BGTransform: a neurophysiologically informed EEG data augmentation framework.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning shows promise for electroencephalography (EEG)-based brain-computer interface (BCI) signal decoding.
• Data scarcity and variability limit deep learning model performance in BCIs.
• Existing data augmentation methods may distort signals or lack physiological validity.

Purpose of the Study:

• To introduce a novel data augmentation strategy, BGTransform, for improving EEG-BCI generalization.
• To preserve the neurophysiological structure of EEG signals during augmentation.
• To address data sparsity challenges in training deep learning models for BCIs.

Main Methods:

• Proposed Background EEG Transform (BGTransform), a framework leveraging neurophysiological dissociation between task-related activity and background EEG.
• Generated new trials by perturbing background EEG while preserving task-related signals.
• Applied BGTransform to three public EEG-BCI datasets (SSVEP and P300) and evaluated with various neural decoding models.

Main Results:

• BGTransform consistently outperformed baseline models and conventional augmentation techniques across datasets and architectures.
• Achieved average classification accuracy improvements ranging from 2.45% to 17.15% compared to models without BGTransform.
• Demonstrated enhanced robustness and stable performance across subjects, tasks, and varying recording conditions.

Conclusions:

• BGTransform offers a principled, neurophysiologically informed approach to EEG data augmentation.
• Effectively addresses data sparsity by introducing controlled variability while preserving discriminative features.
• Supports the utility of BGTransform for enhancing accuracy, robustness, and generalizability of deep learning models in neural engineering.