BrainVAE: Exploring role of white matter BOLD in preclinical Alzheimer’s disease classification

Motivation: Characteristics of white matter (WM) BOLD signals have been reported as potential indicators for classifying preclinical Alzheimer's disease (pre-AD), but their performance and interpretation remain unclear. Goal(s): Develop a novel model for pre-AD classification incorporating WM BOLD signals and quantify WM's contribution versus gray matter (GM). Approach: We introduce BrainVAE, a transformer-based Variational Autoencoder utilizing WM and GM BOLD input and compare its performance against nine models with WM-only, GM-only, and combined inputs, assessing the impact of incorporating WM information. Results: BrainVAE achieved superior accuracy with combined inputs, and WM contributed significantly to success in classification across all models. Impact: This study highlights the potential importance of including analyses of white matter BOLD signals to distinguish subjects with preclinical AD from normal controls subjects, suggesting a critical role of degenerative changes in WM in the etiology of disease.