Personalized Gaussian process-based machine learning models for forecasting Alzheimer's Disease progression

Abstract

In this thesis, I address the problem of predicting behavioral and cognitive metrics from highly heterogeneous datasets (e.g. genetic, clinical/patient history, neuropsychological, biohumoral, molecular) with missing or incomplete data, using Personalized Machine Learning (PML) [71, 72]. In specific, my thesis work focuses on exploring the application of personalized machine learning techniques to the problem of predicting behavioral and cognitive metrics given a pre-organized dataset containing multimodal subject data collected from the longitudinal Alzheimer's Disease Neuroimaging Initiative (ADNI) study. Thus, this thesis explores the impact of PML in the context of predicting the progression of Alzheimer's disease (AD) by predicting various cognitive, clinical, and behavioral metrics known to be indicative of AD diagnosis. To do this, we employ Gaussian Process (GP) Regression as a modeling framework. Using this framework, we design and implement two novel methods for personalized prediction of key cognitive metrics associated with the AD progression (e.g., ADAS-Cog13). Our experimental evaluations show that the proposed personalized model yields significant gains in performance over non-personalized ("one size fits all") approaches applied to the target estimation tasks using the ADNI database. The techniques proposed have the potential to advance and revolutionize disease treatment and clinical research in AD and other health-related domains. We also provide an extensive overview of methods that deal with missing data in ADNI dataset, being one of the main challenges when working with real-world data of AD.