Subject-independent emotion recognition with EEG bispectral quadratic phase coupling features and explainable machine learning

Subject-independent emotion recognition from electroencephalography (EEG) is constrained by nonlinear neural dynamics and inter-subject variability. This study characterises nine bispectral quadratic phase coupling (QPC) descriptors extracted from frontal EEG rhythms of the DEAP dataset, selects a compact subset via a genetic algorithm under nested leave-one-subject-out (LOSO) cross-validation, and evaluates classification with an Explainable Boosting Machine (EBM). Under matched preprocessing, the bispectral features outperform power spectral density and differential entropy baselines by 2.7-3.8 points F 1 for arousal and 2.0-2.8 points F 1 for valence (paired Wilcoxon, p < 0.05 after Bonferroni correction), supporting their utility as compact nonlinear descriptors. The pipeline achieves 70.39% accuracy for arousal (19 features) and 69.98% for valence (25 features) on DEAP, statistically comparable to GRU-Conv and BiDCNN under matched LOSO evaluation, while using approximately an order-of-magnitude faster inference and a substantially smaller model footprint. Application of the identical pipeline to the independent DREAMER dataset yields comparable within-dataset performance, while direct DEAP to DREAMER cross-dataset transfer remains challenging. The additive structure of the EBM provides per-feature shape functions and pairwise interactions that are consistent with established frontal asymmetry and cross-frequency coupling literature. Bispectral QPC descriptors under an inherently interpretable modelling framework thus offer a compact, transparent alternative to higher-dimensional black-box approaches for subject-independent EEG emotion classification.