Self-Supervised Learning for Label- Efficient Sleep Stage Classification: A Comprehensive Evaluation.

Area of Science:

• Neuroscience
• Machine Learning
• Computational Biology

Background:

• Deep learning models for electroencephalogram (EEG)-based sleep stage classification (SSC) require extensive labeled data, limiting practical application.
• Labeling sleep data is costly and time-consuming, creating a bottleneck for model development.

Purpose of the Study:

• To evaluate the effectiveness of self-supervised learning (SSL) in improving the performance of EEG-based SSC models in low-data scenarios.
• To assess the impact of SSL pretraining on model robustness against data imbalance and domain shifts.

Main Methods:

• Conducted a thorough study using three distinct EEG-based sleep stage classification datasets.
• Employed self-supervised learning (SSL) for pretraining models.
• Fine-tuned pretrained models using a small fraction (5%) of labeled data.

Main Results:

• Fine-tuning SSL-pretrained models with only 5% of labeled data achieved performance comparable to models trained with full supervision.
• Self-supervised pretraining enhanced the robustness of SSC models against data imbalance.
• SSL pretraining improved model generalization capabilities when facing domain shifts.

Conclusions:

• Self-supervised learning is a highly effective strategy to overcome data scarcity in EEG-based sleep stage classification.
• SSL enables the development of high-performing SSC models even with limited labeled data, making them more applicable in real-world settings.
• SSL pretraining contributes to more robust and generalizable sleep stage classification models.