Classification of Affective States in the Electroencephalogram

Abstract

The goal of the present work is to investigate the feasibility of automatic affect recognition in the electroencephalogram (EEG) in different populations with a focus on feature validation and machine learning in order to augment brain-computer interface systems by the ability to identify and communicate the users’ inner affective state. Two in-depth studies on affect induction and classification are presented. In the first study, an auditory emotion induction paradigm that easily translates to a clinical population is introduced. Significant above chance group classification is achieved using time domain features for unpleasant vs. pleasant conditions. In the second study, data of an emotion induction paradigm for preverbal infants are investigated. Employing the machine learning framework, cross-participant classification of pleasant vs. neutral conditions is significantly above chance with balanced training data. Furthermore, the machine learning framework is applied to the publicly available physiological affect dataset DEAP for comparison of results. Based on spectral frequency features, the framework introduced outperforms results published by the authors of DEAP. The results strengthen the vision of the feasibility of a BCI that is able to identify and communicate the users’ affective state.