Integrative Modeling of Alzheimer’s Disease Risk: A Multimodal Imaging and Computational Framework

Abstract

Alzheimer’s disease (AD) is a complex neurodegenerative disorder characterized by heterogeneous onset and progression, influenced by both genetic and environmental risk factors. While early onset accounts for 5-10% of the cases of AD, the vast majority of patients suffer from late onset AD. The APOE genotype is the strongest known genetic contributor to late-onset AD, yet its interaction with aging, diet, and immune status in shaping early brain and behavioral vulnerability still remains unclear. This dissertation presents an integrative, multimodal analysis of AD-related risk using genetically humanized APOE2, APOE3, and APOE4 mouse models. Mice were stratified by age, sex, diet, APOE and NOS2 immune background, and evaluated through in vivo neuroimaging and behavioral assays.Quantitative MRI revealed widespread structural and microstructural alterations associated with aging, APOE4 genotype, and high-fat diet. The hippocampus, olfactory bulb, thalamus, and cerebellum consistently emerged as vulnerable regions across modalities. These regions align with observed deficits in odor preference, anhedonia-like responses, and memory performance. These converging findings motivated a multimodal statistical fusion approach, Elastic Multiset Canonical Correlation Analysis (Elastic MCCA), to link structural connectomes with olfactory-related behavioral and phenotypic traits. Our results revealed distributed connectivity patterns particularly involving the olfactory tract that significantly correlated with behavioral performance and AD-related traits.To further explore age-related neurodegeneration, a deep learning framework was developed to predict brain age from structural connectomes, behavioral metrics related to learning, memory and spatial navigation, and AD risk traits. To account for complex, nonlinear relationships between data we developed a Feature Attention Graph Neural Network (FAGNN). This model utilized the graph structure of the connectome, demonstrated high predictive accuracy in comparison to simpler models, and provided interpretable edge-level attention maps. The attention scores highlighted key limbic-cortical connections—particularly the cingulum—that have been linked to accelerated aging, increased Alzheimer’s disease risk, and comorbid mood disorders such as depression.This thesis integrates imaging, behavior, and computational modeling to uncover early markers of AD risk. The developed methods and findings can serve to further computational research in aging and AD.