Data Augmentation: Using Channel-Level Recombination to Improve Classification Performance for Motor Imagery EEG.

Area of Science:

• Neuroscience and Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Deep learning models for brain-computer interfaces (BCI) heavily rely on high-quality and quantity of training data.
• Acquiring extensive electroencephalography (EEG) data through prolonged calibration sessions is impractical and costly.
• Data augmentation (DA) offers a time- and cost-efficient solution for generating artificial EEG data.

Purpose of the Study:

• To propose and evaluate a novel data augmentation (DA) method for motor imagery (MI) EEG signals.
• To introduce the brain-area-recombination (BAR) technique for generating artificial MI-EEG data.
• To assess the effectiveness of BAR across single- and multi-subject scenarios using different schemas.

Main Methods:

• The proposed brain-area-recombination (BAR) method splits EEG samples into halves based on brain channels and recombines them to create artificial samples.
• Two schemas, intra-subject and adaptive-subject, were designed for single- and multi-subject scenarios, respectively.
• Experiments utilized the EEGnet classifier on two public datasets, comparing BAR against noise-added and flipping DA methods, and CSP-SVM.

Main Results:

• The BAR method significantly improved EEGnet's classification performance in both intra- and adaptive-subject schemas (p < 0.01).
• BAR outperformed common DA methods like noise-added and flipping (p < 0.05).
• EEGnet trained with BAR augmentation showed an 8.3% improvement over the CSP-SVM algorithm (p < 0.01).

Conclusions:

• Brain-area-recombination (BAR) is an effective data augmentation technique that significantly enhances deep learning classification of motor imagery EEG signals.
• BAR offers a practical solution to data scarcity in BCI research, improving model performance and potentially accelerating deep learning advancements in the BCI field.