Domain generalized feature embedded learning for calibration-free event-related potentials recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Event-related potentials (ERPs) are crucial for EEG-based Brain Computer Interfaces (BCIs).
• Subject-specific variations in ERP spatio-temporal characteristics pose a significant challenge for calibration-free BCI development.
• Developing BCIs that do not require individual subject calibration is a key research goal.

Purpose of the Study:

• To propose a novel Domain Generalized Feature Embedded Learning (DGFEL) method for calibration-free ERP recognition.
• To address the issue of data distribution across subjects in EEG-based BCIs.
• To enable robust ERP classification without subject-specific training data.

Main Methods:

• ERP alignment using covariance centroids.
• Enhancement of aligned samples via xDAWN filtering for spatio-temporal feature extraction.
• Generalization of features using decomposed adversarial loss and a neural network embedding backbone.

Main Results:

• The DGFEL method demonstrated superior classification performance compared to state-of-the-art methods and deep learning models on benchmark datasets.
• The method successfully extracted robust features from source subjects.
• Features were effectively generalized to new subjects without needing target ERP samples.

Conclusions:

• The proposed DGFEL method offers a novel solution for constructing calibration-free ERP-BCIs.
• This approach effectively overcomes the challenge of subject variability in ERP data.
• The DGFEL method facilitates robust and generalizable ERP recognition for BCI applications.