Emotional Valence Detection based on a Novel Wavelet Feature Extraction Strategy using EEG Signals

This paper presents a novel feature extraction strategy in the time-frequency domain using discrete wavelet transform (DWT) for valence level detection using electroencephalography (EEG) signals. Signals from different EEG electrodes are considered independently for the first time in order to find an optimum combination through different levels of wavelet coefficients based on the genetic algorithm (GA). Thus, we take into consideration useful information obtained from different frequency bands of brain activity along the scalp in valence level detection, and we introduce a new set of features named the cross-level wavelet feature group (CLWF). The effectiveness of this approach is strongly supported by the analytical results of experiments in which EEG signals with valence level labels were collected from 50 healthy subjects. High accuracy was achieved for both 2-level (98%) and 3-level valence detection (90%) by applying leave-one-out cross validation using a probabilistic neural network (PNN). In addition, light-weighted sets with less than half EEG recording electrodes are proposed, which can achieve a high accuracy (86% for 3-level valence detection) with offering convenience of users and reducing computational complexity.