MULTIVARIATE MODELING OF COGNITIVE PERFORMANCE AND CATEGORICAL PERCEPTION FROM NEUROIMAGING DATA

State-of-the-art cognitive-neuroscience mainly uses hypothesis-driven statistical testing to characterize and model neural disorders and diseases. While such techniques have proven to be powerful in understanding diseases and disorders, they are inadequate in explaining causal relationships as well as individuality and variations. In this study, we proposed multivariate data-driven approaches for predictive modeling of cognitive events and disorders. We developed network descriptions of both structural and functional connectivities that are critical in multivariate modeling of cognitive performance (i.e., fluency, attention, and working memory) and categorical perceptions (i.e., emotion, speech perception). We also performed dynamic network analysis on brain connectivity measures to determine the role of different functional areas in relation to categorical perceptions and cognitive events. Our empirical studies of structural connectivity were performed using Diffusion Tensor Imaging (DTI). The main objective was to discover the role of structural connectivity in selecting clinically interpretable features that are consistent over a large range of model parameters in classifying cognitive performances in relation to Acute Lymphoblastic Leukemia (ALL). The proposed approach substantially improved accuracy (13% - 26%) over existing models and also selected a relevant, small subset of features that were verified by domain experts. In summary, the proposed approach produced interpretable models with better generalization.Functional connectivity is related to similar patterns of activation in different brain regions regardless of the apparent physical connectedness of the regions. The proposed data-driven approach to the source localized electroencephalogram (EEG) data includes an array of tools such as graph mining, feature selection, and multivariate analysis to determine the functional connectivity in categorical perceptions. We used the network description to correctly classify listeners behavioral responses with an accuracy over 92% on 35 participants. State-of-the-art network description of human brain assumes static connectivities. However, brain networks in relation to perception and cognition are complex and dynamic. Analysis of transient functional networks with spatiotemporal variations to understand cognitive functions remains challenging. One of the critical missing links is the lack of sophisticated methodologies in understanding dynamics neural activity patterns. We proposed a clustering-based complex dynamic network analysis on source localized EEG data to understand the commonality and differences in gender-specific emotion processing. Besides, we also adopted Bayesian nonparametric framework for segmentation neural activity with a finite number of microstates. This approach enabled us to find the default network and transient pattern of the underlying neural mechanism in relation to categorical perception. In summary, multivariate and dynamic network analysis methods developed in this dissertation to analyze structural and functional connectivities will have a far-reaching impact on computational neuroscience to identify meaningful changes in spatiotemporal brain activities