Constructing large-scale cortical brain networks from scalp EEG with Bayesian nonnegative matrix factorization

Abstract

A large-scale network provides a high hierarchical level for understanding the adaptive adjustment of the human brain during cognition processes. Since high spatial resolution is required, most of the related works are based on functional magnetic resonance imaging (fMRI); however, fMRI lacks the temporal information that is important in investigating the high cognition processes. Although combining electroencephalography (EEG) inverse solution and independent component analysis (ICA), researchers detected large-scale functional subnetworks recently, few researchers focus on the unreasonable negative activation, which is biased from the nonnegative electrical source activations in the brain. In this study, considering the favorable nonnegative property of Bayesian nonnegative matrix factorization (Bayesian NMF) and combining EEG source imaging, we developed a robust approach for EEG large-scale network construction and applied it to two independent real EEG datasets (i.e., decision-making and P300). Eight and nine best-fit networks, including such important subnetworks as the somatosensory-motor network (SMN), the default mode network (DMN), etc., were successfully identified for decision-making and P300, respectively. Compared to the networks acquired with ICA, these networks not only lacked confusing negative activations but also showed clear spatial distributions that are compatible with specific brain function. Based on the constructed large-scale network, we further probed that the self-referential network (SRN), the primary visual network (PVN), and the visual network (VN) demonstrated different interaction patterns with other networks between different responses in decision-making. Our results confirm the possibility of probing the neural mechanisms of high cognition processes at a very high temporal and spatial resolution level.