Adaptive frequency band attention-guided CNN-BiLSTM for Spatial-Spectral-Temporal EEG emotion recognition.

Electroencephalography (EEG)-based emotion recognition has gained increasing attention in affective computing because EEG provides high temporal resolution and reflects intrinsic neural activity. However, reliable emotion recognition remains challenging due to the nonlinear, non-stationery and subject-dependent nature of EEG signals. Moreover, many existing methods insufficiently examine temporal overlap effects, treat frequency bands with equal importance, and offer limited spectral interpretability. To address these issues, this study proposes an interpretable Adaptive Frequency Band Attention (AFBA)-guided multi-band 2D CNN-BiLSTM framework for spatial-spectral-temporal EEG emotion recognition using power spectral density (PSD) features. The proposed framework decomposes EEG signals into delta, theta, alpha, beta, gamma and broadband components using Butterworth bandpass filtering. Welch-based PSD features are then extracted and arranged into structured multi-band representations. The AFBA module adaptively learns the relative contribution of each frequency band before spatial-spectral feature extraction through a 2D convolutional neural network and temporal dependency modeling using a BiLSTM network. To assess temporal continuity, 3-s EEG windows with 0 s, 1 s and 2 s overlap are evaluated. The model is tested on the DEAP dataset using leakage-aware trial-level tenfold cross-validation and further validated on SEED under protocol-aligned conditions. Results show that increased temporal overlap improves performance across bands and emotion dimensions. With 2-s overlap, the gamma band achieves 97.2% valence accuracy, 98.1% arousal accuracy and 88.9% four-class accuracy on DEAP, while reaching 94.1% three-class accuracy on SEED. Ablation and SHAP analyses further support the contribution of temporal modeling, AFBA, and high-frequency beta/gamma activity.