Pitfalls of using numerical predictive checks for population physiologically-based pharmacokinetic model evaluation.

Area of Science:

• Pharmacokinetics and Pharmacometrics
• Computational Biology
• Statistical Modeling

Background:

• Population physiologically-based pharmacokinetic (Pop-PBPK) models are crucial for drug development.
• Model qualification relies on comparing simulations to observed data.
• Numerical predictive checks (NPC) assess model prediction intervals (PIs).

Purpose of the Study:

• To evaluate the impact of data characteristics on NPC performance for Pop-PBPK models.
• To assess the influence of correlated observations, residual error, and demographic discrepancies on NPC.
• To determine the conditions under which NPC reliably qualify Pop-PBPK models.

Main Methods:

• A simulation-based study design was employed.
• Artificial pharmacokinetic (PK) datasets were generated with varying correlations and residual errors.
• Demographic distributions (subject body weights) were manipulated between observed and virtual populations.
• NPC performance was evaluated using a 90% PI for each dataset.

Main Results:

• NPC exhibited inflated type-I error rates (>0.10) when data included correlations or residual error.
• NPC performance was sensitive to demographic distribution mismatches.
• Reliable NPC performance was only observed in an idealized scenario with uncorrelated data, no residual error, and matched demographics.

Conclusions:

• Current NPC methods have limited applicability for qualifying Pop-PBPK models due to sensitivity to data characteristics.
• The presence of correlated observations, residual error, or demographic discrepancies can lead to misleading model qualification.
• Users should exercise caution when interpreting NPC results for Pop-PBPK model evaluation.