BAST-Mamba: a band-aware bidirectional spatio-temporal Mamba network for EEG emotion recognition

EEG-based emotion recognition remains challenging due to the need to jointly capture spectral information, spatial dependencies among electrodes, and temporal dynamics across different scales. Existing methods often emphasize only a subset of these characteristics, limiting their ability to comprehensively model the complex spatiotemporal patterns underlying emotional states. To address these challenges, we propose BAST-Mamba, a Band-Aware Bidirectional Spatio-Temporal Mamba framework that integrates multiband EEG representation, multi-scale temporal modeling, electrode-topology information, and bidirectional state-space sequence modeling. Multi-channel and multiband differential entropy (DE) features are first projected into a compact latent representation through an Inverted Embedding layer. A Multi-Scale Temporal Block (MSTB) then employs temporal convolutions with different receptive fields to capture local temporal patterns at multiple scales, while a lightweight topology-guided spatial module incorporates the physical adjacency relationships among EEG electrodes to enhance spatial representation. Building on these representations, a bidirectional Mamba-based Temporal-Spectral Fusion Block (TSFB) models long-range temporal dependencies in both forward and backward directions and integrates the resulting contextual information for emotion classification. Experiments on the public SEED and SEED-IV datasets demonstrate classification accuracies of 95.71% and 90.28%, respectively. These results support the effectiveness of jointly integrating multiband EEG information, multi-scale temporal patterns, electrode-topology priors, and bidirectional state-space modeling for EEG emotion recognition.