Transfer Learning Based on Optimal Transport for Motor Imagery Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography-based brain-computer interfaces (BCIs) face challenges with cross-session variability.
• Frequent recalibration of decoding models is time-consuming and impractical for real-world use.

Purpose of the Study:

• To develop a novel domain adaptation technique to address signal variability in motor imagery BCIs across sessions.
• To eliminate the need for lengthy recalibration procedures before each BCI use.

Main Methods:

• A backward optimal transport-based domain adaptation method was proposed, transporting new session data to a calibration session.
• Evaluated block-wise and sample-wise online adaptation scenarios using a custom dataset and a public dataset.
• Compared several domain adaptation approaches.

Main Results:

• The backward formulation achieved classification performance equivalent to retraining without requiring model updates.
• Sample-wise adaptation reached up to 90.23% overall accuracy when task labels informed the transport.
• Adaptive time was 10 to 80 times faster than other methods.

Conclusions:

• The proposed method effectively mitigates cross-session variability in motor imagery BCIs.
• This retraining-free approach significantly reduces setup time, enabling rapid BCI deployment.
• The method enhances the practicality of BCIs for applications like motor rehabilitation.