Charoenporn Bouyam, Nannaphat Siribunyaphat, Si Thu Aung, Yunyong Punsawad
Although electroencephalography (EEG)-based emotion recognition is a promising approach for affective brain-computer interface (BCI) applications, substantial inter-subject variability continues to limit its generalizability. This study proposed an EEG-based framework to recognize emotions within a valence-arousal model using auditory stimulation. Predefined emotional states were established using validated affective video clips and subsequently evaluated through EEG responses elicited by instrumental melodies. Three EEG features, which include discrete wavelet transform (DWT), functional connectivity (FC), and effective connectivity (EC), together with their combined feature set, were systematically evaluated using five machine learning classifiers. Performance was assessed under subject-dependent, subject-independent (leave-one-subject-out, LOSO), and few-shot subject-adaptation protocols. The results showed that DWT achieved the highest subject-dependent classification accuracy (0.88), followed by the combined feature set (0.84). In contrast, the subject-independent LOSO evaluation yielded near-chance performance across all feature domains (0.23-0.30), highlighting substantial inter-subject variability. Few-shot subject adaptation using 25-75% subject-specific calibration data substantially improved subject-independent performance, with the highest accuracy of 0.77 achieved by FC at 75% calibration. In conclusion, these findings demonstrate the feasibility of EEG-based emotion recognition using auditory stimulation under predefined affective conditions and provide a foundation for the future development of personalized affective brain-computer interface systems.