Machine Learning Approaches Combined with Power Analysis to Improve the Prediction of Population Pharmacokinetic Modeling

This study aims to determine if integrating machine learning (ML) approaches with power analysis for correlated covariates enhances predictive performance in modeling. Two drugs were selected to apply population pharmacokinetics methodology. Parameter-covariate relationships were estimated using stepwise covariate modeling (SCM). Power analysis identified correlated true covariates for each parameter chosen by SCM. If correlated true covariates were identified, ML methods further selected the most significant covariate. The calibrated model utilizing this significant covariate was compared to the SCM model, evaluating performance using relative error. Non-compartment analysis (NCA) calculated individual reference AUC, enabling comparison of AUCs from calibrated and SCM models with the reference AUC to assess predictive performance. Results showed that calibrated models outperformed SCM models, with R² values exceeding 80%. Overall, the calibrated models predicted AUC more accurately than the SCM models. Power analysis helped eliminate false covariates among correlated ones, while ML methods provided criteria for selecting covariates from the identified true covariates. Thus, combining these methods can enhance population pharmacokinetic model predictions.