Emotion recognition using electroencephalogram signals through Bicomplex Quaternion-Based processing

Abstract

By the time when our predecessors learned to walk upright, a human evolution process began and the social context get stronger until the Homo sapiens evolved and interacted in social events. This social behaviour affected the biology of the humans' brain until presented a drastic change in the subcortical limbic structure emerging new capacities for the nervous system. Nowadays, human emotions significantly influence individual and social interactions, becoming crucial in medical, security, psychological, psychiatric, and educational environments. In this study, an emotion recognition approach is proposed by using a modify Quaternion Signal Analysis algorithm following the bicomplex quaternion form introduced by Cayley-Dixon. This Bicomplex Quaternion Signal Analysis (bQSA) is developed by taking the electroencephalogram (EEG) information of five different emotion recognition databases. Following a channel selection method to find the top-four effective channels per dataset, the bQSA is constructed in order to propose a novel EEG signal processing method and computing their statistical features to feed several machine learning models and testing the performance in two quaternion produc types: (1) the bicomplex and (2) the quaternion. As results, this method highlights that bicomplex product is slightly accurate than the quaternion form in three out of five tested datasets, achieving the kNN and Tree-based kernels as the top classifiers in eight out of ten cross-validation models. A tree-way Analysis of Variance test suggested that the interaction among product type, machine learning model, and dataset significantly affects classification performance (p < 0.00001). Finally, prior literature typically emphasizes fronto-temporal brain regions as crucial for emotion recognition, this approach identifies a significant relationship among fronto-temporal-parietal regions based on the selected effective channels. Numerical results followed a 10-fold cross-validation to increase the reliability of the Bicomplex Quaternion Signal Analysis and positioning this electroencephalogram signal processing method as one of the top approaches in the state-of-the-art.