Nonlinear classification of emotion from EEG signal based on maximized mutual information

Abstract

Recent research reveals that continuous efforts are being made to explore the relationship between EEG signals and manually scored emotions through feature extraction or emotion extraction. In order to achieve emotion extraction, wavelet transforms, Support Vector Machine (SVM), higher-order crossing, short term Fourier transform and ANOVA as classifiers are commonly used. This paper presents for the first time, the determination of maximally informative dimensions from EEG signals, the application of the same for the prediction of human emotions and assessment of the prediction for manually scored emotions. This is an alternative approach of emotion extraction compared to traditional approaches such as Support Vector Machines (SVM) and random forest. The information space, that is available in an EEG database does not usually map into the emotion space entirely. Thus a relevant subspace needs to be developed, which satisfactorily defines the target emotional space. Feature vectors of the dataset are reoriented to the directions, which are relevant informative directions for identifiable emotion. The correlation of manually scored emotion with EEG signal is assessed using mutual Information concerning emotional space. There is no hidden assumption in this method and hence the method is generic in nature. Our method predicts 82% and 72% in ‘two-class emotional scoring’ and ‘three-class emotional scoring’ methods of emotions respectively in a limited dataset of 32 subjects. The maximum prediction recorded is 95.87% for the dominance component of emotion.