Multi-Source Discriminant Dynamic Domain Adaptation for Cross-Subject Motor Imagery EEG Recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is crucial for motor imagery (MI) brain-computer interfaces (BCI).
• Deep learning advances BCI, but traditional methods struggle with cross-subject generalization.
• Limited generalization capability across subjects hinders EEG-based BCI performance.

Purpose of the Study:

• To propose a multi-source discriminant dynamic domain adaptation (MSD-DDA) model.
• To enhance motor imagery classification accuracy by leveraging domain adaptation.
• To address global and local disparities in EEG data for improved BCI.

Main Methods:

• Developed a multi-source discriminant dynamic domain adaptation (MSD-DDA) model.
• Dynamically minimized differences between global and local subdomains.
• Introduced batch kernel norm maximization for target domain discriminability and prediction diversity.
• Devised a weighted joint prediction mechanism to adapt source domain contributions based on similarity.

Main Results:

• Achieved high average classification accuracies: 92.43% (BCI Competition Dataset 1), 79.24% (BCI Competition Dataset 2a), and 71.96% (openBMI dataset).
• Demonstrated superior performance compared to classical and recent algorithms.
• Effectively handled global and local disparities in motor imagery classification.

Conclusions:

• The proposed MSD-DDA model significantly improves cross-subject motor imagery classification accuracy in BCIs.
• The model's dynamic adaptation and weighted prediction mechanisms enhance robustness and adaptability.
• This approach offers a promising solution for overcoming generalization challenges in EEG-based BCIs.