SAND: Spectral-Attention Neural Decoding of Hand Kinematics from Low-Frequency EEG Dynamics

Brain-Computer Interface (BCI) technology, integrating neuroscience and artificial intelligence, has been widely applied in neural rehabilitation. However, hand kinematics decoding via electroencephalography (EEG) is constrained by limited precision and cross-subject adaptability. This study proposes the Spectral-Attention Neural Decoder (SAND) - a hybrid framework synergizing spectral decomposition and adaptive deep learning for robust 2D/3D trajectory reconstruction. Systematic analysis of the WAY EEG Grasp-and-Lift dataset revealed that hand movement information is primarily encoded in low-frequency EEG bands. Therefore, a dual-branch continuous decoding architecture was developed: (1) a frequency-domain pathway for noise-resistant spectral embedding, and (2) a temporal-attention pathway utilizing transformer networks to capture dynamic neural modulations. Five-fold cross-validation results demonstrate that SAND achieves state-of-the-art performance in hand-trajectory decoding. The Pearson correlation coefficients for the x, y, and z axes reach 0.9595 ± 0.0148, 0.9534 ± 0.0151, and 0.9293 ± 0.0250, respectively, representing an improvement of 0.07-0.13 over baselines. To assess cross-task generalization, we validate SAND on a self-collected dataset, where it attains average correlation coefficients of 0.90 (x-axis) and 0.96 (y-axis) in 2D trajectory reconstruction. The temporal alignment with ground-truth kinematic recordings was validated by remarkable performance in dynamic time warping analysis. These results confirm SAND as an effective solution for precise hand motion decoding advances non-invasive BCI applications.