

目的:在阿立哌唑(aripiprazole,ARI)抽动障碍(tic disorders,TD)儿童群体药动学(population pharmacokinetics,PPK)模型的引导下建立机器学习模型,预测体内ARI及脱氢阿立哌唑(dehydroaripiprazole,DARI)的清除率。方法:收集81例4.8~17.3岁TD儿童口服ARI后体内ARI和DARI的药物浓度和临床资料,采用PPK模型结合最大后验贝叶斯法计算ARI和DARI的清除率,运用随机森林(RF)算法对协变量重要性进行排序,并通过RF、决策树回归(DTR)和神经网络(NNET)3种机器学习算法建立患者生理、遗传特征与ARI、DARI清除率之间的映射关系,以及计算模型预测误差和可视化检验方法评价机器学习算法的预测性能。结果:RF结果显示,影响ARI清除率的协变量重要性排序为CYP2D6代谢型>体质量>体表面积>身高>年龄>血清肌酐浓度;影响DARI清除率的协变量重要性排序为体表面积>身高>体质量>年龄>血清肌酐浓度。NNET算法用于预测ARI清除率的平均绝对预测误差和平均预测误差平方最低,清除率预测值与参比值线性回归决定系数最大。DTR、RF和NNET算法用于预测DARI清除率的预测性能相当。3种机器学习算法预测ARI和DARI清除率的平均相对预测误差和中位相对预测误差均≤20%。结论:该研究在PPK模型引导下建立了3种机器学习模型,可先验性预测TD儿童ARI和DARI清除率,为临床精准用药提供帮助。

OBJECTIVE To establish a machine learning model under the guidance of population pharmacokinetic(PPK)to predict the clearance of aripiprazole(ARI)and dehydroaripiprazole(DARI)in children of tic disorders(TD).METHODS Drug concentrations and clinical data of ARI and DARI in 81 TD children aged 4.8—17.3 years were collected.Clearance rate of ARI/DARI was calculated by PPK model plus maximum posterior Bayesian method,the importance of covariates ranked by random forest(RF)algorithm and mapping relationship between physiological and genetic characteristics and ARI/DARI clearance rate established by three machine learning algorithms of RF,decision tree regression(DTR)and neural network(NNET).Prediction error of the model was calculated and visual test method utilized for evaluating the prediction performance of machine learning algorithms.RESULTS RF results indicated that importance order of covariates affecting ARI clearance was CYP2D6 metabolic type>body weight>body surface area>height>age>serum creatinine concentration;order of importance of covariates affecting DARI clearance was body surface area>height>body weight>age>serum creatinine concentration.NNET algorithm was employed for predicting ARI clearance with the lowest mean absolute prediction error and average prediction error squared and linear regression determination coefficient between predicted value and reference value of clearance was the largest.Prediction performance of DTR,RF and NNET algorithms for predicting DARI clearance was comparable.Mean relative prediction error and median relative prediction error of three machine learning algorithms for predicting ARI/DARI clearance were both≤20%.CONCLUSION In this study,three kinds of machine learning models have been established under the guidance of PPK.It may a priori predict the clearance of ARI/DARI and provide rationales for clinical precision medicine in TD children.