S4Sleep: Elucidating the design space of deep-learning-based sleep stage classification models.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Sleep Medicine

Background:

• Manual sleep stage scoring of polysomnography (PSG) is labor-intensive and time-consuming.
• Existing machine learning algorithms for automatic sleep scoring lack systematic exploration of architectural design choices.
• There is a need for robust and efficient automated methods to analyze sleep patterns.

Purpose of the Study:

• To comprehensively investigate architectural design decisions in encoder-predictor models for automatic sleep stage scoring.
• To develop and evaluate novel machine learning architectures leveraging structured state space models.
• To establish new benchmarks for performance on widely used sleep datasets.

Main Methods:

• Exploration of encoder-predictor architectures for time series and spectrogram representations of PSG data.
• Integration of structured state space models within the proposed architectures.
• Systematic evaluation of S4Sleep(spec) and S4Sleep(ts) models on benchmark datasets without hyperparameter optimization.

Main Results:

• The proposed S4Sleep(spec) and S4Sleep(ts) models consistently outperformed all existing approaches across multiple benchmark datasets.
• Superior performance was demonstrated on the Sleep EDF, Montreal Archive of Sleep Studies, and Sleep Heart Health Study datasets.
• The models achieved high accuracy without requiring additional hyperparameter tuning.

Conclusions:

• The study provides critical architectural insights for developing advanced machine learning models for sleep staging.
• The refined methodology for architecture search offers a valuable tool for future research in time series annotation.
• The developed S4Sleep models represent a significant advancement in automated sleep analysis and have broader applicability to other time series tasks.