Multi-View Self-Supervised Learning Enhances Automatic Sleep Staging From EEG Signals.

Area of Science:

• Artificial Intelligence
• Neuroscience
• Biomedical Engineering

Background:

• Manual sleep staging is time-consuming and subjective.
• Deep learning models require extensive labeled data, limiting their practical use.
• Self-supervised learning (SSL) offers a way to learn from unlabeled data.

Purpose of the Study:

• To evaluate the effectiveness of a customized SSL approach for automated sleep staging.
• To develop a multi-view deep learning model for robust sleep staging.
• To reduce the dependency on large labeled datasets in sleep analysis.

Main Methods:

• A multi-view SSL model was trained using temporal and spectral EEG features.
• Cross-view contrastive loss and dynamic weighting were incorporated during pretraining.
• The pretrained model was finetuned on labeled data for sleep staging.

Main Results:

• The SSL-pretrained model achieved competitive accuracy (86.4%, 83.8%, 85.5%) on three public datasets.
• Near-equivalent performance was obtained using only 5% of labeled data compared to full supervised training.
• The method demonstrated enhanced feature transferability and robustness.

Conclusions:

• SSL is a powerful technique for improving automated sleep staging efficiency and reducing data requirements.
• The proposed multi-view SSL framework offers a promising direction for objective sleep analysis.
• This approach has the potential to increase the adoption of automated sleep staging tools.