An Adaptive Spatial Filter for User-Independent Single Trial Detection of Event-Related Potentials.

Area of Science:

• Neuroscience
• Signal Processing
• Biomedical Engineering

Background:

• Current brain-computer interfaces (BCIs) often rely on supervised learning methods.
• These methods require extensive subject-specific training data for calibration, which is time-consuming.

Purpose of the Study:

• To develop a novel dimensionality reduction algorithm for event-related potential (ERP) detection.
• To create a spatial filter adaptable to new subjects, minimizing or eliminating calibration time.

Main Methods:

• The algorithm utilizes generalized eigendecomposition, building upon the xDAWN filter.
• Recursive least squares (RLS) updates enable incremental training of filter coefficients.
• The spatial filter's effectiveness is analyzed in various transfer learning scenarios with adaptive classifiers.

Main Results:

• The developed spatial filter effectively compensates for user-switching variations.
• Previously recorded training data from different subjects can be reused.
• The system demonstrates adaptability to new users.

Conclusions:

• The novel approach significantly reduces or eliminates the BCI calibration phase.
• Instantaneous BCI system use is possible with only a minor performance reduction.
• The BCI system can be made user-independent by adapting precomputed spatial filters.