Associate Director, Pharmacometrics Pfizer Inc. Redwood City, California, United States
Disclosure(s):
Yuchen Wang, PhD: No relevant disclosure to display
Objectives: In pediatric population PK (PopPK) model development, sparse PK data and limited covariate information can prevent accurate estimation for model parameters and their variability. We explored three methods to utilize prior model information for pediatric PopPK model development. The objectives are to summarize the model performance from each method and highlight key differences and advantages associated with these methods.
Methods: A PopPK model was previously established for Drug A using PK samples of adults and adolescents from two Phase 1, one Phase 2b/3, and one Phase 3 studies. The PK of Drug A was described adequately by a 2-compartment model with first-order oral absorption with non-stationary clearance (CL) and bioavailability (F). The model results were then utilized as prior information in the pediatric PopPK model development in three ways: 1) All existing PK parameters were fixed to the previous estimates, and pediatric effects on key PK parameters were estimated. 2) Using the PRIOR subroutine in NONMEM, the point estimates and covariance matrix from the previous model were employed as priors, and pediatric effects were estimated using the method of FOCEI. 3) Similar to 2), with a full Bayesian analysis conducted, and posterior distributions were sampled. Across all three methods, comparisons were made on parameter estimates, sampled posterior distributions, and model diagnostics.
Results: In all three approaches, pediatric effect as a factor was identified on F. In approach 1, the shrinkage of inter-individual variability (IIV) on CL was nearly 70% due to the limited sample size. For the same reason, IIV on CL was not re-estimable in approach 1. In approach 2 and 3, informative priors were implemented for all parameters except CL and IIV on CL, and the modeling fitting was improved in both approaches compared to approach 1. Additionally, approach 2 and 3 shared similar point estimates and VPC results. In approach 3, the four chains sampled via Markov chain Monte Carlo algorithm (MCMC) mixed well. Most of the observations were within the 90% posterior prediction credible intervals.
Conclusions: We compared three approaches to implement prior information in pediatric PopPK model development with limited pediatric data. Simply fixing parameters to previous estimates (approach 1) can lead to suboptimal model performance due to inherent differences between pediatric and reference populations, while utilizing previous model results as priors with FOCEI method (approach 2) or with Bayes method (approach 3) can better capture pediatric PK characteristics. While both of the approaches 2 and 3 borrow information from the previous model results and allow some adjustments in model parameterizations with similar results, approach 2 is easier and more efficient to implement since no posterior samplings nor MCMC diagnostics are needed and the point estimates in approach 2 are easily interpreted for use in subsequent simulation analysis.
