SMANet: A Model Combining SincNet, Multi-Branch Spatial-Temporal CNN, and Attention Mechanism for Motor Imagery BCI

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Decoding motor imagery (MI) tasks for brain-computer interfaces (BCIs) is challenging due to individual variability and low EEG signal-to-noise ratio.
• Accurate MI decoding is crucial for developing effective BCI applications.

Purpose of the Study:

• To propose an end-to-end deep learning model, Sinc-multibranch-attention network (SMANet), for enhanced MI-BCI classification.
• To address the challenges of individual discrepancy and low signal-to-noise ratio in EEG signals for BCI applications.

Main Methods:

• SMANet integrates SincNet for band-pass filtering, multibranch spatial-temporal convolutional neural networks (MBSTCNN) for feature learning, and an attention mechanism (ECA) for feature calibration.
• A multi-objective optimization scheme using cross-entropy and central loss enhances discriminative features.
• The model processes spatial-spectral-temporal information from EEG signals.

Main Results:

• SMANet achieved high average accuracies: 80.21% on BCI Competition IV 2a (4-class), 84.02% on BCI Competition IV 2b (2-class), and 72.70% on OpenBMI (2-class).
• These results surpass current state-of-the-art methods in MI-BCI classification.
• The model effectively decodes spatial-spectral-temporal EEG features.

Conclusions:

• SMANet demonstrates superior performance in MI-BCI classification compared to existing methods.
• The proposed deep learning architecture effectively enhances the decoding of EEG signals for BCI applications.
• SMANet offers a promising approach for improving the accuracy and reliability of motor imagery-based BCIs.