Comprehensive Analysis of Feature Extraction Methods for Emotion Recognition on Motor Imagery from Multichannel EEG Recordings

Electroencephalogram (EEG) related analyses have a vast extent of enquiries, preliminary from Brain-Computer Interfaces (BCI) on motor imagery tasks, neurotechnology, human-computer interaction, clinical and medical applications, sleep research, as well as the neural basis of emotional recognition and behavioural responses. This study purposely investigates the application of feature extraction methods from the emotion recognition arena against the motor imagery arena, using the electroencephalogram (EEG) data from the BCI Competition IV. As the same methods are used in both fields, binary classification has been embraced, emotion recognition’s valence and arousal, using its equivalent motor imageries of left and right classes. Essentially, six EEG feature extraction methods were used in the classification accuracy process, including statistical features, wavelet analysis, higher-order spectra (HOS), Hjorth, fractal dimension (Katz, Higuchi, Petrosian) and three-dimensional combination of fractal, wavelet and Hjorth. Further, the classifier performance methods were the Gaussian radial basis function RBF SVM (GSVM) and the regression tree (CART). Remarkably, the statistical method has better accuracy than the other feature sets, precisely the fractal dimension, which was the highest considering the emotion recognition span. Also, GSVM generally has better accuracy on the BCI IV dataset than CART. By contemplating the novel outcome notices, we can deduce that in the methods used in emotion recognition when applied to motor imagery, unique results are obtained, feature-wise and classifier-wise.