Machine-learning models for sleep stage and respiratory event detection showed performance comparable to expert scorers in healthy and suspected sleep-disordered breathing participantsCan computer programs match human experts at reading your sleep data?

medRxiv Published April 9, 2026 Study authors: Kaneda, M.; Ogaki, S.; Nohara, T.; Fujita, S.; Osako, N.; Yagi, T.; Tomita, Y.; Ogata, T. DOI ↗ Editorial oversight: Dr. Ji-eun Park, MD · Brain, Mind & Pain

AI-generated summary of the cited source, checked by automated accuracy review. How we work

Key Takeaway

Note that machine-learning models for sleep staging and arousals matched expert agreement, but respiratory event detection lagged behind human scorers.

The study involved the development and evaluation of machine-learning models designed to classify sleep stages, detect arousals, and identify respiratory events from overnight polysomnography data. The population included healthy participants and individuals referred for suspected sleep-disordered breathing, though the specific sample size and study setting were not reported. The models were compared against expert scoring performed by four certified scorers.

For sleep stage classification, the models achieved an accuracy of 0.840, a Cohen's kappa of 0.791, and an F1-score of 0.841. Limits of agreement for total sleep time were approximately +/- 0.5 hours. For arousal detection, the models yielded an F1-score of 0.733, with limits of agreement for the arousal index of approximately +/- 15 events per hour.

Respiratory event detection resulted in an F1-score of 0.818. Limits of agreement for the apnea-hypopnea index were also within approximately +/- 15 events per hour. The study did not report adverse events, discontinuations, or specific tolerability data. Key limitations include the absence of reported sample size, study setting, and primary outcome definitions, as well as the lack of reported funding or conflict of interest information.

While absolute performance was high relative to prior studies, model performance for respiratory events remained below human inter-scorer agreement. The evidence is observational in nature, and the study phase and publication type were not reported. Clinicians should interpret these results with caution, particularly regarding the reliability of automated respiratory event detection compared to expert consensus.

Getting a proper sleep diagnosis often means spending hours reviewing complex data from overnight monitors. This process is expensive and relies on human experts who can get tired or disagree with each other. A new study asked if smart computer models could do this job just as well without the cost and delay. The team tested these models on data from healthy people and those suspected of having sleep-disordered breathing. They compared the computer's results against four certified human experts who manually scored the same data. The goal was to see if the machines could be trusted to help doctors make faster, more consistent decisions.

When it came to identifying sleep stages and detecting moments of waking up, the computer models performed very well. Their accuracy was comparable to the agreement seen between different human experts reviewing the same data. This is a big deal because it means the technology is reliable for these specific tasks. However, the results for detecting breathing pauses, known as respiratory events, were different. While the models still performed better than previous studies, they did not match the agreement seen between human experts for this specific job.

This study shows that computers are excellent at some parts of sleep analysis but not all. For breathing events, the computer's limits of agreement were wider than those of human experts. This means the technology is not yet ready to fully replace human review for counting breathing pauses. Until this gap is closed, doctors will likely need to keep using their expertise to ensure accurate diagnosis. The findings offer hope for faster care, but they also remind us that current technology has clear limits.

What this means for you:

Computers match humans for sleep stages but not yet for counting breathing pauses.

Study Details

EvidenceLevel 5

PublishedApr 2026

View Original Abstract ↓

Study Objectives 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. Methods 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. Results Sleep stage classification achieved accuracy 0.840, Cohen's kappa 0.791, and F1-score 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. Conclusions 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.

Machine-learning models for sleep stage and respiratory event detection showed performance comparable to expert scorers in healthy and suspected sleep-disordered breathing participantsCan computer programs match human experts at reading your sleep data?

Study Details

Intravenous tenecteplase improves functional outcomes but increases hemorrhage risk in non-large vessel occlusion acute ischemic stroke

Tenecteplase improves stroke recovery but increases bleeding risk in late treatment window

Clinical research that matters. Delivered to your inbox.

Machine-learning models for sleep stage and respiratory event detection showed performance comparable to expert scorers in healthy and suspected sleep-disordered breathing participantsCan computer programs match human experts at reading your sleep data?

More on sleep-disordered breathing

Study Details

Intravenous tenecteplase improves functional outcomes but increases hemorrhage risk in non-large vessel occlusion acute ischemic stroke

Tenecteplase improves stroke recovery but increases bleeding risk in late treatment window

Clinical research that matters. Delivered to your inbox.

Related in Neurology

From Other Specialties