Channel selection method for efficient EEG-based emotion analysis

Abstract

Emotion can be defined as a complex reaction pattern that takes into account a person’s behavioral and physiological characteristics. Neural impulses from different regions of the human brain are integrally responsible for generating and processing emotion. Among the existing emotion recognition systems electroencephalogram (EEG) offers the most effectiveness in capturing emotional states in humans. However, recognition of human emotions using EEG signals is quite a challenging task because it requires a balanced tradeoff between retaining the important features of the EEG signals and maintaining computational efficiency at the same time. Working on the DEAP dataset, this paper introduces a novel method of brain source localization (BSL), AgLORETA, as a refinement of the popular BSL method eLORETA. The focus of this designed AgLORETA algorithm is to iteratively refine a core element used in the inverse problem solution calculations of the conventional source localization problems, the lead field matrix. This refinement is done with the view of achieving accurate source localization results allowing us to find the regions of the human brain that generate most neural activity. The utilization of these regions is made while mapping the activities to the electrode channels placed across the scalp in 10-20 systems. The top 10 most active EEG channels have been selected in this approach. To investigate the performance of the selected channels, a combination of two EEG-based feature engineering techniques Hjorth parameters and Permutation Entropy have been applied to the EEG data. The two feature extraction methods have been chosen through various preliminary experimental trials. Finally, the EEG data is validated after feature extraction using some ML classifiers such as SVM, Decision Tree, Random Forest, and KNN. We have compared the AgLORETA algorithm with traditional dSPM, sLORETA, and eLORETA, in a similar approach and achieved promising performance, showing the highest accuracy amongst all. Thus computational complexity is reduced through the 10 selected channels out of the 32 channels of the dataset using AgLORETA.