Channel Selection Method for EEG Emotion Recognition Using Normalized Mutual Information

Abstract

Electroencephalography (EEG) signals can reflect activities of the human brain and represent different emotional states. However, recognizing emotions based on full-channel EEG signals will lead to redundant data and hardware complexity, thus it is not suitable for designing wearable devices for daily-life emotion recognition. This paper proposes a channel selection method to select an optimal subset of EEG channels by using normalized mutual information (NMI). Compared with other methods, the proposed method solves the problem of obtaining a higher recognition rate while reducing EEG channels sharply. First, EEG signals are sliced into fixed-length pieces with a sliding window, and short-time Fourier transform is adopted to capture EEG spectrogram. Then inter-channel connection matrix is calculated based on NMI, and channel reduction is conducted by using thresholding and connection matrix analysis. The experiments are based on the widely-used emotion recognition database DEAP. It can be derived from the experimental results that the proposed method can select optimal EEG channel subsets to a certain number while maintaining high accuracy of 74.41% for valence and 73.64% for arousal with support vector machines. Further analysis also reveals that the distribution of the selected channels is consistent with cortical areas for general emotion tasks.