Neuroinformed spectro-temporal-spatial attention with bi-hemispheric learning for EEG emotion recognition.

Accurate recognition of human emotions from electroencephalogram (EEG) signals is fundamental to affective computing, yet it remains challenging due to the complex spatial-temporal-spectral dynamics of brain activity. In this paper, we propose a neuroinformed hybrid framework for EEG emotion recognition. The framework is centered on Eformer, a neuroinformed hybrid model that integrates Convolutional Neural Networks and Transformers to capture multi-dimensional EEG representations. Eformer constructs a 3D spatio-spectral data tensor by mapping fused Differential Entropy (DE) and Power Spectral Density (PSD) features onto a 2D sensor-level topology. Within the Eformer model, a Multi-band Spatial Alignment module is first employed to calibrate these features using horizontal and vertical directional encodings, ensuring spatial-spectral consistency. To explicitly model the functional lateralization of the brain, Eformer proposes a bi-hemispheric collaborative representation consisting of Hemisphere Spatial Encoder and Hemispheric Cross-Attention modules. Furthermore, a Multi-Scale Temporal Pyramid is integrated into Eformer to capture hierarchical temporal dynamics across varying temporal resolutions. Evaluations on the DEAP and DREAMER datasets show that Eformer achieves competitive performance under the more rigorous Leave-One-Trial-Out protocol. Subject-dependent results are also reported as an upper-bound estimate of within-subject performance. Visualization analysis further suggests that Eformer prioritizes sensor-level regions, such as frontal and temporal electrodes, that are commonly discussed in relation to emotion processing. Our code will be available at the following address:\url{https://anonymous.4open.science/r/Eformer-8B35}.