Dual-Branch Self-Supervised Contrastive Pre-Training Framework for Sleep Stage Classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Accurate sleep staging is crucial for diagnosing sleep disorders and assessing sleep quality.
• Current automated methods require extensive expert-labeled datasets, which are costly and subjective.
• This limits the development and accessibility of advanced sleep analysis tools.

Purpose of the Study:

• To develop a few-shot learning framework for automated sleep staging using single-channel electroencephalogram (EEG) data.
• To reduce the dependency on large, expert-annotated datasets for sleep staging model training.
• To enable effective sleep classification with minimal labeled data.

Main Methods:

• A dual-branch contrastive pre-training framework was proposed for few-shot, single-channel EEG-based sleep staging.
• The framework employed fully self-supervised pre-training on unlabeled data.
• Subsequent fine-tuning required only a small subset of labeled samples.

Main Results:

• The framework achieved state-of-the-art results on the Sleep-EDF-v2 dataset using only 1% labeled data (76.10% accuracy, 61.34% Macro F1-score).
• Performance was comparable to supervised models trained on 100% labeled data.
• Robust and consistent results were observed on the ISRUC-1 and ISRUC-3 datasets.

Conclusions:

• The proposed few-shot learning framework effectively enables sleep classification with minimal labeled data.
• This approach addresses the bottleneck of expert annotation in automated sleep staging.
• The framework holds significant potential for diverse clinical applications and broader adoption of sleep analysis tools.