An improved self-supervised learning for EEG classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor Imagery Electroencephalography (MI-EEG) classification is crucial for Brain-Computer Interface (BCI) systems.
• Deep learning models for MI-EEG classification typically require substantial labeled data, which is often scarce.
• Insufficient labeled data leads to diminished classification performance in deep learning models.

Purpose of the Study:

• To develop a Self-Supervised Learning (SSL) based method for MI-EEG classification.
• To reduce the dependency on large amounts of labeled training samples in MI-EEG classification.
• To improve the classification performance of BCI systems using limited labeled data.

Main Methods:

• A novel SSL approach involving a pretext task and a downstream classification task was proposed.
• The pretext task involved pre-training a network using original and temporally rearranged MI-EEG data.
• The downstream task utilized a small set of labeled MI-EEG samples to fine-tune the pre-trained network.

Main Results:

• The proposed SSL method demonstrated significant performance improvements compared to baseline methods without SSL, especially when using only one-third of the labeled training samples.
• Experiments on BCI competition datasets (BCI IV 1, 2b, and BCI III IVa) validated the effectiveness of the SSL strategy.
• Consistent performance gains were observed across various percentages of labeled training data, highlighting the method's robustness.

Conclusions:

• The developed SSL strategy effectively enhances MI-EEG classification performance.
• This method offers a viable solution to overcome the challenge of limited labeled data in BCI research.
• The SSL approach is beneficial for improving the overall accuracy and reliability of BCI systems.