Motor Imagery Classification via Kernel-Based Domain Adaptation on an SPD Manifold

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interface (BCI) calibration is time-consuming and burdensome for users.
• Domain adaptation offers a solution by leveraging existing data, but often requires complex feature extraction from EEG signals.
• Existing methods face challenges with small sample sizes or missing target data labels.

Purpose of the Study:

• To introduce a novel domain adaptation framework, kernel-based Riemannian manifold domain adaptation (KMDA), for BCI.
• To eliminate the need for manual EEG signal feature extraction.
• To improve BCI classification accuracy and reduce calibration time.

Main Methods:

• KMDA analyzes electroencephalogram (EEG) signal covariance matrices within a Riemannian manifold.
• Covariance matrices are aligned and mapped to a high-dimensional space using a log-Euclidean metric Gaussian kernel.
• Subspace learning minimizes conditional distribution distance while preserving target discriminative information. EEG trials are converted to 2D frames (E-frames) for dimensionality reduction.

Main Results:

• KMDA outperformed state-of-the-art domain adaptation methods on three EEG datasets.
• Achieved average Kappa scores of 0.56 (BCI IV IIa) and 0.75 (BCI IV IIIa), and 81.56% accuracy (BCI III IVa).
• Using E-frames further improved overall accuracy by 5.28%.

Conclusions:

• KMDA effectively addresses subject dependence in motor imagery-based BCIs.
• The method significantly reduces calibration time, enhancing user experience.
• KMDA demonstrates a promising approach for practical BCI development.