MSCANet: a cross-attention-based multi-scale convolutional fusion neural network for EEG motor imagery classification

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Brain-computer interfaces (BCIs) offer restoration of function for individuals with motor impairments.
• Electroencephalogram (EEG) signals present challenges due to non-stationarity, low signal-to-noise ratio, and inter-subject variability.
• Accurate decoding of EEG signals is crucial for effective BCI control.

Purpose of the Study:

• To develop an advanced deep learning model for decoding complex EEG signals.
• To integrate multi-scale spatiotemporal features for enhanced BCI performance.
• To address the limitations of current EEG signal processing in BCI applications.

Main Methods:

• Proposed a cross-attention-based multi-scale convolutional fusion neural network (MSCANet).
• Employed multi-scale spatio-temporal convolutions and attention mechanisms for feature extraction.
• Utilized temporal convolution with residual connections and cross-attention for dependency capture and feature fusion.

Main Results:

• MSCANet achieved high classification accuracies: 82.06% on BCI IV-2a and 87.45% on BCI IV-2b datasets.
• Demonstrated superior performance compared to several existing BCI decoding models.
• Achieved significant kappa values of 0.76 on both datasets.

Conclusions:

• MSCANet effectively integrates local and global features, capturing temporal dependencies across multiple scales.
• The proposed model shows significant promise for advancing BCI technology.
• This approach offers a robust solution for decoding challenging EEG signals in BCI applications.