Supervised and Semisupervised Manifold Embedded Knowledge Transfer in Motor Imagery-Based BCI.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Motor imagery (MI) brain-computer interfaces (BCIs) require extensive calibration, limiting practical use.
• Existing transfer learning methods struggle with high inter-subject variability in electroencephalographic (EEG) signals.
• Unsupervised manifold embedded knowledge transfer (MEKT) provides a framework for addressing these limitations.

Purpose of the Study:

• To develop novel supervised (sMEKT) and semisupervised (ssMEKT) MEKT algorithms for MI-BCI.
• To reduce the need for extensive subject-specific labeled data in MI-BCI systems.
• To improve the adaptability and performance of MI-BCI systems across different subjects.

Main Methods:

• Proposed sMEKT and ssMEKT algorithms based on the MEKT framework.
• Utilized Riemannian alignment (RA) and tangent space mapping (TSM) for domain adaptation.
• Implemented domain adaptation strategies to preserve source discriminability and target domain structure, minimizing distribution shifts.

Main Results:

• Both sMEKT and ssMEKT outperformed six competing algorithms on two public MI datasets.
• ssMEKT demonstrated significant accuracy improvements (5.27% and 2.69%) over the previous best semisupervised method with limited labeled and abundant unlabeled target data.
• The proposed algorithms effectively handle variable sizes of labeled and unlabeled target domains.

Conclusions:

• sMEKT and ssMEKT significantly reduce the requirement for labeled data from target subjects in MI-BCI.
• These algorithms enhance the practical applicability of MI-BCI systems by addressing calibration time and inter-subject variability.
• The developed methods offer a promising solution for more accessible and efficient brain-computer interfaces.