Machine Learning for Modeling Progression and Heterogeneity in Alzheimer's Disease

Machine Learning for Modeling Progression and Heterogeneity in Alzheimer’s Disease Raghav Tandon 185 Pages Directed by Dr. Cassie S. Mitchell This thesis uses existing and novel machine learning algorithms to study three related questions in Alzheimer’s disease (AD) research. These questions are - 1) What are the proteomic markers of change in AD? 2) In what sequence do changes take place as AD progresses?, and 3) What are the heterogeneities in disease progression among the AD population? These questions are important due to some key challenges faced in AD research. Disease aetiology in AD is incompletely understood which has prevented the development of effective therapies. Though recent drug developments have resulted in disease modifying therapies (Lecanemab), these drugs have been found to be effective in early disease stages. Statistics show that early AD detection is a challenge and only a small fraction of people with AD are diagnosed early enough when available disease modifying treatments can show effectiveness. Further, AD is known to have a heterogeneous clinical presentation which complicates its clinical management. Finally, given the long disease evolution timelines which span several decades, collecting data throughout the disease evolution process has practical challenges and limitations. The three questions this thesis aims to address are relevant to the above challenges. Identifying proteomic markers of change in the disease helps to elucidate the proteopathic changes underlying AD. The identified markers also have diagnostic utility and can be potentially useful for early disease detection. Understanding the sequence of progressive changes helps to develop new disease staging systems which are data driven and serve as early warning systems for the disease. Studying heterogeneities in AD progression helps to understand variations in important clinical variables such as rate of progression and cognitive decline, comorbidities, age of onset, and markers of disease pathology. This understanding about disease heterogeneity can be useful in two important ways. First, it can aid clinical decision making which helps to move towards a personalized medicine approach in AD. Second, it can be useful in clinical trial designs to control for disease related phenotypic heterogeneities and making the trial more sensitive to treatment effects. These questions are answered using machine learning models applied to cross-sectional data, which makes the approaches data economical. Despite using limited data, the learned models generalize well to new data from external sources and future follow up visits from patients, which shows the potential utility of these approaches in solving important clinical challenges. The first question (“markers of change?”) is addressed using classic supervised learning approaches. The second (“sequence of disease related changes?”) develops new probabilistic generative algorithms which attempt to construct a disease progression trajectory from a large number of disease related neuroimaging biomarkers. Finally, the third question (“heterogeneities in progression”) extends the previous generative algorithm to model disease progression to take place over one of potentially many disease trajectories. This thesis makes two main contributions. First, it contributes to a better understanding of AD. More specifically, these are – 1. Identifying the role of dysfunctional sugar metabolism in disease development. 2. Identifying important peptide markers which can be useful for early disease detection. 3. Inferring a trajectory of disease related neuroanatomical changes which can be useful in disease staging. 4. Understanding the heterogeneities in disease progression rates, brain regions affected and cognitive performance which can aid clinical decision making. The second contribution of this thesis is to introduce new ML algorithms to model disease progression in AD. These algorithms are – 1. scaled Event Based Model (sEBM): A probabilistic generative approach which models disease progression as a sequence of biomarker abnormalities. It extends previous algorithms by solving important computational challenges and makes the resulting algorithm scalable to a much larger number of disease markers. 2. scaled Subtype and Stage Inference (s-SuStaIn): Extends sEBM to infer varying disease progression trajectories. Experiments show that including a larger set of disease features to model progression can be helpful in identifying significantly varying disease progression trajectories. While these algorithms have been applied to AD in this thesis, these are equally applicable to other neurodegenerative conditions such as Parkinson’s Disease, frontotemporal dementia, and progressive supranuclear palsy.

Date

2024-07-18

Resource Type

Text

Resource Subtype

Dissertation

Full item page