Exploiting frontotemporal asymmetry in low-channel EEG emotion recognition via spectral fusion network

Portable EEG devices hold promise for pervasive affective computing, yet sparse channel counts limit the resolution for capturing global cortical dynamics of emotional processing. To address this, we propose the Spatiotemporal Spectral Asymmetric Fusion Network (STSANet), a neurophysiologically grounded framework designed to decode emotional states from low-density recordings. Unlike traditional methods limited by spatial resolution, we introduce a dual-branch architecture prioritizing Frequency and Spatial Asymmetry Modeling. Instead of merely extracting features, the model explicitly quantifies non-linear hemispheric lateralization-a core mechanism of valence regulation. By modeling differential activation between homologous frontotemporal electrodes (e.g., prefrontal asymmetry), STSANet captures critical lateralized neurodynamics without high-density coverage. A cross-modal attention mechanism then adaptively fuses these lateralized signatures with spectral oscillations. Validating ecological feasibility, we confirm consistent spectral energy distributions between portable Xmuse and professional-grade Enobio devices. Experiments on SEED and a self-collected dataset demonstrate STSANet's superior performance, achieving accuracies of 86.80% and 87.43%, respectively. Crucially, the model maintains high robustness even under rigorous trial-level cross-validation, confirming its generalization capability against temporal distribution shifts. Ablation studies reveal success hinges on explicit spatial asymmetry modeling. Thus, capturing lateralized neurodynamics serves as a resource-efficient strategy for accurate, neurophysiologically interpretable pervasive BCI applications.