Mapping Lifespan Functional Connectivity with Vision Transformer and Attention-Guided Trajectory Analysis

Abstract

Brain aging involves widespread reorganization of functional networks, yet characterizing these changes across the lifespan remains challenging. This study presents an innovative application of a Vision Transformer (ViT) to restingstate fMRI (rs-fMRI) functional connectivity for age group classification using the Human Connectome Project (HCP) Lifespan datasets. A ViT model was fine-tuned on 379×379 ROI-to-ROI connectivity matrices to perform three binary age group classifications (Young vs Old, Adult vs Aging, Young vs Aging). The model achieved accuracies of 91%, 71%, and 93% respectively under five-fold cross-validation. To interpret model decisions, attention weight maps from the final transformer encoder layer were extracted using attention rollout and thresholded with a scree plot analysis, identifying 48 high-salience connections across 28 brain regions. These features were analyzed across the lifespan using generalized additive models (GAM), and k-means clustering of the trajectories revealed three dominant lifespan connectivity patterns: Age-Declining, Mid-Life Peaking, and Age-Increasing. This study demonstrates the viability of the ViT for functional connectome analysis and shows that attention-based interpretability combined with GAM trajectory analysis can enrich insights into age-related changes in brain network connectivity, offering a framework for interpretable brain-age modeling.