Vision Transformer and Brain Connectivity Patterns for Estimating Cognitive States

The transformer model is excellent at handling time series signals (such as electroencephalography: EEG) because it can extract information from long-term dependencies effectively. This work combines binarization of EEG connectivity features, cognitive state classification using the vision transformer (ViT), and identifying graphical connectivity patterns for each cognitive state of the mental arithmetic task. The common spatial pattern (CSP) filter coefficient-based channel selection method selects the optimum EEG channels from the input channel set. Then, the Singular Value Decomposition (SVD) method is applied to prepare the binarized connectivity feature matrices, eliminating noisy connections between the optimum channels. The binarized functional-effective connectivity features are passed to the ViT model for cognitive state classification. The ViT model achieves the maximum classification accuracy of 94.86% with the phase-based connectivity feature. The proposed model improves classification accuracy by 6.15% compared to the state-of-the-art studies. This study also suggests a robust brain connectivity network to build a graphical connectivity pattern for each cognitive state. My findings of the EEG-based graphical patterns will bring further understanding of the scalp-level EEG channel patterns among different brain regions for other cognitive tasks.