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

To develop machine learning models for sleep stage classification, arousal detection, and respiratory event detection from overnight polysomnography, and to evaluate their performance relative to expert scorers.

Overnight polysomnography recordings were obtained from healthy participants and participants referred for suspected sleep-disordered breathing. Four certified scorers completed calibration sessions and generated reference annotations for sleep stages, arousals, and respiratory events. A subset of recordings was independently annotated by all scorers to support consensus analyses, enabling direct comparison between model outputs and human inter-scorer agreement. Gradient-boosted decision tree models were trained using hand-crafted features derived from standard physiological signals.

Sleep stage classification achieved an accuracy of 0.840, a Cohen's kappa of 0.791, and an F1-score of 0.841, with limits of agreement for total sleep time of approximately ±0.5 h. Arousal detection achieved an F1-score of 0.733, with limits of agreement for the arousal index of approximately ±15 events/h. Respiratory event detection achieved an F1-score of 0.818, with limits of agreement for the apnea-hypopnea index also within approximately ±15 events/h. In consensus analyses, model performance was comparable to human inter-scorer agreement for sleep stages and arousals, while remaining below human inter-scorer agreement for respiratory events, despite high absolute performance relative to prior studies.

The proposed models achieved performance approaching human-level agreement across major sleep scoring tasks. These findings indicate that high consistency in expert annotations is a key factor underlying robust model performance and support the use of quality-controlled annotations for developing reliable automated sleep analysis systems.