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

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) decoding from electroencephalography (EEG) is crucial for brain-computer interfaces (BCIs) but faces challenges like low signal-to-noise ratio and complex dynamics.
• Existing deep learning models (CNNs, Transformers, RNNs) struggle with capturing long-range temporal dependencies, positional coherence, and computational efficiency in MI-EEG decoding.

Purpose of the Study:

• To introduce MIFNet, a hybrid deep learning architecture combining MamBa-based selective state-space models (SSMs) with interactive frequency convolutional neural networks.
• To systematically integrate spectral, spatial, and temporal feature extraction for improved MI-EEG decoding.

Main Methods:

• MIFNet employs non-overlapping frequency decomposition to extract mu and beta rhythms.
• A ConvEncoder module fuses spectral-spatial features across frequency bands.
• A MamBa-based temporal module utilizes selective SSMs for efficient long-range dependency modeling with linear complexity.

Main Results:

• MIFNet achieved superior performance on three public MI-EEG datasets (BCIC-IV-2A, OpenBMI, High Gamma), outperforming existing models.
• Average accuracy improvements were 12.3% over EEGNet, 8.3% over FBCNet, 4.7% over IFNet, and 5.5% over Conformer.
• Ablation studies confirmed the contribution of each component, with the MamBa module alone improving accuracy by 5.5% on BCIC-IV-2A.

Conclusions:

• MIFNet demonstrates significant advancements in MI-EEG decoding accuracy and generalization.
• The hybridization of CNNs and SSMs offers a promising direction for robust real-time BCI applications.
• MIFNet effectively bridges localized feature extraction with global temporal modeling for enhanced EEG signal processing.