Symbolic neural networks for automated covariate modeling in a mixed-effects framework

Mixed-effects models are used to describe the inter-patient variability in drugs. Modeling of these variabilities include both fixed and random effects. Fixed effects relate covariates such as age and weight to compartment volumes and clearances, whereas random effects account for unexplained variability. Traditionally, the development of fixed effects models is an inefficient process where covariate relationships are evaluated in a step-wise manner. In this study, we implemented a symbolic neural network (SNN) to automate the development of a fixed effects model and used it to develop a population pharmacokinetic model for propofol. With the SNN, we can find covariate relationships that are traditionally not evaluated. Then, we apply random effects and estimate parameters in the standard mixed-effects modeling framework. Our final model shows comparable predictive performance to a published model for propofol, despite having fewer covariates and model parameters.