Accurate Emotion Recognition Utilizing Extracted EEG Sources as Graph Neural Network Nodes

Automated analysis and recognition of human emotion play an important role in the development of a human-computer interface. High temporal resolution of EEG signals enables us to noninvasively study the emotional brain activities. However, one major obstacle in this procedure is extracting the essential information in presence of the low spatial resolution of EEG recordings. The pattern of each emotion is clearly defined by mapping from scalp sensors to brain sources using the standardized low-resolution electromagnetic tomography (sLORETA) method. A graph neural network (GNN) is then used for EEG-based emotion recognition in which sLORETA sources are considered as the nodes of the underlying graph. In the proposed method, the inter-source relations in EEG source signals are encoded in the adjacency matrix of GNN. Finally, the labels of the unseen emotions are recognized using a GNN classifier. The experiments on the recorded EEG dataset by inducing excitement through music (recorded in brain-computer interface research lab, University of Tabriz) indicate that the brain source activity modeling by ESB-G3N significantly improves the accuracy of emotion recognition. Experimental results show classification accuracy of 98.35% for two-class classification of positive and negative emotions. In this paper, we concentrate on extracting active emotional cortical sources using EEG source imaging (ESI) techniques. Auditory stimuli are used to rapidly and efficiently induce emotions in participants (visual stimuli in terms of video/image are either slow or inefficient in inducing emotions). We propose the use of active EEG sources as graph nodes by EEG source-based GNN node (ESB-G3N) algorithm. The results show an absolute improvement of 1-2% over subject-dependent and subject-independent scenarions compared to the existing approaches.

Automated summary

Automated analysis and recognition of human emotion is important for human-computer interface development. EEG signals with high temporal resolution enable the study of emotional brain activities, but the low spatial resolution in EEG recordings is a major obstacle. This paper proposes a method that uses graph neural networks to recognize emotions based on EEG data.