Towards Sleep Scoring Generalization Through Self-Supervised Meta-Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Biomedical Engineering

Background:

• Automatic sleep scoring models struggle with generalization across different patients and recording facilities.
• This lack of generalization is a major barrier to the clinical and research adoption of automated sleep analysis.

Purpose of the Study:

• To develop a novel meta-learning method for sleep scoring that enhances model generalization without requiring target data adaptation.
• To address the limitations of current sleep scoring models in diverse clinical settings.

Main Methods:

• Introduced S2MAML, a new approach combining Model Agnostic Meta-Learning (MAML) with a self-supervised learning (SSL) stage.
• The SSL stage utilizes a general-purpose pseudo-task to mitigate overfitting to subject-specific patterns.
• Evaluated S2MAML on multiple public sleep scoring datasets (SC, ST, ISRUC, UCD, CAP).

Main Results:

• S2MAML significantly outperformed the standard MAML framework.
• The self-supervised learning component was crucial for the performance gains.
• S2MAML demonstrated superior performance compared to standard supervised learning and MAML across all tested datasets.

Conclusions:

• S2MAML offers a robust solution to the generalization problem in automatic sleep scoring.
• The proposed method shows promise for improving the reliability and applicability of sleep scoring in clinical and research environments.
• This work advances the development of more adaptable and accurate automated sleep analysis tools.