Kekic, M.; Stepanov, O.; Wang, W.; Richardson, S.; Olabode, D.; Traynor, C.; Dearden, R.; Zhou, D.; Tang, W.; Gibbs, M.; Nowojewski, A.

Covariate selection in population pharmacokinetics modelling is essential for understanding interindividual variability in drug response and optimizing dosing. Traditional stepwise covariate modelling is often time-consuming, compared to the new machine learning alternatives. This study investigates the use of Neural Networks with Stochastic Gates for automated covariate selection, aiming to efficiently identify relevant covariates while penalizing excessive covariate inclusion. On various synthetic datasets the approach demonstrated robustness in detecting important covariates, overcoming challenges such as high correlations, low covariate frequencies, high interindividual variability and complex covariate dependencies. In real clinical data from a monalizumab study, the method successfully identified covariates that matched those found by experts. However, for tixagevimab/cilgavimab, it identified a superset of covariates, indicating a potential need for further pruning. This machine learning-based method enhances the covariate pre-selection process in population pharmacokinetics model development, offering significant time savings and improving efficiency even under challenging scenarios.