SleepAC: Less Dependency on Manual Annotations, More Reliable Sampling for Automatic Sleep Staging.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Sleep Medicine

Background:

• Accurate sleep staging is crucial for diagnosing sleep disorders but requires extensive, costly manual annotations.
• Current methods for reducing annotation reliance often lack data diversity, limiting model generalization.
• Limited labeled data hinders the development of robust sleep analysis models.

Purpose of the Study:

• To develop a novel model, SleepAC, that minimizes the need for manual sleep scoring annotations.
• To improve the generalization and accuracy of sleep staging models in low-resource settings.
• To reduce the significant costs and time associated with creating large, expertly labeled sleep datasets.

Main Methods:

• SleepAC utilizes an adaptive sample selection strategy, prioritizing informative and diverse data points.
• The model incorporates sleep-specific factors to enhance classification accuracy with fewer labels.
• A contrastive learning framework generates hard negative samples to improve classification of challenging transitional sleep stages.

Main Results:

• SleepAC achieves competitive accuracy and F1-scores across four public datasets.
• The model reaches approximately 95% of fully supervised performance using only 20% of labeled data.
• Demonstrated effectiveness in low-resource scenarios, showing promising generalization across complex sleep dynamics.

Conclusions:

• SleepAC significantly reduces annotation costs and time for sleep staging.
• The model offers a viable solution for developing accurate sleep analysis tools with limited labeled data.
• SleepAC shows potential for widespread application in sleep medicine and research, enhancing diagnostic capabilities.