Toward in-silico data assessment for passive BCIs: generating EEG rhythms with GANs.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Passive brain-computer interfaces (BCIs) using electroencephalography (EEG) are crucial for monitoring vigilance in demanding situations.
• Limited public EEG datasets hinder the advancement of AI research in this field.
• Generative adversarial networks (GANs) offer a potential solution for data augmentation.

Purpose of the Study:

• To augment electroencephalography (EEG) datasets using generative adversarial networks (GANs).
• To establish a quality-assessment pipeline for synthetic EEG data.
• To evaluate the impact of GAN-based data augmentation on AI model performance for vigilance monitoring.

Main Methods:

• Augmented two EEG datasets (resting-state and repetitive task simulations) using GANs.
• Extracted time-variant rhythms via continuous wavelet transform and compared real vs. synthetic data using L2 distance and cross-correlation.
• Trained forecasting models on original and augmented data, evaluating performance using Mean Absolute Error (MAE) and symmetric mean absolute percentage error (SMAPE).
• Assessed model embeddings using a Fréchet Inception Distance (FID)-inspired metric.

Main Results:

• GAN-augmented data significantly improved signal prediction, reducing MAE by up to 46.4% for the resting-state dataset.
• Forecasting models trained on augmented data showed substantial MAE reductions (15.4%-21.2%) for larger data splits.
• Linear interpolation yielded worse performance and indicated overspecialization, unlike GAN-generated data.

Conclusions:

• GAN-based data augmentation is effective for enhancing EEG datasets for BCI applications.
• The proposed quality-assessment pipeline provides a robust method for evaluating synthetic EEG data.
• This framework addresses critical data limitations, paving the way for scalable AI in BCI research.