RSCNet: A Rhythmic Supervised Contrastive Network for Motor Imagery EEG Decoding

Numerous studies have underscored the crucial role of rhythmic components in decoding motor imagery electroencephalogram (MI-EEG) signals. However, most existing approaches rely on generic convolutional operations to process broadband EEG signals, neglecting critical rhythm-specific spatiotemporal features and cross-rhythm coupling dynamics. To address these limitations, we propose the Rhythm Supervised Contrastive Network (RSCNet) for enhanced MI-EEG decoding. Specifically, we introduce a rhythm encoder inspired by neural oscillatory mechanisms to extract rhythm-specific information, combined with a pointwise convolutional structure to capture rhythm coupling features. Additionally, we propose a novel hybrid-loss classifier that leverages supervised contrastive learning to strengthen rhythm coupling features and utilize cross-entropy loss for classification. Experimental results demonstrate that RSCNet significantly outperforms state-of-the-art models in terms of classification accuracy and kappa coefficient. Notably, RSCNet achieves a 3.86% improvement in classification accuracy over existing methods in multi-class tasks.