Feature Extraction Approach Based on Statistical Methods and Wavelet Packet Decomposition for Emotion Recognition using EEG Signals

Abstract

Classification of human emotions via EEG signals is a hot topic today. In this study, a method for feature extraction from EEG signals is presented. It is applied for the first time on the GAMEEMO dataset, using a combination of the wavelet packet decomposition (WPD) method with the statistical feature method (SF). The GAMEEMO dataset was classified using the discrete model (2 classes) and the dimensional model (4 classes). This study passed through three sequential main steps: In the first step, four methods were used to extract the features from EEG signals, which are the SF, the combination of the SF with the Welsh power spectral density using (PSD), the combination of the SF with the fast Fourier transform (FFT), and the combination of the SF with the WPD method. In a second step, the Decision Tree Classifier (DT), and the recurrent neural network (RNN) with the long short-term memory (LSTM) algorithm, were applied to classify both emotion models. In the third step, the model performance was evaluated by calculating accuracy, sensitivity, and specificity. The proposed method achieved 98.31% accuracy in the binary-class classification and 99.48% in the multi-class classification. The proposed method may be used in other EEG datasets.