Subject Separation Network for Reducing Calibration Time of MI-Based BCI.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery brain-computer interfaces (MI-based BCIs) show promise but require lengthy calibration due to inter-subject variability in EEG signals.
• This calibration process is a significant barrier to wider adoption of MI-based BCIs beyond laboratory settings.

Purpose of the Study:

• To develop a novel domain adaptation framework to minimize the calibration time for MI-based BCIs.
• To adapt labeled data from multiple source subjects to unseen target subjects' trials, improving generalization.

Main Methods:

• A Subject Separation Network (SSN) was trained for each source subject.
• Adversarial domain adaptation was employed to create a shared encoder for similar data representations and subject-specific private feature spaces.
• A shared decoder validated task-relevant information extraction, and an ensemble classifier integrated SSNs.

Main Results:

• The framework demonstrated effective domain adaptation, visualized through transformed features.
• Analysis of the accuracy-calibration cost trade-off showed significant improvements.
• Experimental results on the BCI Competition IV-IIa dataset confirmed the framework's effectiveness against other classification methods.

Conclusions:

• The proposed framework significantly reduces the need for extensive calibration in MI-based BCIs.
• This research highlights the potential of deep learning and domain adaptation for inter-subject EEG decoding.
• Users may soon control MI-based BCIs with minimal calibration, facilitating broader application.