Data-driven Methods for Resting-state fMRI Biomarker Discovery in Mental Illness

Thumbnail Image
Salman, Mustafa S.
Calhoun, Vince D.
Associated Organization(s)
Supplementary to
The objective of this dissertation is to develop data-driven methods for discovering the biological markers of mental illness from functional neuroimaging data. We focused on two broader research areas, the first of which is the characterization of brain networks in functional magnetic resonance imaging (fMRI) and dynamic functional network connectivity (dFNC) states. We posed three specific research questions to address several limitations in the current literature and enhance the characterization process through our contributions. Firstly, there is a need for methods for estimating brain networks to improve their robustness against noise and artifacts in the data. We developed an efficient and automated method, called Autolabeller, which successfully identifies and labels neuroscientifically meaningful intrinsic connectivity networks (ICNs) from artifacts in fMRI data, achieving high accuracy and reproducibility in a time-effective manner. Secondly, the existing methods for identifying brain networks should be adapted to better handle individual differences in the data. We investigated the effectiveness of two methods, spatio-temporal regression (STR) and group information guided ICA (GIG-ICA), for reconstructing subject-specific networks from fMRI data in schizophrenia patients and controls. We showed that adaptive brain networks exhibit greater sensitivity in statistical analysis and classification performance, indicating their potential for identifying image-derived biomarkers of brain dis- orders, which is consistent across independent datasets. Lastly, further investigation is required to address the ambiguities in the identification of brain networks that have been overlooked by existing methods. We evaluated twenty-four cluster validity indexes and identified Davies-Bouldin and Ray-Turi as the most effective methods for determining the optimal number of clusters in dFNC analysis, surpassing commonly used approaches. The other broad research area is the identification of biomarkers of brain disorders using novel approaches and frameworks. We pose three relevant research questions to guide our explorations and contributions. Firstly, we examined the potential of information theory in identifying biomarkers specifically for schizophrenia. We utilized the dynamic functional domain connectivity (DFDC) framework to analyze resting-state fMRI data from schizophrenia patients and controls, revealing increased entropy and reduced cross-domain mutual information in specific DFDC pairs, suggesting higher uncertainty and decreased interdependence in brain function among schizophrenia patients. Additionally, we investigated the use of treatment response as a means of identifying biomarkers specifically for mood disorders. In the context of diagnosing bipolar disorder patients with predominant depressive symptoms, which may lead to misdiagnosis as major depressive disorder (MDD), we introduced a biologically-based classification algorithm utilizing neuroimaging data and the neuromark framework, achieving high accuracy in predicting treatment response and demonstrating the potential for identifying biomarkers associated with medication class response within mood disorders. Lastly, we explored the utility of topological data analysis in identifying biomarkers for various brain disorders. We analyzed the spatial dynamics of fMRI data, revealing lower Betti numbers and higher Wasserstein distance in schizophrenia patients compared to controls, indicating reduced dynamism in resting-state fMRI studies.
Date Issued
Resource Type
Resource Subtype
Rights Statement
Rights URI