PhD thesis defense to be held on December 16, 2024, at 17:00 (virtually)
Thesis title: Unravel Heterogeneity in Human Brain Aging with Neuroimaging and Artificial Intelligence: Clinical, Lifestyle, Cognitive, and Genetic Associations.
Abstract: The present thesis investigates the complex and multifaceted brain changes associated with aging, which lead to cognitive decline and the development of Alzheimer’s disease (AD). Utilizing and advancing state-of-the-art machine learning techniques and harnessing large-scale datasets, distinct and homogeneous imaging patterns linked to various brain aging trajectories are identified. The overarching objective is to disentangle the neuroanatomical heterogeneity across the brain aging spectrum, examining the variability driven by AD-related degeneration and the influence of co-existing pathologies, lifestyle, environmental, and genetic risk factors. Additionally, this work seeks to leverage the identified dimensions of brain changes to predict future cognitive decline and clinical progression, providing insights that may ultimately improve early diagnosis, risk stratification, and intervention strategies in aging and neurodegenerative diseases.
First, the heterogeneity of neuroanatomical brain changes in aging at early asymptomatic phases is investigated by leveraging recent advancements in deep learning and big data analytics. Collectively, there has been an increasing understanding of the neurobiological processes related to various neuropathologies that affect the human brain, including AD and cerebrovascular disease. However, little is known about how people, at the individual level, transition from normal aging to pathologic manifestation. This knowledge gap is partly due to the lack of sufficiently large-scale neuroimaging datasets and the tools to model and validate such complex processes. Unraveling the neuroanatomical heterogeneity in aging at early stages before the emergence of clinical symptoms may provide prognostic information about susceptibility to or presence of neurodegenerative disease and influence patient management and clinical trial recruitment. To this end, a novel semi-supervised clustering method based on generative adversarial networks (GAN), termed Smile-GAN, is applied to cross-sectional T1- and T2-weighted magnetic resonance imaging (MRI) data, consolidated by the iSTAGING (imaging‐based coordinate SysTem for AGIng and NeurodeGenerative diseases) consortium for a large-scale and diverse harmonized multi-cohort sample of middle-to-late age cognitively unimpaired individuals (N=27,402). Neuroanatomical subgroups are independently examined in four decade-long age intervals spanning 45-85 years; decade intervals are used to mitigate age-related effects during clustering. The derived subgroups are correlated with genetic and lifestyle risk factors, biomedical measures, and cognitive decline trajectories.
Three subgroups, consistent across decades, are identified within the cognitively unimpaired population. Briefly, a typical aging subgroup characterized by low atrophy and white matter lesions and a genetic profile protective against vascular disease and two accelerated aging subgroups are found: one characterized by elevated cardiovascular disease risk factors, disruption of white matter integrity, and increased cerebral amyloid β deposition, while the other displays diffuse and severe atrophy, likely driven by lifestyle and exposure factors. These subgroups may reflect differential susceptibility to AD and other neurodegenerative conditions, cognitive decline, and clinical progression.
Next, a novel methodology for the study of heterogeneity is introduced. Unlike current approaches relying solely on cross-sectional data, thus neglecting dynamic observations of pathological changes, the proposed approach, termed Coupled Cross-sectional and Longitudinal Non-negative Matrix Factorization (CCL-NMF), develops a mutually constrained NMF framework to delineate components that encapsulate distinct patterns of brain alterations derived jointly from cross-sectional and longitudinal data. A cross-sectional map (C-map) captures the cumulative brain changes due to aging or disease over long periods inferred from broader population-level comparisons, while a longitudinal map (L-map) captures the dynamic patterns of brain change on an individual basis. CCL-NMF identifies components shared by C- and L-maps based on the assumption that an aging or disease effect estimated cross-sectionally at a population level should be compatible with dynamic brain changes captured by longitudinal data. It also estimates corresponding coefficients (loadings), representing the degree of expression of each component from each individual by optimizing the reconstruction of both data types, thereby capturing the complex interplay between static and dynamic aspects of brain alterations. Notably, CCL-NMF avoids rigid classification into mutually exclusive categorical subtypes, allowing individuals to exhibit varying degrees of co-expression across multiple patterns.
The proposed methodology is formulated in a general framework, enabling its application to analyses of heterogeneity of any disease characterized by monotonic brain alterations (e.g., gradual gray matter atrophy and cerebrospinal fluid expansion, progressive white matter lesion accumulation, or increasing deposition of neuropathologies such as amyloid and tau). This thesis applies CCL-NMF to parse the heterogeneity of aging-related atrophy using T1-weighted MRI data. The method is first validated using semi-synthetic data with predefined atrophy patterns and severity levels. Then, it is applied to delineate the heterogeneity in an aging population (N=48,949) with a healthy middle-aged cohort (N=977) as a reference. Both populations are drawn from the iSTAGING consortium. The identified components are correlated with AD biomarkers, cognition, cardiovascular disease risk factors, and disease progression, revealing meaningful patterns closely aligned with clinical phenotypes, highlighting the method’s ability to offer deeper insights into the biological processes underlying aging. Importantly, by deriving individualized expression levels across these components, the approach facilitates personalized therapeutic interventions tailored to individual patient profiles, paving the way for more targeted and effective treatment strategies.
Moreover, comparisons with state-of-the-art deep learning models applied to the same dataset demonstrate that the CCL-NMF components provide improved predictive power for biomarkers and clinical variables, refining our understanding of brain aging pathways. Finally, the model facilitates out-of-sample application through regression-based loading estimation, broadening its utility in research and clinical contexts.
Supervisor: Professor Konstantina S. Nikita
PhD Student: Ioanna Skampardoni