Real-time EEG-based emotion recognition using Riemannian quantification learning

Emotion recognition via electroencephalogram (EEG) has obtained significant attention in many fields, including brain-computer interfaces. It is crucial to accurately identify emotional states in real-time scenarios with few-channel EEG signals and minimal user training data. This study tries to use symmetric positive definite (SPD) matrices to capture the characteristics of emotional EEG signals. We propose a classifier, termed as Riemannian robust soft learning vector quantization with the log-Euclidean metric learning (RRSLVQ-LEML), to distinguish the SPD matrices. To provide thorough analyses, we first evaluated the offline classification performance with channel selection and training data reduction across three public datasets, i.e., SEED, SEED-IV, and SEED-V. The results demonstrated that the proposed model outperformed the SPD network (SPDnet) significantly ( p < 0 . 05 ). The classification accuracies on the three datasets using selected universal channels (22 channels) and a limited number of training samples (150 samples for each category) were 88.33% ± 5.77%, 76.05% ± 7.85%, and 70.62% ± 11.66%, respectively. We further validated the proposed model under the selected channel setup and a small amount of training in the online scenario by recruiting seven participants to conduct experiments for four-category emotion recognition. The average online classification accuracy was 77.28% ± 6.68%, which exhibited a strong linear correlation with the offline results ( r = 0 . 82 ). These findings indicate that the RRSLVQ-LEML model is suitable for real-time emotion recognition with few EEG channels and minimal training samples. Furthermore, the offline performance evaluated by the RRSLVQ-LEML model is transferable to the online scenario. • A novel Riemannian classifier is utilized for EEG-based emotion recognition. • The number of EEG channels and training samples was optimized across three datasets. • The effectiveness of the proposed approach is validated in an online scenario.