Augmenting brain-computer interfaces with ART: An artifact removal transformer for reconstructing multichannel EEG...

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Artifact removal in electroencephalography (EEG) is crucial for accurate neuroscientific analysis and brain-computer interface (BCI) performance.
• Existing methods often struggle with transient dynamics and multiple artifact types in multichannel EEG data.

Purpose of the Study:

• To introduce the Artifact Removal Transformer (ART), an end-to-end deep learning model for comprehensive EEG denoising.
• To enhance supervised learning by improving the generation of noisy-clean EEG training data.

Main Methods:

• Utilized a transformer architecture to capture millisecond-scale EEG signal dynamics.
• Employed independent component analysis to generate robust noisy-clean EEG data pairs for training.
• Developed a holistic, end-to-end denoising solution for multichannel EEG data.

Main Results:

• ART demonstrated superior performance compared to other deep learning artifact removal methods across diverse BCI datasets.
• Validated using metrics like mean squared error and signal-to-noise ratio, alongside source localization and component classification.
• Established a new benchmark for EEG artifact removal accuracy and reliability.

Conclusions:

• ART offers a significant advancement in EEG signal processing, improving the reliability of neuroscientific studies.
• The model's effectiveness in handling multiple artifact types facilitates brain dynamics research in naturalistic settings.
• This work is expected to drive further innovations in EEG analysis and BCI development.