Abstract
INTRODUCTION: Like gray matter (GM), white matter (WM) BOLD functional signals change in preclinical AD. However, the potential of WM BOLD for identifying preclinical AD remains underexplored. METHODS: We developed BrainVAE, a transformer-based variational autoencoder with interpretability, to classify preclinical AD and normal controls using resting-state fMRI data. We benchmarked BrainVAE against nine alternative models under three input configurations: WM-only, GM-only, and combined WM+GM. Interpretability analysis was also performed to investigate each brain region's contribution to the classification task. RESULTS: BrainVAE outperformed other models and performed well (accuracy = 83.42%, F1-score = 91.62%, AUC = 64.50%) using the combined input compared to WM-only and GM-only. Specific WM bundles--including corpus callosum, fornix, and corticospinal tract-were among the most influential features contributing to the classification. DISCUSSION: Incorporating WM BOLD signals improves the distinction of preclinical AD from controls, underscoring the potential role of WM BOLD features for detecting early-stage AD.