Stochastic Gate Neural Networks for Automatic Covariate Selection in Pharmacometrics Population Modeling

Marija Kekic¹, Andrzej Nowojewski², Oleg Stepanov³, Sam Richardson², Itziar Irurzun³, Diansong Zhou⁴, Weifeng Tang⁴, Richard Dearden²

¹Imaging & Data Analytics, Clinical Pharmacology & Safety Sciences, R&D BioPharmaceuticals, AstraZeneca, Barcelona, Spain

²Imaging & Data Analytics, Clinical Pharmacology & Safety Sciences, R&D BioPharmaceuticals, AstraZeneca, Cambridge, UK

³Clinical Pharmacology & Quantitative Pharmacology​, Clinical Pharmacology & Safety Sciences, R&D BioPharmaceuticals, AstraZeneca, Cambridge, UK

⁴Clinical Pharmacology & Quantitative Pharmacology​, Clinical Pharmacology & Safety Sciences, R&D BioPharmaceuticals, AstraZeneca, Gaithersburg, MD, USA

 

Abstract:
Patients’ covariates can significantly influence drug absorption, distribution, and elimination processes, making their identification a crucial step in pharmacometrics model development. Stepwise covariate selection (SCM) is the commonly used tool to automatically identify influential covariates. However, this procedure has several drawbacks, e.g., the requirement to invoke the parameter estimation repeatedly, leading to extended computation times.

Supervised Machine Learning has emerged as a promising alternative for covariate selection.
In this study, we employ Neural Networks with Stochastic Gate layer for automatic feature selection. We evaluate the model performance on simulated data with a known covariate structure. Initially, we establish the proof of concept using a simple one-compartment pharmacokinetic (PK) model, where we vary the number of relevant covariates, their effect sizes, and the correlation coefficient between them.

Finally, we present the performance of this approach on data simulated for a study related to the development of a PK model for a specific compound.