Learning Interpretable Prototype Trajectories for Patients with Alzheimer’s Disease

Abstract

lzheimer’s Disease (AD) patients are known to have subtypes with distinctive progression traits. But progressions differ with distinctive patterns emerging even before patients transition to AD. Understanding these differences could help physicians understand what normal vs abnormal progressions look like. One way of characterizing differences in trends is by identifying patients that represent prototypical progressions of the defining features. We trained models to learn low-dimensional representations of all patient sequences and then, among the patients who transition to AD, used unsupervised learning to identify prototypical sequences that resembled and represented subsets of the representations. We examined trends in these subsets, searching for distinctive feature progression, and compared the prototypes to the subsets they described. We found that there are unique trends for subgroups of patients. In particular, we found a cohort that was predominantly female (79%) with distinctly high verbal memory retention relative to other cohorts. This was signifi cant because verbal memory is not a significant predictive factor, but seemed to be an axis of variation when distinguishing among subgroups. Additionally, we looked at how prototypes learned during training improve model performance compared to prototypes selected after training.