Differential entropy feature for EEG-based emotion classification

Abstract

EEG-based emotion recognition has been studied for a long time. In this paper, a new effective EEG feature named differential entropy is proposed to represent the characteristics associated with emotional states. Differential entropy (DE) and its combination on symmetrical electrodes (Differential asymmetry, DASM; and rational asymmetry, RASM) are compared with traditional frequency domain feature (energy spectrum, ES). The average classification accuracies using features DE, DASM, RASM, and ES on EEG data collected in our experiment are 84.22%, 80.96%, 83.28%, and 76.56%, respectively. This result indicates that DE is more suited for emotion recognition than traditional feature, ES. It is also confirmed that EEG signals on frequency band Gamma relates to emotional states more closely than other frequency bands. Feature smoothing method- linear dynamical system (LDS), and feature selection algorithm- minimal-redundancy-maximal-relevance (MRMR) algorithm also help to increase the accuracies and efficiencies of EEG-based emotion classifiers.