A Transfer Learning Algorithm to Reduce Brain-Computer Interface Calibration Time for Long-Term Users.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery-based brain-computer interfaces (BCI) demand extensive initial calibration.
• This prolonged calibration poses a significant challenge for long-term BCI users.
• Reducing calibration time is crucial for practical BCI application.

Purpose of the Study:

• To introduce a novel transfer learning algorithm, r-KLwDSA, to decrease BCI calibration duration.
• To enhance classification accuracy for long-term BCI users by leveraging past data.
• To improve the usability and effectiveness of BCI for stroke rehabilitation.

Main Methods:

• Developed the r-KLwDSA algorithm employing a novel linear alignment technique.
• Aligned historical EEG data with current session data.
• Fused aligned historical and current EEG trials using a weighting mechanism for model calibration.

Main Results:

• Achieved over 4% improvement in classification accuracy compared to session-specific methods with minimal current data.
• Demonstrated substantial accuracy gains (approx. 10%) for sessions with initial accuracy below 60%.
• Validated on a dataset of 11 stroke patients across 18 BCI sessions.

Conclusions:

• The r-KLwDSA algorithm effectively reduces BCI calibration time and improves accuracy.
• This method is particularly beneficial for long-term users and stroke patients with lower initial BCI performance.
• Enables more meaningful BCI rehabilitation by increasing accessibility and performance.