Cardiorespiratory Sleep Staging Hits 70% Accuracy but Faces N1 Blind Spot and Validation Gap, Systematic Review Finds
Date: 2026-07-08
Automatic sleep staging using only cardiorespiratory signals — derived from wearable sensors rather than full electroencephalography — achieves a practically meaningful accuracy of approximately 70%, but the field faces systematic shortcomings that limit clinical deployment, according to a systematic review published July 7 in the Journal of Medical Systems.
The review, covering 35 studies published since 2010, found that no single signal modality or algorithmic approach outperformed others: cardiac signals alone, combined cardiorespiratory signals, and cardiorespiratory signals augmented with other non-EEG channels all produced comparable results, as did traditional machine learning versus deep learning architectures.
What they found
Researchers from Zhejiang University and Hangzhou City University searched four databases to identify studies developing automatic sleep staging models based on cardiorespiratory inputs — electrocardiography, photoplethysmography, respiratory inductance plethysmography, and comparable signals — without any EEG channels.
The core finding is a plateau of approximately 70% accuracy across the field, with no significant differences between:
- Signal modalities — cardiac-only, cardiorespiratory, or cardiorespiratory plus other non-EEG signals
- Algorithm families — traditional machine learning versus deep learning
Three systematic failure modes emerged:
1. Widespread lack of external validation. The vast majority of studies trained and tested on the same dataset, typically a single public repository. Models have not been shown to generalize across different populations, sensors, or recording environments.
2. Consistently poor classification of N1 sleep. The lightest stage of non-REM sleep — already the most difficult for human scorers to agree on — was systematically the worst-performing category across all reviewed approaches.
3. Limited generalization across diverse patient populations. Studies enrolling healthy young adults dominated the literature. Performance in older adults, individuals with sleep disorders, and pediatric populations remains largely unknown.
Why it matters
Full polysomnography is the gold standard for sleep staging, but it requires specialized equipment, a sleep laboratory, and trained technicians — making it impractical for large-scale or longitudinal use. Consumer wearables that capture heart rate and respiratory signals are already widely used, but their sleep staging algorithms are typically proprietary and unvalidated.
A validated, open cardiorespiratory-based staging pipeline could transform population sleep health monitoring. The 70% accuracy ceiling identified by this review sets a realistic benchmark for the field and defines where future improvements are needed: external validation, N1 detection, and population diversity.
Limits
The review itself includes only published studies and may reflect publication bias. The 70% figure is an aggregate — individual study performance varied. The review does not address proprietary algorithms used in commercial wearables, which may differ from academic models.
Bottom line
Cardiorespiratory sleep staging has reached a stable accuracy plateau that is sufficient for population-level screening but not yet for clinical diagnosis. The field’s next step is clear: rigorous external validation on diverse populations, not further algorithm tweaking.
Source
Chen W, He X, Zheng J, Chen S, Tian X. “Automatic sleep staging using cardiorespiratory signals: A systematic review of methodologies and performance.” Journal of Medical Systems. 2026 Jul 7;50(1):110. DOI: 10.1007/s10916-026-02435-9