Brain Networks derived from neuroimaging data have been widely studied recently to analyze the underlying complex spatial relationships between different regions of the brain. Functional connectivity (FC) derived from brain networks have been used as a potential biomarkers for delineating between healthy and mentally ill groups or subgroups with pre-defined features. Although there has been an intense development of statistical methods for computing brain networks, further advances are needed for developing novel and flexible statistical approaches are needed to address several gaps in the literature.
In chapter one, we propose a novel graph-theoretic approach for estimating a population of individual-specific dynamic functional connectivity that is able to systematically borrow information across multiple heterogeneous samples in a data-adaptive manner and guided by supplementary covariate information. In one of the first such approaches in literature, we develop a Bayesian product mixture model that uses covariates to model the mixture weights, which is able to cluster across heterogeneous samples independently at each time scan. An application to a fMRI block task experiment with behavioral interventions in veterans reveals sub-groups of individuals with homogeneous dynamic connectivity patterns and identifies significant dynamic network changes resulting from the intervention.
In chapter two, we proposed a novel semi-parametric Bayesian Support Vector Machine (SVM) approach that incorporates high-dimensional networks as covariates and is able to assign varying levels of shrinkage to the coefficients in an unsupervised manner via a Dirichlet process mixture of double exponential priors. Although SVM-based methods are heavily used in classifying mental disorders, there are few, if any, semi-parametric Bayesian SVM approaches for classification based on high-dimensional brain networks that naturally provides the ability to conduct inferences. We apply the approach to a connectome fingerprinting problem using the Human Connectome Project (HCP) data as well as a second application involving the classification of individuals with attention deficiency hyperactivity disorder (ADHD) and showcase the superior classification accuracy of the proposed approach.
In chapter three, we examine the potential of multimodal dynamic FC, computed by fusing functional magnetic resonance imaging (fMRI) and diffusion tensor imaging data, in terms of predicting continuous clinical measures of disease severity. We develop concrete measures of temporal network variability that are directly linked with disease severity and identify regions whose temporal connectivity fluctuations are significantly related to the disease. Our results illustrate the distinct advantages of prediction of disease severity compared to the usual analysis based on disease phenotype categories, it shows that the multimodal approach is more sensitive to connectivity changes and highlights the predictive prowess of multimodal dynamic FC over existing static and dynamic network models.
Table of Contents
Integrative Learning for Population of Dynamic Networks with Covariates: 11
Non-parametric Bayesian Support Vector Machines for Brain Network-based Classification: 54
Estimating Dynamic Connectivity Correlates Of PTSD Resilience Using MultiModal Imaging: 84
Summary and Future work: 115
About this Dissertation
|Subfield / Discipline|
|Committee Chair / Thesis Advisor|
|Brain Network-Based Statistical Approaches for Neuroimaging Data ()||2022-04-28 09:11:00 -0400||