Baseline-Guided Representation Learning for Noise-Robust EEG Signal Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) are limited by noisy electroencephalography (EEG) signals.
• Current denoising methods can remove important task-related information.

Purpose of the Study:

• To develop a novel deep learning framework, BGNet, for dynamic noise mitigation and feature extraction in motor imagery (MI) EEG classification.
• To leverage underutilized baseline EEG signals to enhance BCI accuracy.

Main Methods:

• BGNet utilizes data augmentation, an autoencoder for feature extraction from baseline and MI signals, and a feature alignment module.
• The framework separates task-specific information from noise for improved classification.

Main Results:

• Achieved state-of-the-art performance with a 5.9% and 3.7% improvement on BCIC IV 2a and 2b datasets, respectively.
• Learned features demonstrated superior representational power compared to baseline models in noisy EEG conditions.

Conclusions:

• Baseline EEG signals can be effectively utilized to enhance BCI performance.
• BGNet offers a promising approach for robust feature extraction and noise mitigation, potentially simplifying BCI systems for brain-based communication.