CMS2-Net: Semi-Supervised Sleep Staging for Diverse Obstructive Sleep Apnea Severity.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Supervised deep learning for sleep staging faces challenges due to insufficient labeled data.
• Semi-supervised learning (SSL) can mitigate data scarcity but struggles with discriminative feature extraction for varying obstructive sleep apnea (OSA) severity.
• Domain adaptation in SSL models can exacerbate performance degradation across different OSA conditions.

Purpose of the Study:

• To propose a novel co-attention meta sleep staging network (CMS2-net) to address inter-class disparity and intra-class selection problems in sleep staging.
• To enhance feature representation and improve the robustness of SSL models for OSA severity classification.
• To alleviate the domain-shift issue in SSL models for diverse clinical OSA conditions.

Main Methods:

• Development of the co-attention meta sleep staging network (CMS2-net) integrating a co-attention module and a triple-classifier.
• Explicit refinement of feature representations by identifying class boundary inconsistencies.
• Introduction of mutual information with meta contrastive variance for multi-scale gradient supervision.

Main Results:

• The proposed CMS2-net framework demonstrates strong performance on both public and local datasets.
• The model effectively refines feature representations, addressing inter-class disparity and intra-class selection issues.
• The approach achieves state-of-the-art results for semi-supervised learning in sleep staging.

Conclusions:

• CMS2-net offers a robust solution for sleep staging, particularly in the context of limited labeled data and diverse OSA severities.
• The co-attention mechanism and meta contrastive learning effectively improve feature discriminability and model generalization.
• This work advances the application of SSL in sleep disorder detection and highlights potential for clinical translation.