SSL-SE-EEG: A Framework for Robust Learning from Unlabeled EEG Data with Self-Supervised Learning and...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Electroencephalography (EEG) is vital for brain-computer interfaces (BCIs) and neurological diagnostics.
• Real-world EEG deployment is hindered by noise, missing data, and high annotation costs.
• Current EEG processing methods often require extensive labeled data and struggle with artifacts.

Purpose of the Study:

• To introduce SSL-SE-EEG, a framework combining Self-Supervised Learning (SSL) and Squeeze-Excitation Networks (SE-Nets).
• To enhance EEG feature extraction, improve noise robustness, and minimize the need for labeled data.
• To enable scalable and low-power processing of EEG signals for advanced applications.

Main Methods:

• EEG signals are transformed into structured 2D image representations for deep learning.
• Integration of Self-Supervised Learning (SSL) for unsupervised feature learning.
• Utilizes Squeeze-Excitation Networks (SE-Nets) for improved feature representation.

Main Results:

• Achieved state-of-the-art accuracy: 91% on MindBigData and 85% on TUH-AB datasets.
• Demonstrated enhanced noise robustness and reduced reliance on labeled data.
• Validated performance across multiple benchmark EEG datasets (MindBigData, TUH-AB, SEED-IV, BCI-IV).

Conclusions:

• SSL-SE-EEG offers a promising solution for biomedical signal analysis and neural engineering.
• The framework is well-suited for real-time brain-computer interface applications.
• Enables low-power, scalable EEG processing, advancing next-generation BCIs.