Single-trial ERP classification of emotional processing

This paper investigates human emotion recognition based on event-related potentials (ERPs) in EEG elicited by picture presentation. Emotion is manipulated through arousal and valence with a calibrated picture dataset. A classification framework is designed for single-trial ERP classification. The most discriminative spatio-temporal features of emotional states were selected and fed to a shrinkage linear discriminant classifier. Various binary classifications were tested according to the emotional valence (positive, negative, neutral) and the arousal level (low, high and no excitation). High classification rate (87%) was obtained for the discrimination between the high-arousal (HA) and low-arousal (LA) negative conditions. Relative good performances were also observed for the (extreme) case “HA negative versus neutral conditions” (66%). Our results suggest that the discrimination of emotional states is better when it is mainly based on an arousal difference between stimuli rather than on a valence difference.