Electroencephalogram Emotion Recognition Using Combined Features in Variational Mode Decomposition Domain

Using electroencephalogram (EEG) to recognize human emotion has attracted increasing attention. However, feature extraction from EEG is a challenging work because it is a nonstationary continuous sequential signal. To obtain more pattern information, a combined feature extraction method in the variational mode decomposition (VMD) domain is proposed, which can extract local features of EEG signals to overcome the effects caused by nonstationarity. This method first decomposes EEG into several components using VMD and then combined features of differential entropy (DE) and short-time energy (STE) are extracted from each component. To optimize combined features, the important features are selected by tree modes, and the feature set is dimensionally reduced by further using linear discriminant analysis (LDA). Moreover, an XGBoost classifier with Bayesian optimization is presented to classify different emotional states. Binary-class and multiclass EEG emotion recognition are conducted on the DEAP data set, from which the experimental results show that accuracy of binary-class classification is 81.77% for high/low valence and 80.47% for high/low arousal, and accuracy of 91.41%, 94.27%, 94.27%, and 93.49% are obtained for HVHA, LVHA, LVLA, and HVLA, respectively, which demonstrate its effectiveness.