Abstract
Motor imagery-based brain computer interface (MI-BCI) have been increasingly adopted in neurorehabilitation and related fields. The performance of MI-electroencephalogram (MI-EEG) decoding algorithms is central to the advancement of MI-BCI. However, current studies often lack rigorous investigation into the brain's complex network organization. Moreover, most existing methods do not incorporate the cross-frequency coupling (CFC) phenomena that occur during MI into their algorithmic designs, nor do they adequately account for how temporal dynamics across different MI stages influence decoding outcomes. To address these limitations, we propose the Dynamic Spectral-Spatial Interaction Convolution Neural Network (DSSICNN), a parameter-efficient MI-EEG decoding framework that jointly extracts temporal-spectral-spatial features. DSSICNN adopts a dual-branch parallel architecture to concurrently learn spatial representations in both Euclidean and non-Euclidean domains. It further integrates a CFC-inspired attention module to model cross-spectral interactions, followed by an additional attention mechanism that quantifies the contributions of distinct MI stages to decoding performance. DSSICNN achieves decoding performance on two public datasets that surpasses the current state-of-the-art (SOTA) under both session-dependent and session-independent settings. Beyond its empirical advantages, DSSICNN offers design insights for developing Graph Neural Network (GNN)-based MI-EEG decoding algorithms and provides a network neuroscience-inspired perspective for understanding the neurophysiological mechanisms underlying MI.