Recognition of Emotional States from EEG Signals with Nonlinear Regularity- and Predictability-Based Entropy Metrics

Abstract

Recently, the recognition of emotions with electroencephalographic (EEG) signals has received increasing attention. Furthermore, the nonstationarity of brain has intensified the application of nonlinear methods. Nonetheless, metrics like quadratic sample entropy (QSE), amplitude-aware permutation entropy (AAPE) and permutation min-entropy (PME) have never been applied to discern between more than two emotions. Therefore, this study computes for the first time QSE, AAPE and PME for recognition of four groups of emotions. After preprocessing the EEG recordings, the three entropy metrics were computed. Then, a tenfold classification approach based on a sequential forward selection scheme and a support vector machine classifier was implemented. This procedure was applied in a multi-class scheme including the four groups of study simultaneously, and in a binary-class approach for discerning emotions two by two, regarding their levels of arousal and valence. For both schemes, QSE+AAPE and QSE+PME were combined. In both multi-class and binary-class schemes, the best results were obtained in frontal and parietal brain areas. Furthermore, in most of the cases channels from QSE and AAPE/PME were selected in the classification models, thus highlighting the complementarity between those different types of entropy indices and achieving global accuracy results higher than 90% in multi-class and binary-class schemes. The combination of regularity- and predictability-based entropy indices denoted a high degree of complementarity between those nonlinear methods. Finally, the relevance of frontal and parietal areas for recognition of emotions has revealed the essential role of those brain regions in emotional processes.