Statistical Machine Learning to Develop Connectome-Based Biomarkers of Alzheimer's Disease and Related Dementias

University Of Washington, Seattle WA

Investigators

Abstract

Project Summary Alzheimer’s disease (AD) is a prevalent chronic neurodegenerative disorder, projected to affect between 11 to 16 million individuals in the US by 2050. Neurodegeneration in AD initiates with the formation of amyloid beta (A) plaques decades before observable symptoms. This latent, pre-clinical phase represents a crucial "window of opportunity" when new A lowering therapies can slow the disease progression and mitigate its clinical impacts. However, existing biomarkers of preclinical AD, such as cerebrospinal fluid (CSF), plasma and positron- emission tomography (PET), face significant limitations. This underscores an urgent need for readily available and noninvasive biomarkers capable of detecting AD at its latent, preclinical, stage when disease-modifying therapies are most effective. Resting state functional MRI (rs-fMRI) is gaining recognition for its potential to provide non- invasive biomarkers for AD pathology, particularly through the assessment of alterations in functional connectivity (FC) dynamics and networks related to A deposition. However, current computational approaches for analyzing rs-FC networks and their dynamics have notable limitations, including the use of ad-hoc or black-box analytical methods, over looking heterogeneity in rs-fMRI data and dynamic FC across subjects and populations, and limited sample sizes, impacting the applicability, generalizability, and replicability of existing biomarkers and their ability to reveal underlying disease mechanisms. This project aims to rectify the deficiencies in current computational methods and is grounded in the fundamental hypothesis that improved biomarkers for early detection of AD and related dementias (ADRD) can be obtained by leveraging the dynamics of rs-fMRI data and FC, while accounting for heterogeneity across subjects and studies. To achieve this goal, the first aim develops a new model to comprehensively capture the dynamics of FC and offer robust biomarkers for AD pathology by assessing the accumulation of 𝐴. To account for individual level heterogeneity, the second aim develops a multivariate dynamic models featuring both local and global structures to incorporate heterogeneous brain connectivity among different brain states in different individuals. The third aim then develops effective techniques for detecting alterations in brain connectivity networks, while accounting for subject-level heterogeneity through random effects, helping uncover underlying processes of ADRD initiation and progression. Finally, to address the limited sample sizes in individual studies, the fourth aim proposes an innovative transfer learning framework that leverages the inherent similarities among multiple datasets to improve the reliability of connectome-based ADRD biomarkers. Upon evaluation and validation, the above methods will be implemented as efficient, open-source software tools in form of R-packages and python libraries, accompanied with extensive documentations, illustrative examples, and interactive visualization capabilities, to maximize the adoption of the proposed methods by the broader community.