ARX-based EEG data balancing for error potential BCI.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning algorithms in brain-computer interfaces (BCIs) require large electroencephalographic (EEG) datasets for training.
• EEG datasets are often imbalanced, especially in error-related potential (ErrP) experiments where ErrP epochs are rare.

Purpose of the Study:

• To address the challenge of imbalanced datasets in BCI research.
• To present a novel data balancing method based on AutoRegressive with eXogenous input (ARX) modeling for rare epoch classification.

Main Methods:

• ARX models were identified using EEG data from the 'Monitoring error-related potentials' dataset.
• Synthetic data for the minority class (ErrP epochs) were generated using ARX models.
• A balanced dataset was used to train an EEGNet classifier for non-ErrP vs. ErrP epochs.

Main Results:

• The ARX-based balancing method significantly outperformed classical techniques like class weights (CW).
• Achieved higher accuracy (91.5% vs. 88.3%), F1-score (78.3% vs. 73.7%), and balanced accuracy (87.0% vs. 81.1%).
• Reduced false positive detections (51 vs. 104) and demonstrated better model generalization.

Conclusions:

• The proposed ARX-based data balancing method effectively tackles data imbalance in BCI applications.
• This approach leads to more reliable and robust classification performances for rare events.
• The method holds promise for improving BCI system efficacy and user experience.