Abstract/Summary

Parallel independent component analysis (ICA) is a framework for analysing concurrent electroencephalogram (EEG) and functional magnetic resonance imaging (fMRI) signals recorded from the brain that involves performing ICA in each modality and then matching the independent components (ICs) across modalities based on their statistical similarities. Together, the matched ICs are understood to provide information about the same neural sources (i.e. functional networks), with the EEG IC providing a high resolution temporal description and the fMRI IC providing a high resolution spatial description. In this thesis, EEG ICA, fMRI ICA, and parallel ICA are evaluated in terms of their accuracy at providing source information using synthetic data generated with The Virtual Brain (Sanz-Leon et al., 2013). Two novel extensions to parallel ICA, which are matching the ICs across modalities using spatial features and mutual information, are also proposed and evaluated. The results of this work indicate that EEG ICA, fMRI ICA, and parallel ICA performances increase with the number of orthogonal sources in the absence of noise, and decrease with the level of noise dispersion when the number of sources are fixed. In the absence of noise, EEG and fMRI ICA performances do not vary largely with source network size (in regions), but in the presence of noise, they vary without clear trends. The incorporation of spatial features improves parallel ICA performance at matching the ICs across modalities, whereas the incorporation of mutual information, in comparison with correlation, deteriorates it. An important observation is that the single-modality and parallel ICAs do not always perform well in best-case conditions. That said, it must be acknowledged that this work is an initial investigation in this direction and further work with more diverse simulation parameters is needed to assess the generalisability of these results. This thesis contributes to the existing body of ICA literature by performing the first evaluation of EEG ICA and fMRI ICA in terms of the number of neural sources and source network size. It is also the first evaluation of the parallel ICA approach that matches ICs across modalities using within-subject temporal dynamics, and the first application of parallel ICA that matches the ICs using spatial features and mutual information. This thesis is also the first demonstration of how The Virtual Brain can be used to evaluate unimodal and multimodal neuroimaging methods.