Bayesian Pharmacokinetic Extrapolation from Dense Adult to Sparse Pediatric Data
Sebastian Weber
Novartis Pharma AG
Objectives: To characterize the pharmacokinetics (PK) of pediatric patients has recently gained much in importance due to regulatory requirements. Often we find ourselves in the situation of extensive data on adults while data on pediatric patients are very sparse due to: (i) adult centric drug development, (ii) a changing metabolism due to aging and (iii) ethical and practical difficulties in sample collection.
In this work we explore how a Bayesian approach can bridge the two populations. The objective is to establish an approach suitable to extrapolate between adult and pediatric PK models.
Methods: The pooled study data on adult patients compromises about 600 patients. 25% of the PK samples were collected in the absorption and distribution phase, 50% around the Ctrough at 12h and 25% in the elimination and washout phase. The population PK model adequately describing the data was a 2cmt model with a time-changing clearance. The pediatric data included only 22 patients in a wide age range. The PK samples were timed around 12h after dosing while few measurements were taken in the absorption phase and almost no measurements in the elimination or washout phase.
The base adult model was fit using NONMEM and has been converted into a Bayesian model using Stan [2]. In the analysis we compare the frequentist NONMEM result to the Bayesian Stan posterior estimate. The key step is to use the adult posterior as prior in the pediatric analysis with discounting.
Results: A reparametrized 2cmt PK model with standard allometric 3/4 power-scaling [1] using weight combined with weakly-informative priors was able to robustly fit and describe the adult data set. The derived posterior has been discounted by assuming that weight standardized estimates may at most deviate by 100% from the adult parameter for the pediatric analysis.
The pediatric data slightly updates the adult estimates insofar that the clearance is increased. The respective frequentist result on the pediatric data set revealed that the 2cmt model was not identifiable due to lack of data.
Conclusions: The pediatric data set was well compatible with the adult model. The analysis revealed that the clearance for the pediatric population is likely increased which leads to a decreased steady-state concentration.
In summary, the Bayesian approach warranted model identifiability of a complex PK model despite a very sparse data situation. This was possible by using the discounted adult posterior as prior in the pediatric analysis.
References:
[1] West GB, Brown JH, Enquist BJ, Science. 1997;276(5309):122–6
[2] Stan Development Team (2017). Stan: A C++ library for probability and sampling.
