Leveraging machine learning strategies for nonlinear mixed effects model selection (using pyDarwin)

The application of machine learning approaches to pharmacokinetic and pharmacodynamic measurements is becoming more widespread. Most of these approaches are used to predict these measurements without providing inferential insights. This presentation focuses on the application of machine learning for nonlinear mixed effects PK model selection. It was recognized almost three decades ago that there are significant interactions between structural, statistical and covariate models that result in very different inferences with respect to model structure, covariate and random effects (Wade 1994). Classical stepwise approaches to model development are particularly susceptible to these interactions. Previous applications of a genetic algorithm based model search strategy illustrated superiority of this approach to stepwise model evaluation based on typical model fitness considerations (Bies 2006, Sherer 2012, Sale 2015). The machine learning strategies described in this talk present an alternative means of evaluating the model search space that may provide greater insight into these interactions while optimizing the numerical characteristics of the solutions obtained. This is illustrated using the implementation of these techniques (Random Forest, Gaussian Process/Bayesian Optimization, Genetic Algorithm among others) in the open source pyDarwin software.