Weighted Transfer Learning for Improving Motor Imagery-Based Brain-Computer Interface.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Motor imagery (MI)-based brain-computer interfaces (BCIs) face significant calibration challenges due to session and subject variations in brain signals.
• Extensive training data collection is typically required for each user and session to calibrate BCI system parameters, limiting practical application.

Purpose of the Study:

• To propose a novel transfer learning approach to reduce the calibration time for MI-BCI systems.
• To maintain or improve classification accuracy even with limited subject-specific training data.

Main Methods:

• A transfer learning approach was developed, incorporating a regularization parameter into the classifier's objective function.
• A new similarity measure using Kullback-Leibler (KL) divergence was employed to assess feature space similarity between users.
• The approach was applied to a logistic regression classifier and validated on three datasets using subject-specific Common Spatial Patterns (CSP).

Main Results:

• The proposed weighted transfer learning classifier demonstrated improved classification results compared to subject-specific classifiers, particularly when training data was scarce (p < 0.05).
• Performance gains were more significant for users exhibiting medium and poor initial accuracy.
• Statistical analysis confirmed the proposed method's superior performance over comparable baseline algorithms.

Conclusions:

• The developed transfer learning approach effectively reduces MI-BCI calibration time without compromising classification accuracy.
• This method offers a significant advantage for users with limited training data or lower performance levels.
• The approach shows promise for enhancing the practical usability and accessibility of brain-computer interfaces.