Thesis Oral Defense: Assessing reproducibility of Brain-behavior associations using bootstrap aggregation methods

Abstract: In this thesis, amidst growing utilization of resting-state functional connectivity MRI (rs-fcMRI) for linking neural activity to pathological conditions, we confront the prevalent concerns regarding the reliability of such data. Our exploration concentrates on improving the reproducibility of brain-behavior associations within the framework of the Human Connectome Project (HCP) dataset. We employ two distinct bootstrap aggregation approaches to investigate the enhancement of functional connectivity reliability: individual time series bagging using Circular Block Bootstrap (CBB) and subject-level bagging utilizing Linear Support Vector Regression (LSVR) models. Our investigation into individual time series bagging with CBB reveals that this method does not significantly bolster the reproducibility of brain-behavior associations. This finding points to the complexity of achieving reliable functional connectivity measures and the limitations of certain aggregation methods in overcoming this challenge. In contrast, our examination of subject-level bagging through LSVR models presents a more promising outcome. This approach markedly enhances the reliability of model weights between analyses, demonstrating its efficacy in improving data robustness and reproducibility. This differential impact of the two methodologies underscores the critical role of appropriate analytical strategies in enhancing the reliability of neuroimaging data. By delineating the outcomes of these two methodologies, this thesis contributes to the broader discussion on data reliability in the field of neuroimaging. It underscores the necessity for continued methodological innovations and validations across varied datasets to advance the reliability and interpretability of rs-fcMRI studies.

Advisors: Dr. Wheelock and Dr. Lahiri

Committee Members: Soumendra Lahiri, Muriah Wheelock, Robert Lunde