Decoding of human emotional states from neural signal

Abstract

Emotions are separated by emotional arousal and valence. The valence represents positive or negative emotional state, and the emotional arousal means that the arousal of strong emotions and emotional behavior. In the present study, we have done two studies to decode the emotional arousal and the valence. In the first study, we investigated whether global EEG synchronization can be a new index to track emotional changes during continuous EEG recording. Global synchronization index (GFS), a measure known to reflect the human cognitive process, was evaluated to quantify the global synchronization of multichannel EEG data recorded from a group of participants (n=20) while they were watching two short video clips. The two video clips were about 5-min long each and designed to respectively evoke happy and fearful emotions. Other participants (n=37) were asked to select two most impressive scenes in each clip, and the questionnaire results were then compared to the grand averaged GFS waveforms. In the study I, beta-band GFS value was decreased when people experienced high emotional arousal regardless of the types of emotional stimuli, and we concluded the GFS could be used to represent for the emotional arousal. The GFS measure might be used as a new index to track emotional changes during video watching and would be potentially used for evaluating movies, TV commercials, and other broadcast products. In the second study, we studied to make new features to represent the valence. We recruited 20 subjects. International Affective Picture System (IAPS), which is a database of pictures used to evoke a three dimensional emotion, such as the valence, emotional arousal, and dominance was used as stimuli. The stimuli, which were selected based on the valence score in a state that was fixed the score of the emotional arousal and dominance, were presented to subjects. Many features, which were calculated in EEG signal, were extracted by using three methods (normalized power spectral density, normalized asymmetry, cross-frequency coupling). To observe the continuous changes of the valence, the features were applied to EEG data which was used in study I, and we calculated the correlation between the tilt of the trend line and the valence score. Finally,we could find two features that the one was a normalized power spectral density on beta-band in the occipital lobe, and the other was a theta-gamma phase-amplitude coupling in the central sulcus.