Mifnet: a MamBa-based interactive frequency convolutional neural network for motor imagery decoding.

Abstract

Motor imagery (MI) decoding remains a critical challenge in brain-computer interface (BCI) systems due to the low signal-to-noise ratio, non-stationarity, and complex spatiotemporal dynamics of electroencephalography (EEG) signals. Although deep learning architectures have advanced MI-EEG decoding, existing approaches-including convolutional neural networks (CNNs), Transformers, and recurrent neural networks (RNNs)-still face limitations in capturing global temporal dependencies, maintaining positional coherence, and ensuring computational efficiency. To address these challenges, we propose MIFNet, a MamBa-based Interactive Frequency Convolutional Neural Network that systematically integrates spectral, spatial, and temporal feature extraction. Specifically, MIFNet incorporates: non-overlapping frequency decomposition, which selectively extracts motor imagery-related mu (8-12 Hz) and beta (12-32 Hz) rhythms; a ConvEncoder module, which autonomously learns to fuse spectral-spatial features from both frequency bands; and a MamBa-based temporal module, leveraging selective state-space models (SSMs) to efficiently capture long-range dependencies with linear complexity. Extensive experiments on three public MI-EEG datasets (BCIC-IV-2A, OpenBMI, and High Gamma) demonstrate that MIFNet outperforms existing models, achieving an average classification accuracy improvement of 12.3%, 8.3%, 4.7%, and 5.5% over EEGNet, FBCNet, IFNet, and Conformer, respectively. Ablation studies further validate the necessity of each component, with the MamBa module contributing a 5.5% improvement in accuracy on the BCIC-IV-2A dataset. Moreover, MIFNet exhibits strong generalization performance in cross-validation settings, establishing a robust foundation for real-time BCI applications. Our findings highlight the potential of hybridizing CNNs with state-space models (SSMs) for improving EEG decoding performance, effectively bridging the gap between localized feature extraction and global temporal modeling.