Ανάπτυξη ερμηνεύσιμων μοντέλων για τη μελέτη της ετερογένειας της γήρανσης του εγκεφάλου αξιοποιώντας δεδομένα λειτουργικής απεικόνισης

Abstract

Normal brain aging is accompanied by changes in the functional network organization of the brain, which exhibit substantial heterogeneity among cognitively healthy individuals. This thesis investigates whether functional connectivity is organized into distinct and reproducible patterns beyond the systematic effects of age and sex. To this end, data from 1,814 cognitively healthy participants from the UK Biobank were analyzed. For each participant, a 156 × 156 functional connectivity matrix derived from resting-state functional magnetic resonance imaging was available. Following matrix preprocessing, a BrainNetCNN-based autoencoder was employed to extract latent representations. Standardized, 50-component principal component analysis (PCA50), and L2-normalized representations were compared using multiple clustering algorithms and selection criteria, while the final partition was evaluated through resampling-based consensus analysis. Functional interpretation of the groups combined Integrated Gradients with graph-theoretic metrics, whereas their structural and clinical characterization was based on morphometric, cognitive, clinical, and genetic measures. The results identified the L2-normalized representation and the three-subtype solution as the most comprehensively supported configuration, with a mean pairwise adjusted Rand index of 0.749 and an overall weighted stability of 0.8204. C0 was characterized by an emphasis on the Visual network and high data-driven modularity. C1 exhibited increased network strength and efficiency, with stronger involvement of connections of the Somatomotor network and the Subcortical and Cerebellar categories. C2 displayed a more distributed, multi-network pattern, with reduced local cohesion and increased participation across different functional networks. Structural and clinical comparisons did not provide evidence of broad differentiation among the subtypes in markers of neurodegeneration, brain age, cognitive performance, clinical risk factors, or genetic burden related to apolipoprotein E (APOE). Overall, the findings support the coexistence of reproducible configurations of functional network organization among cognitively healthy individuals. These configurations are more appropriately interpreted as data-driven functional connectivity profiles than as sequential stages of aging or definitively established clinical or biological entities.