Self-Supervised Electroencephalogram Representation Learning for Automatic Sleep Staging: Model Development and...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning models excel in sleep medicine by analyzing electroencephalogram (EEG) data.
• A significant challenge remains in effectively utilizing vast amounts of raw EEG data for model training.

Purpose of the Study:

• To develop robust vector representations from massive unlabeled EEG signals.
• To ensure these learned features are suitable for sleep staging and outperform supervised models with limited or noisy data.

Main Methods:

• Introduction of a self-supervised model named Contrast with the World Representation (ContraWR) for EEG signal representation learning.
• ContraWR utilizes global data statistics, unlike prior methods relying on negative samples, to differentiate sleep stages.
• Model validation on three diverse, real-world EEG datasets (at-home and in-laboratory recordings).

Main Results:

• ContraWR demonstrated superior performance compared to four other self-supervised learning methods in sleep staging across three large EEG datasets.
• The model outperformed supervised learning, particularly in low-label scenarios, showing a 4% accuracy improvement with less than 2% labeled data on the Sleep EDF dataset.
• Generated 2D projections revealed informative and representative feature structures from the learned representations.

Conclusions:

• ContraWR exhibits robustness to noise, yielding high-quality EEG representations for subsequent prediction tasks.
• The model's applicability extends to other unsupervised physiological signal learning tasks.
• Future research will focus on task-specific data augmentations and hybrid self-supervised/supervised approaches.