Modeling and predicting tablet dissolution slowdown using an acceleration factor approach and constrained neural network

Slowdown in dissolution of tablets upon storage can negatively impact drug release and bioavailability. Predictive stability models, developed from accelerated stability data, can provide insight into the risk of dissolution change during long-term storage. This paper compares science-based and machine learning models for predicting dissolution slowdown in tablets. An empirical model focusing on an acceleration factor (AF) was applied to open dish stability data to describe dissolution rates relative to a reference dissolution profile (i.e., unstressed tablets). As the AF model leverages "humidity extended" Arrhenius relationship and first-order decay curves that are commonly observed in "relaxation" or "equilibration" processes, using it to constrain a neural network was expected to improve prediction performance relative to the unconstrained neural network. The AF approach and the constrained neural network effectively predicted dissolution profiles of tablets stored in packaged configuration. Additionally, the AF approach identified boundary conditions below which dissolution slowdown would not occur, offering insights for controlling humidity inside the packages. These accelerated stability modeling approaches show promise as Modeling Approaches to Reimagine Stability (MARS) tools for pharmaceutical programs where dissolution slowdown presents challenges to formulation development, packaging selection, and stability evaluations.