Tensor-CSPNet: A Novel Geometric Deep Learning Framework for Motor Imagery Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning (DL) is prevalent in electroencephalography (EEG)-based brain-computer interfaces (BCIs), particularly for motor imagery (MI) classification.
• Convolutional neural networks (CNNs), effective for visual data, are the mainstream DL approach for MI-EEG classification, despite differing signal characteristics.

Purpose of the Study:

• To investigate alternative deep learning architectures beyond CNNs for MI-EEG classification.
• To propose and validate a novel geometric deep learning (GDL) framework for characterizing EEG signals.

Main Methods:

• Introduced Tensor-CSPNet, a GDL framework utilizing symmetric positive definite (SPD) manifolds.
• Characterized spatial covariance matrices of EEG signals on SPD manifolds.
• Integrated established MI-EEG classification techniques to optimize the GDL framework.

Main Results:

• Tensor-CSPNet achieved state-of-the-art or superior performance on two standard MI-EEG datasets in cross-validation and holdout tests.
• Visualization and interpretability analyses confirmed the framework's effectiveness for MI-EEG classification.
• Demonstrated the potential of DL on SPD manifolds for MI-EEG classification.

Conclusions:

• The study presents Tensor-CSPNet as a viable GDL methodology for MI-EEG classification.
• Generalizing DL on SPD manifolds offers a new direction for BCI research.
• Tensor-CSPNet provides a feasible alternative to CNNs for analyzing complex EEG temporospatial patterns.