Martensen, C. J.; Korsbo, N.; Ivaturi, V.; Sager, S.

In pharmacometrics, developing and selecting models is crucial for quantitatively assessing drug-biological interactions, treatment planning, and gaining insights into underlying processes. These validated models are essential for predictive analytics and strategic decision-making in drug development and clinical practice. Unlike traditional methods, machine learning (ML) offers a data-driven alternative to conventional, first-principle approaches. This paper presents an automatic method to derive unknown or uncertain (sub)models using longitudinal, heterogeneous data. Initially, we employ deep nonlinear mixed effect (DeepNLME) to train a neural network as a universal approximator, which then generates a parameterized representation of the underlying process. Subsequently, we apply symbolic regression to identify a set of potential models expressed as equations. Within the study parameters, the proposed method outperforms the baseline models and demonstrates the validity of both the DeepNLME approach and the symbolic regression approach.