A dual-branch network with brain region-constrained attention for EEG emotion recognition.

Introduction: Electroencephalography (EEG)-based emotion recognition provides an objective avenue for affective computing. However, the complexity of EEG signals across temporal, frequency, and spatial domains makes any single dimension inadequate. Methods: To overcome these limitations, we propose the Brain Region-Constrained Attention Dual-Branch Network (BRAD-Net). This network adopts a parallel Spatio-Temporal and Spectral-Spatial dual-branch architecture to achieve synergistic multi-domain EEG feature learning. Within the spatio-temporal branch, we introduce a novel Brain Region-Constrained Attention mechanism, which strictly confines self-attention computation to channels belonging to the same brain region. This design not only suppresses irrelevant cross-region interference but also incorporates neuroanatomical priors of brain parcellation, thereby enabling effective and interpretable representation learning. Results: In subject-dependent experiments using 10-fold cross-validation on DEAP and DREAMER datasets, BRAD-Net achieves high accuracies of 97.44%, 97.70%, and 97.97% for valence, arousal, and dominance on DEAP, and 99.66%, 99.78%, and 99.80% on DREAMER, respectively. Leave-one-subject-out validation on DREAMER dataset achieves accuracies of 72.80% and 75.66% for arousal and dominance, respectively. Additionally, the BRAD-Net demonstrates strong cross-paradigm adaptability, achieving 98.21% accuracy on a depression classification dataset. Conclusions: These findings confirm that integrating neuroanatomical priors into a dual-branch multi-dimensional learning framework effectively extracts robust and interpretable neural representations. BRAD-Net not only advances high-performance EEG emotion recognition but also provides a novel, biologically-constrained design paradigm for developing more interpretable brain-computer interface models. By demonstrating that restricting attention to within-brain-region interactions suffices for accurate emotion recognition, our work offers a new theoretical perspective on the application of brain parcellation knowledge in classification models.