EEG-Based Emotion Identification Using 1-D Deep Residual Shrinkage Network With Microstate Features
Junhui Wang, Yu Song, Zemin Mao, Junjie Liu, Qiang Gao
Abstract
Previous studies on emotion identification from electroencephalogram (EEG) mostly focused on normal and depressed people. However, hearing-impaired subjects may require emotional identification due to their chronic lack of perception of auditory information. In this article, we designed an experiment to collect EEG signals from 15 hearing-impaired subjects when they are watching the four kinds of emotional movie clips (happiness, calmness, sadness, and fear). The novel <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${K}$ </tex-math></inline-formula> -means method is used to extract the ten kinds of microstates (as A, B, C, D, E, F, G, H, I, and J) from the raw EEG signal, and then the new EEG single will be retrofitted by those ten microstates. For feature extraction, six kinds of microstate features (global explained variance (GEV), GEV total (GEVT), global field power (GFP), coverage, duration, and occurrence) are calculated. To classify the microstate features, a 1-D deep residual shrinkage network (1-D-DRSN) is utilized, which can filter the emotional irrelevant noise information, and capture emotional representational information. Experimental results show that the proposed model can significantly improve performance compared with other machine learning methods, with an average accuracy of 87.48%. Moreover, we explore different combinations of microstate features to reduce redundant information, and the combination of occurrence, GEV, and coverage reaches 90.15%. From the exploration of each microstate, we find that microstate C has the advantage with an average accuracy of 49.07%.