Towards Generalizability of Neural ODEs in predictive applications in pharmacometrics

Neural Ordinary Differential Equations (NeuralODEs) are a powerful emerging AI/ML approach for longitudinal modelling, notable in pharmacometrics for the potential to identify pharmacological models directly from data and greatly speed model development. These methods are very new, with examples in pharmacometrics only emerging in the last two years. While many approaches have been proposed utilizing different neural network architectures, encoder/decoder strategies, etc., it is uncertain if they are ready for broad deployment in industry or what gaps need to be addressed to enable widespread adoption. In this talk we seek to answer these questions by illustrating how different NeuralODE implementations perform on a test bed of simulated data from a range of realistic models. In particular we focus on the ability of NeuralODE models to generalize beyond training data and how much/what types of data are needed during training to achieve generalization, a crucial task for the real-world use of these models. For example, how we test how well can NeuralODE methods predict drug concentrations from different repeated dosing regimens from single dose data. We note that nonlinear pharmacokinetics can be particularly challenging to generalize from and discuss strategies for improving success in these scenarios.