PSNSleep: a self-supervised learning method for sleep staging based on Siamese networks with only positive sample...

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Sleep Medicine

Background:

• Supervised learning for sleep staging demands extensive, expert-labeled polysomnography data, which is time-consuming and costly.
• Self-supervised learning (SSL) offers a promising alternative by reducing reliance on labeled data for feature extraction.
• Existing SSL methods often depend heavily on negative sample construction, posing challenges for sleep staging applications.

Purpose of the Study:

• To introduce PSNSleep, a novel self-supervised learning method for automated sleep staging.
• To eliminate the need for negative samples in SSL-based sleep staging by utilizing carefully selected data augmentations.
• To demonstrate the efficacy of PSNSleep on public sleep datasets.

Main Methods:

• Developed PSNSleep, a Siamese network-based self-supervised learning framework tailored for sleep staging.
• Employed a time-shift data augmentation strategy to create positive sample pairs for training.
• Validated the method on the Sleep-EDF and ISRUC-Sleep datasets without using any negative samples.

Main Results:

• PSNSleep achieved an accuracy of 80.0% on the Sleep-EDF dataset.
• The method attained an accuracy of 74.4% on the ISRUC-Sleep dataset.
• Satisfactory performance was obtained without relying on negative sample mining.

Conclusions:

• PSNSleep presents an effective self-supervised approach for sleep staging, reducing the burden of data labeling.
• The proposed time-shift augmentation effectively constructs positive sample pairs, enabling high performance without negative samples.
• The public availability of the source code facilitates further research and application in automated sleep analysis.