A Regression Approach to Visual Predictive Checks for Population Pharmacokinetic and Pharmacodynamic Models

Objectives: To demonstrate how regression techniques can be used to perform visual predictive checks (VPCs) and prediction-corrected VPCs (pcVPCs) for population pharmacokinetic (PK) and pharmacodynamic (PD) models.  This approach negates the need for specification of intervals, or “bins”, of the independent variable.

Methods: VPCs were derived using additive quantile regression (AQR). pcVPCs were generated by normalising model-derived population predictions to expected values from local regression (LOESS), then applying AQR.  Three evaluation scenarios were considered: (i) a hypothetical one-compartment model with a drug-interaction effect on clearance, where VPCs and a pcVPC were derived, (ii) two warfarin population PK-PD models, where VPCs were used for model discrimination, and (iii) a one-compartment model of phenobarbital in neonates, where a pcVPC was derived. All VPCs and pcVPCs were generated using 500 replicates of the original datasets and re-derived using bins. This work was performed in NONMEM, WFN, PsN, R and Xpose.

Results: For all scenarios, VPCs and pcVPCs via LOESS and/or AQR were accurate and similar to those derived using bins. For each scenario, it took <2 seconds to optimise and apply LOESS and/or AQR to the original datasets and ~30 seconds to apply across the 500 replicated datasets.

Conclusions: This work provides support for deriving VPCs and pcVPCs via regression techniques, which negates the need for bin specification.