Augmenting Electroencephalogram Transformer for Steady-State Visually Evoked Potential-Based Brain-Computer...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• High-speed steady-state visually evoked potential (SSVEP) brain-computer interface (BCI) systems require advanced electroencephalogram (EEG) decoding.
• Current deep learning methods face challenges with data sparsity and unclear neural underpinnings of augmentation techniques.
• Processing dynamic EEG signals and augmented data necessitates sophisticated models tailored to EEG characteristics.

Purpose of the Study:

• Introduce Background EEG Mixing (BGMix), a novel, neurally-grounded data augmentation technique for EEG.
• Propose the Augment EEG Transformer (AETF), a Transformer-based deep learning model for EEG signal processing.
• Enhance the performance and practicality of high-speed SSVEP-based BCI systems.

Main Methods:

• Developed BGMix to augment training samples by replacing background noise between classes.
• Designed the AETF model to capture temporal, spatial, and frequential EEG features using Transformer architecture.
• Evaluated BGMix and AETF on two public SSVEP datasets.

Main Results:

• BGMix improved classification accuracy across four deep learning models by 4.81%–25.17%.
• AETF outperformed state-of-the-art models, especially with limited training data.
• AETF achieved high information transfer rates (ITRs) of 205.82 ± 15.81 and 240.03 ± 14.91 bits/min.

Conclusions:

• BGMix and AETF represent significant advancements in EEG augmentation and deep learning model design.
• These innovations are informed by neural processes, improving EEG decoding.
• The study enhances the performance and applicability of high-speed SSVEP BCIs.