Multivariate Pattern Analysis Based On Functional Connectivity with Machine Learning in Neuroimaging

Abstract

This dissertation discusses machine learning and multivariate pattern analysis (MVPA) in neuroimaging as new sophisticated tools to understand complex cognitive functions in human brains as well as to identify optimal disease-specific biomarkers. MVPA refers to a number of multivariate analytical techniques that are able to exploit patterns of EEG/fMRI signals to extract fine-grained information in the brain. In the first part of this dissertation, we proposed a novel MVPA-based model combined with Bayesian variational learning to decode brain dynamics in visual recognition and speech perception tasks using EEG data. Our multivariate Gaussian hidden Markov model decomposed stochastic multi-subject event-related potentials into distinct quasi-stable EEG microstates. Each of the underlying EEG microstates is expected to represent a unique cortical localization and functional connectivity pattern. Evidence of significant improvements of microstate correlations and improved tendency of functional connectivity distinction over reported methods make this approach a preferred methodology to study brain dynamics and guarantee its use for further clinical applications. In the second part, we discussed how hybrid MVPA techniques help to select the most informative features associated with Alzheimer’s dementia in fMRI data and when combined with appropriate machine learning algorithms, we were able to achieve the maximum classification accuracies on the benchmarks datasets than all the competing methods. We also evaluated novel three-dimensional deep learning methods for the diagnosis of Alzheimer’s disease and jointly predicted the Mini-Mental State Examination (MMSE) scores of patients. Finally, our results suggested that, in the absence of trained clinicians, Alzheimer’s dementia status and clinical MMSE scores could be jointly predicted.