Automated sleep stage and event detection algorithms using quality-controlled polysomnography annotations.

Area of Science:

• Sleep Medicine
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Automated sleep analysis using polysomnography (PSG) is essential for diagnosing sleep disorders.
• Machine learning (ML) models offer potential for objective and efficient sleep scoring.
• Evaluating ML model performance against expert human scorers is critical for clinical translation.

Purpose of the Study:

• To develop and evaluate ML models for sleep stage classification, arousal detection, and respiratory event detection from overnight PSG recordings.
• To compare the performance of these ML models against certified human scorers and assess inter-scorer agreement.

Main Methods:

• Overnight PSG data from healthy and sleep-disordered breathing participants were collected.
• Four certified scorers provided reference annotations for sleep stages, arousals, and respiratory events.
• Gradient-boosted decision tree models were trained using hand-crafted features from physiological signals.

Main Results:

• Sleep stage classification achieved 0.840 accuracy and 0.791 Cohen's kappa.
• Arousal detection yielded an F1-score of 0.733, with limits of agreement for arousal index of ±15 events/h.
• Respiratory event detection achieved an F1-score of 0.818, with limits of agreement for apnea-hypopnea index of ±15 events/h.
• Model performance approached human inter-scorer agreement for sleep stages and arousals.

Conclusions:

• Developed ML models demonstrate performance comparable to human experts in key sleep scoring tasks.
• High consistency in expert annotations is vital for robust ML model development in sleep analysis.
• Quality-controlled annotations are recommended for building reliable automated sleep analysis systems.