SincMSNet: a Sinc filter convolutional neural network for EEG motor imagery classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) is crucial for brain-computer interfaces (BCIs).
• Decoding MI-EEG signals using convolutional neural networks (CNNs) faces challenges due to individual variability.
• Existing methods struggle to adapt to subject-specific EEG patterns.

Purpose of the Study:

• To propose SincMSNet, a fully end-to-end CNN designed to overcome individual variability in MI-EEG decoding.
• To enhance the accuracy and robustness of MI-BCIs.
• To develop a model capable of extracting subject-specific frequency and temporal features from EEG data.

Main Methods:

• SincMSNet utilizes Sinc filters for subject-specific frequency band extraction.
• Mixed-depth convolutions capture multi-scale temporal information within each frequency band.
• Spatial convolutional blocks and temporal log-variance blocks extract spatial and classification features, respectively.
• Joint supervision with cross-entropy and center loss ensures discriminative feature representations.

Main Results:

• SincMSNet achieved high accuracies on the BCIC-IV-2a (four-class) and OpenBMI (two-class) datasets.
• Average accuracies reached 80.70% (four-class inter-session) and 71.50% (two-class inter-session).
• Single-session accuracies were 84.69% (four-class) and 76.99% (two-class).
• Visualizations confirmed SincMSNet's ability to extract subject-specific frequency bands.

Conclusions:

• SincMSNet demonstrates superior performance in MI-EEG decoding compared to benchmark methods.
• The proposed network shows significant potential for improving the robustness and performance of MI-BCIs.
• The findings suggest SincMSNet is a promising approach for personalized BCI applications.