A transformer-based network with second-order pooling for motor imagery EEG classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) signals are crucial for decoding motor intentions.
• Motor imagery (MI)-based brain-computer interfaces (BCIs) are gaining traction in neuroinformatics.
• Existing deep learning models for EEG decoding often overlook high-order statistical dependencies.

Purpose of the Study:

• To introduce SecTNet, a novel neural network for EEG decoding that integrates transpose-attention and second-order pooling.
• To address limitations in current deep learning models by capturing complex EEG data structures.
• To improve the accuracy and generalizability of MI-BCI systems.

Main Methods:

• SecTNet utilizes a multi-scale spatial-temporal convolution module for feature extraction.
• A transpose-attention mechanism adaptively models inter-channel dependencies.
• Second-order pooling captures high-order statistical correlations using Riemannian geometry on SPD matrices.

Main Results:

• SecTNet achieved 86.88% accuracy on the BCI competition IV 2a dataset.
• The model reached 74.99% accuracy on the OpenBMI dataset.
• SecTNet demonstrated strong generalization, maintaining performance with 50% less training data.

Conclusions:

• SecTNet offers a robust and generalizable framework for EEG decoding.
• The model effectively enhances MI-BCI performance across diverse applications.
• This approach supports the development of advanced BCI technologies.