Channel- and Label-Flip Data Augmentation for Motor Imagery-Based Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• High classification accuracy in motor-imagery brain-computer interfaces (BCIs) necessitates extensive training data.
• Acquiring large datasets from users is often impractical, creating a bottleneck for BCI development.
• Data augmentation (DA) offers a viable solution to expand limited training datasets.

Purpose of the Study:

• To introduce a novel data augmentation technique, channel&label-flip DA, for motor-imagery BCIs.
• To leverage the neuroscientific principle of symmetry between left- and right-hand motor imagery.
• To enhance the performance of BCI classification models with limited data.

Main Methods:

• Proposed a novel channel&label-flip data augmentation method.
• Utilized the OpenBMI dataset with electroencephalograms from 54 participants performing left- and right-hand motor imagery tasks.
• Evaluated performance using three classical machine learning models (filter bank common spatial pattern features) and one deep learning model (raw signal input).

Main Results:

• The proposed channel&label-flip DA method significantly improved average classification accuracy.
• In contrast, a simple channel-flipping DA method without label alteration led to a decrease in classification accuracy.
• The findings demonstrate the effectiveness of incorporating label-flipping alongside channel-flipping for BCI data augmentation.

Conclusions:

• Channel&label-flip DA is an effective strategy for enhancing motor-imagery BCI performance.
• This method addresses the challenge of limited training data by exploiting inherent neural signal properties.
• The study highlights the importance of considering label manipulation in DA for BCIs.