Investigating the Effect of EEG Channel Selection on Inter-subject Emotion Classification

Inter-subject emotion classification from physiological brain signal (electroencephalogram (EEG)) is often challenging due to lack of generalizability in computed features across the subjects. A great deal of research effort has been devoted to EEG-based emotion recognition. However, the effect of channel selection and relative efficacy of different domains of EEG features on multilevel emotion classification in the inter-subject scenario is still unclear. This work aims to investigate the effect of channel selection/reduction on inter-subject multilevel emotion classification. The analysis is performed on a publicly available DEAP dataset and four groups of channels are selected from the literature. Furthermore, we compute 6 number of time and frequency domain EEG features, and 3 number of non-linear EEG features and study the relative efficacy of these features towards capturing the generalizable component in EEG signals to classify emotions in inter-subject scenarios. The results indicate that the emotions are better classifiable after 20 seconds time from the beginning of the stimulus. Also, the inter-subject emotion classification accuracy increases significantly with increasing the number of channels beyond 10. Furthermore, between the time and frequency domain features, and nonlinear EEG features, the prior shows better efficacy in classifying multilevel emotions in the inter-subject domain.