Dynamic time warping-based transfer learning for improving common spatial patterns in brain-computer interface

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Common Spatial Patterns (CSP) is crucial for motor imagery (MI)-based brain-computer interfaces (BCIs).
• CSP's performance degrades with limited training data due to sample-based covariance estimation.
• Addressing data scarcity is vital for robust BCI system development.

Purpose of the Study:

• To propose a novel regularized covariance matrix estimation framework for CSP, named DTW-RCSP.
• To enhance CSP performance in MI-BCIs, especially when training data is scarce.
• To leverage dynamic time warping (DTW) and transfer learning for improved feature extraction.

Main Methods:

• DTW-RCSP combines subject-specific and DTW-aligned transferred covariance matrices.
• Dynamic Time Warping (DTW) aligns previous subjects' trials to reduce temporal variations.
• An online method selects regularization parameters based on classifier confidence scores.

Main Results:

• DTW-RCSP significantly outperformed baseline algorithms across various testing scenarios.
• Performance improvements were most notable with limited training trials.
• Successful BCI interactions were achieved with minimal calibration (one trial per class).

Conclusions:

• The proposed DTW-RCSP framework effectively addresses the limitations of traditional CSP with small datasets.
• This method enhances the practicality and accessibility of MI-BCI systems.
• The findings suggest a significant advancement in BCI technology, particularly for users with limited training capacity.