Automated covariate modeling using efficient simulation of pharmacokinetics

Pharmacometric modeling plays an important role in drug development and personalized medicine. Pharmacometric covariate models can be used to describe the relationships between patient characteristics (such as age and weight) and pharmacokinetic (PK) parameters. Traditionally, the functional structure of these relationships are obtained manually. This is a time-consuming task, and consequently limits the search space of covariate relationships. The use of data-driven machine learning (ML) in pharmacometrics has the potential to automate the search for adequate model structures, which can speed up the modeling process and enable the evaluation of a wider range of model candidates. Even with moderately sized data sets, ML approaches require millions of simulations of pharmacokinetic (PK) models, which dictates the need for an efficient simulator. In this paper, we demonstrate how to automate covariate modeling using neural networks (NNs), that are trained using efficient PK simulation techniques. We apply the methodology to a propofol data set with 1031 individuals and compare the results to previously published covariate models for propofol. We use the NN as a function approximator that relates covariates to the parameters of a three-compartment PK model, and train it on dose and plasma concentration time series. Our study demonstrates that NN-based covariate modeling allows for automation of the otherwise time-consuming task of identifying which of available covariates to include in the model, and what functional mappings from these covariates to PK model parameters to consider in the model search. Additional to this saving in modeler effort, the NN-based model obtained in our clinical data set example has PK parameters within a clinically reasonable range, and slightly enhanced predictive precision than a previously published state-of-the-art covariate models for propofol model.