MtCLSS: Multi-Task Contrastive Learning for Semi-Supervised Pediatric Sleep Staging.

Area of Science:

• Neuroscience
• Machine Learning
• Pediatrics

Background:

• Increasing incidence and recognition of pediatric sleep disorders necessitate automated sleep staging.
• Supervised learning for pediatric sleep staging faces challenges like limited expert availability and data heterogeneity.
• Existing methods struggle with the need for extensive labeled data.

Purpose of the Study:

• To propose a novel multi-task contrastive learning strategy for semi-supervised pediatric sleep stage recognition (MtCLSS).
• To enhance automatic sleep staging accuracy in children, especially with limited labeled data.
• To improve model robustness by leveraging both supervised and self-supervised learning.

Main Methods:

• Developed a multi-task contrastive learning framework (MtCLSS) integrating semi-supervised and self-supervised contrastive learning.
• Applied signal-adapted transformations to electroencephalogram (EEG) recordings for improved representation learning.
• Introduced an extended contrastive loss function adapted for the semi-supervised setting.

Main Results:

• Achieved promising performance on a real-world pediatric sleep dataset: 0.80 accuracy, 0.78 F1-score, and 0.74 kappa.
• Demonstrated the framework's effectiveness on a public dataset, confirming its generality.
• Showcased significant improvements in automatic pediatric sleep staging with very limited labeled data.

Conclusions:

• The MtCLSS framework effectively addresses the limitations of supervised learning in pediatric sleep staging.
• This approach enhances model robustness and accuracy by learning general features from signal transformations.
• MtCLSS offers a viable solution for automatic pediatric sleep staging in data-scarce environments.