Abstract

Alzheimer’s disease (AD) is a fatal neurodegenerative disease that induces deficits in multiple cognitive domains and reduces the overall quality of life. Among various physiological processes, AD is characterized by structural changes to gray matter incurred from gross neuronal loss due to the accumulation and spread of tau proteins across the brain. Neuroimaging modalities such as T1-weighted magnetic resonance imaging (MRI) and tau positron emission tomography (PET) can quantify these disease processes. Though tau neurofibrillary tangles (NFTs) follow a stereotypical pattern following the Braak staging scheme, the recent network degeneration hypothesis (NDH) has suggested that NFTs spread selectively along functional networks of the brain. This thesis investigates whether it is beneficial to incorporate network-level information into an ROI-based classification
model. To this end, we employed a Bayesian hierarchical multinomial logistic regression model and developed four statistical models with increasing levels of complexity. We assessed each model’s out-of-sample predictive ability and goodness of fit using the leave-one-out cross-validation and posterior predictive check procedures. Our results
validated the initial hypothesis and found that a hierarchical model incorporating ROI-level and network-level information was best. Furthermore, posterior predictions from the best model revealed laterality patterns across hemispheres within specific regions and
networks. Moreover, the predictions also suggested that tau precedes cortical atrophy for particular regions and networks. Lastly, tau-PET predictions suggested how though the disease may globally affect functional networks, the regions that comprise a given network display their own patterns of heterogeneity and progression.