Decoding Functional Brain Data for Emotion Recognition: A Machine Learning Approach

Abstract

The identification of emotions is an open research area and has a potential leading role in the improvement of socio-emotional skills such as empathy, sensitivity, and emotion recognition in humans. The current study aimed at using Event Related Potential (ERP) components (N100, N200, P200, P300, early Late Positive Potential (LPP), middle LPP, and late LPP) of EEG data for the classification of emotional states (positive, negative, neutral). EEG data were collected from 62 healthy individuals over 18 electrodes. An emotional paradigm with pictures from the International Affective Picture System (IAPS) was used to record the EEG data. A linear Support Vector Machine (C = 0.1) was used to classify emotions, and a forward feature selection approach was used to eliminate irrelevant features. The early LPP component, which was the most discriminative among all ERP components, had the highest classification accuracy (70.16%) for identifying negative and neutral stimuli. The classification of negative versus neutral stimuli had the best accuracy (79.84%) when all ERP components were used as a combined feature set, followed by positive versus negative stimuli (75.00%) and positive versus neutral stimuli (68.55%). Overall, the combined ERP component feature sets outperformed single ERP component feature sets for all stimulus pairings in terms of accuracy. These findings are promising for further research and development of EEG-based emotion recognition systems.