Denoising EEG Signals for Real-World BCI Applications Using GANs

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Electroencephalography (EEG) is crucial for brain-computer interfaces (BCIs), translating brain activity into actions.
• Real-world BCI performance is significantly hampered by noise and artefacts in EEG signals, limiting practical use.
• Existing denoising methods struggle with the complexity and variability of artefacts in portable EEG devices.

Purpose of the Study:

• To develop and evaluate a Generative Adversarial Network (GAN) pipeline for robust EEG artefact removal.
• To enhance the performance and applicability of portable BCIs in real-world environments.
• To demonstrate the GAN's capability in handling multiple artefact types and its potential for unsupervised learning tasks.

Main Methods:

• Utilized Generative Adversarial Networks (GANs) for supervised EEG signal denoising, mapping noisy to clean EEG signals.
• Developed a pipeline leveraging GANs to address artefact removal in EEG time series data.
• Validated the model on a toy dataset with mains noise and a benchmark dataset with induced myogenic and ocular artefacts.

Main Results:

• The GAN framework successfully denoised EEG signals, significantly improving signal-to-noise ratio (SNR) and power spectral density (PSD).
• The model demonstrated competitive performance compared to current state-of-the-art deep learning denoising techniques.
• Qualitative and quantitative evidence confirmed the GAN's effectiveness in removing multiple artefact types.

Conclusions:

• This GAN-based framework is the first to achieve generalizable EEG artefact removal across multiple artefact types.
• The proposed method advances deep learning techniques for EEG denoising, offering a robust solution for real-world BCI applications.
• Enables the development of portable EEG devices with cleaner, more stable brain signals, integrating AI with wearable technology.