Time and Frequency Domain Feature Selection Using Mutual Information for EEG-based Emotion Recognition

Emotion Recognition using EEG signals remains a challenging task. Usually, feature extraction and channel selection are determined based on neuro-scientific assumptions. Too many features during the EEG-based human emotion recognition will lead to reduced classification accuracy and consume high computational costs. This study analyzes time and frequency domain features such as Mean, Mean Absolute Value, Standard Deviation, and Power Spectral Density. In this study, an EEG Recording session involved 25 subjects consisting of 12 males and 13 females. Video with two emotions, happy and sad, were stimulated to the subjects. The electrodes were placed in channels F7, F8, FP1, and FP2 based on the 10/20 EEG system. The EEG pre-processing, such as signal filtering, Automatic Artifact Removal EOG, Artifact Subspace Reconstruction, and Independent Component Analysis, were done using MATLAB Toolbox, followed by Infinite Impulse Response with Butterworth was applied to separate the EEG signal into alpha, beta, and gamma sub-band. Therefore, 48 numbers of features were extracted to perform emotion recognition. Mutual Information is used for calculating the degree of importance of each feature. Then, the features were ranked to eliminate features with a minimal contribution. We implemented a Random Forest algorithm to classify human emotions based on the EEG signal. The experimental results show that reducing the number of utilized features from 48 to 12 can increase the accuracy score from 82.61 % to 95.65 %.