Low dimensional brain connectome representations for alzheimer's disease dementia

Alzheimer's Disease Dementia (ADD) progresses along a continuum, typically beginning with Subjective Cognitive Impairment (SCI), advancing through Mild Cognitive Impairment (MCI), and culminating in dementia. Capturing this trajectory in a compact and interpretable form remains a key challenge in brain network analysis. This thesis introduces a Graph Neural Network (GNN)-based framework that embeds structural brain connectomes into a two-dimensional manifold, offering a low-dimensional yet informative representation of disease staging. The architecture integrates a Graph Isomorphism Network (GIN) and a Graph Attention Network (GAT), followed by attention-based pooling to generate graph-level embeddings. These are projected using Principal Component Analysis (PCA) to construct a two-dimensional staging space. The learned manifold reveals a coherent spatial organization of ADD, MCI, and SCI populations, with MCI cases generally located between the two clinical extremes—consistent with a gradual model of cognitive decline. The model is validated through a detailed ablation study and random analysis, supporting the reliability of the learned embeddings. Additionally, key anatomical regions—such as the precuneus, angular gyrus, and superior parietal lobule—are identified as dominant contributors to variance across the manifold, in line with established biomarkers of ADD. Together, the proposed framework offers a robust and interpretable approach for visualizing neurodegenerative progression and supports biologically grounded models of disease staging.