Brain Multi-Region Information Fusion using Attentional Transformer for EEG Based Affective Computing

In this paper, we worked on the fusion of multiple brain regions in order to combine information from different brain regions. The idea is that considering the dynamic processing of emotional video stimulus will involve different brain regions, and hence, fusion of information from these brain regions can increase emotion recognition accuracy significantly. We utilized EEG datasets from Indian (DENS) and European (DEAP) populations, comprising 128 and 32 channels, respectively. We categorized EEG channels into different brain regions and divided them into five sub-regions. Then, we used five transformer models to extract information from these sub-regions and concatenated the results to obtain a fused information vector for classification. The idea resembles the theoretical findings that brain association areas that fuse information from different brain lobes contribute significantly to emotion processing. For the DENS dataset, we achieved 97.78% accuracy in valence and 95.74% accuracy in arousal. For the DEAP dataset, we achieved 90.07% accuracy in valence and 84.52% accuracy in arousal. We propose that deep learning models simulating information fusion in the association regions of the brain can enhance emotion recognition accuracy.