This thesis focuses on developing innovative graph theory-based data analysis pipelines using fMRI data to study neuropsychiatric disorders. Three distinct analysis pipelines have been developed using non-parametric statistics, static and dynamic graph theory, and machine learning approaches. The first step involves optimizing fMRI preprocessing pipelines by combining preprocessing steps available in the CONN toolbox and systematically investigating and optimizing parameter settings for field distortion correction, voxel size, and smoothing. The comparisons have demonstrated that the chosen CONN settings yield reliable and reproducible results. Quality checks have been integrated into each pipeline using CONN’s quality analysis tools, and the thesis provides detailed recommendations on preprocessing and quality assurance tailored to different analysis types. The three pipelines developed address: group-level static-network comparisons using non-parametric statistics, machine-learning-based dynamic-network predictors for identifying disease characteristics such as symptom severity, and individual-level static-network comparisons for analyzing longitudinal measurements. Each pipeline has been applied to clinical data. The first pipeline identified group-level static-network differences in Tourette syndrome patients. The second uncovered dynamic-network predictors of tic severity. The third analyzed static-network changes in a single major depressive episode patient after treatment. These applications have revealed the critical role of the basal ganglia-thalamo-cortical network and the amygdala-mediated social decision-making network in Tourette syndrome, and the potential of precision psychiatry approaches in treatment monitoring in major depressive disorder. The findings have been discussed from both methodological and clinical perspectives, alongside challenges and prospects for future research.
Shukti Ramkiran (Wed,) studied this question.