Abstract:
Normal brain aging refers to the gradual, natural changes in brain structure and function that
occur with advancing age, in the absence of neurodegenerative disease. Identifying biomarkers of normal aging
is essential for distinguishing healthy brain development from pathological conditions and for understanding
the mechanisms underlying cognitive resilience, especially given the global increase in the aging population.
Most existing studies employ deep learning for brain age prediction rather than classification into discrete age
groups, even though aging is a gradual process in which group-based analysis can better capture meaningful
structural brain changes than individual age-based analysis. In this study, we propose a deep learning
framework with integrated explainable AI to identify and interpret biomarkers of brain aging. We analyzed
diffusion-weighted and T1-weighted MRI data from healthy individuals in the publicly available IXI dataset,
categorized into three age groups: Group 1 (early adulthood, 20–39 years, n = 119), Group 2 (middle age,
40–59 years, n = 124), and Group 3 (later life, 60–79 years, n = 133). Structural brain networks were
constructed, and graph attention networks (GATs) were trained to perform pairwise classification between
Group 1 vs. Group 2 and Group 2 vs. Group 3, achieving accuracies of 86% and 87%, respectively. Using a
GNN explainer, we identified discriminant brain regions and connections contributing to aging. From early
adulthood to middle age, we found six highly robust discriminant connections indicating broad network
reorganization involving multimodal integration, social cognition, and visual–memory pathways. In contrast,
the transition from middle age to later life revealed only four discriminant connections, suggesting that many
midlife changes stabilize, but specific left-hemisphere memory and semantic/cognitive networks continue to
decline. These structural biomarkers have clear functional relevance for monitoring healthy cognitive aging.
