Electrical and Computer Engineering ETDs

Publication Date

Spring 5-13-2017

Abstract

Mental illnesses are serious disorders of the brain that have devastating effects on individuals and society. In addition to their disabling and impairing effects, mental illnesses have deep social and economical implications, accounting for an estimated loss of 12 billion working days and a care cost surge to $6 trillion a year by 2030. For diseases such as depression and anxiety, enhancing preventive programs and treatment accessibility, in combination with accurate early diagnosis and personalized treatments, are projected to result in a four-fold return on every dollar invested, a strategy that can drastically help curtail those losses. Notably, within the neuroimaging community, blind source separation (BSS) methods have been at the center stage of a push towards guided diagnosis and prediction, playing a key role as tools for disentangling brain "networks". Moreover, current trends show an increased interest in joint analyses of multiple datasets, both multi-subject and multimodal, as a means to enhance individualized assessments, driving efforts toward data sharing and multi-site collaborations. A growth in investments such as the BRAIN initiative and the UK Biobank enhanced imaging study further highlight a requirement for innovation in multidataset image analysis. This work adds a key component to this picture: a new unifying framework for modeling and analysis, combining numerous methods and re-framing the entire BSS field. With new connections among both classical and upcoming methods, highlighting their differences and advantages, and a multidataset mindset, I present a new, general BSS model for joint analysis of multiple datasets called Multidataset Independent Subspace Analysis (MISA). Besides demonstrating MISA's capacity to solve the classical independent component analysis (ICA) problem, we show that it can also solve its independent vector analysis (IVA) and independent subspace analysis (ISA) extensions, plus a novel problem of ISA over multiple datasets. In addition, the present work also includes an extensive review of the BSS literature, a thorough description of good practices on synthetic data generation and model verification, and extensions of joint ICA and IVA to the case of multi-site decentralized data. The results demonstrate MISA's capabilities in challenging noisy scenarios, hybrid simulations, and real multi-subject, multimodal data, highlighting its flexibility and potential to shed new light on the underpinnings of complex mental disorders and improve early diagnosis and treatment outcomes.

Keywords

Multidataset Multidimensional Multimodal Neuroimaging Independent Subspace

Sponsors

NIH grants R01EB006841 & P20GM103472

Document Type

Dissertation

Language

English

Degree Name

Computer Engineering

Level of Degree

Doctoral

Department Name

Electrical and Computer Engineering

First Committee Member (Chair)

Vince D Calhoun

Second Committee Member

Marios S Pattichis

Third Committee Member

Gregory L Heileman

Third Advisor

Sergey M Plis

Fourth Committee Member

Tulay Adali

Available for download on Monday, May 13, 2019

Share

COinS