Data augmentation for invasive brain-computer interfaces based on stereo-electroencephalography (SEEG).

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning is crucial for brain-computer interfaces (BCIs), but invasive BCIs suffer from limited data.
• Existing data augmentation (DA) methods for brain signals often overlook temporal dependencies, relying on convolutional neural networks.

Purpose of the Study:

• To enhance generative models for BCIs by incorporating temporal relationships from a time-series perspective.
• To introduce a novel conditional transformer-based generative adversarial network (cTGAN) for improving data augmentation in invasive BCIs.

Main Methods:

• A conditional transformer-based generative adversarial network (cTGAN) was developed to capture temporal dependencies in stereo-electroencephalography (SEEG) data.
• The cTGAN was evaluated against noise injection (NI), variational autoencoder (VAE), and conditional Wasserstein generative adversarial network with gradient penalty (cWGANGP) using SEEG data from eight epileptic patients.
• Data quality was assessed through visual inspection, cosine similarity (CS), Jensen-Shannon distance (JSD), and impact on deep learning classifier performance.

Main Results:

• The cTGAN and cWGANGP generated realistic SEEG data, outperforming NI and VAE.
• cTGAN produced superior samples based on CS and JSD metrics.
• cTGAN significantly improved deep learning classifier performance by 6%, compared to cWGANGP's 3.4% improvement.

Conclusions:

• This study is the first to apply DA methods to invasive BCIs using SEEG data.
• The proposed cTGAN demonstrates the advantage of preserving temporal dependencies for enhanced BCI performance.