EEG-DG: A Multi-Source Domain Generalization Framework for Motor Imagery EEG Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motorimagery Electroencephalography (EEG) classification is vital for non-invasive Brain-Computer Interfaces (BCIs).
• EEG signal non-stationarity and individual variability hinder classification model generalization.
• Existing domain adaptation methods require test data during training, which is often impractical.

Purpose of the Study:

• To propose a novel multi-source domain generalization framework (EEG-DG) for robust EEG classification.
• To build generalizable models for unseen target EEG data by leveraging multiple source domains.
• To achieve domain-invariant feature representation and reduce calibration efforts.

Main Methods:

• Developed a multi-source domain generalization framework (EEG-DG).
• Optimized marginal and conditional distributions for joint distribution stability across domains.
• Extended the framework to achieve domain-invariant feature representation.

Main Results:

• EEG-DG demonstrated superior performance on simulative, BCI competition IV (2a, 2b), and OpenBMI datasets.
• Achieved average accuracies of 81.79% (IV-2a) and 87.12% (IV-2b).
• Outperformed state-of-the-art methods in inter-session (78.37%) and inter-subject (76.94%) evaluations on OpenBMI.

Conclusions:

• The proposed EEG-DG framework effectively addresses EEG signal non-stationarity and individual variations.
• EEG-DG offers a practical solution for building generalizable EEG classification models without requiring test data during training.
• The framework shows significant potential for advancing non-invasive Brain-Computer Interface research and applications.