Abstract
Currently employed methods for qualifying population physiologically-based pharmacokinetic (Pop-PBPK) model predictions of continuous outcomes (e.g., concentration-time data) fail to account for within-subject correlations and the presence of residual error. In this study, we propose a new method for evaluating Pop-PBPK model predictions that account for such features. The approach focuses on deriving Pop-PBPK-specific normalized prediction distribution errors (NPDE), a metric that is commonly used for population pharmacokinetic model validation. We describe specific methodological steps for computing NPDE for Pop-PBPK models and define three measures for evaluating model performance: mean of NPDE, goodness-of-fit plots, and the magnitude of residual error. Utility of the proposed evaluation approach was demonstrated using two simulation-based study designs (positive and negative control studies) as well as pharmacokinetic data from a real-world clinical trial. For the positive-control simulation study, where observations and model simulations were generated under the same Pop-PBPK model, the NPDE-based approach denoted a congruency between model predictions and observed data (mean of NPDE = - 0.01). In contrast, for the negative-control simulation study, where model simulations and observed data were generated under different Pop-PBPK models, the NPDE-based method asserted that model simulations and observed data were incongruent (mean of NPDE = - 0.29). When employed to evaluate a previously developed clindamycin PBPK model against prospectively collected plasma concentration data from 29 children, the NPDE-based method qualified the model predictions as successful (mean of NPDE = 0). However, when pediatric subpopulations (e.g., infants) were evaluated, the approach revealed potential biases that should be explored.