Semi-supervised multi-source transfer learning for cross-subject EEG motor imagery classification.

Area of Science:

• Neuroscience and Biomedical Engineering
• Machine Learning for Healthcare

Background:

• Electroencephalogram (EEG) motor imagery (MI) classification is vital for brain-computer interfaces (BCIs).
• Collecting labeled EEG data is time-consuming and labor-intensive, hindering new subject model training.
• Significant inter-subject variability in EEG signals degrades cross-subject classification performance.

Purpose of the Study:

• To develop a model that leverages existing labeled EEG data and unlabeled data from new subjects.
• To improve motor imagery classification accuracy for new subjects in cross-subject scenarios.
• To address the challenge of data scarcity and individual differences in EEG-based BCIs.

Main Methods:

• Proposed a semi-supervised multi-source transfer (SSMT) learning model.
• Focused on learning informative and domain-invariant representations for cross-subject MI-EEG classification.
• Implemented a dynamic transferred weighting schema to integrate multi-source domain features for final predictions.

Main Results:

• Achieved average accuracies of 83.57% and 85.09% on two public EEG datasets.
• Demonstrated the effectiveness of the SSMT approach in cross-subject MI classification.
• Validated the importance of domain-invariant representations in maximizing data utility.

Conclusions:

• The SSMT model successfully enhances motor imagery classification for new subjects.
• The study highlights the significance of domain-invariant representations for robust EEG-based BCIs.
• The proposed method offers a viable solution for data-scarce and personalized BCI applications.