PPG-Based Sleep Staging Algorithms
PPG-based sleep staging classifies sleep into wake, light (N1/N2), deep (N3), and REM stages using heart rate dynamics, HRV spectral changes, respiratory rate modulation, and accelerometer-derived movement patterns. Consumer wearables achieve 65–80% epoch-by-epoch agreement with polysomnography.
Each sleep stage produces characteristic autonomic signatures detectable in PPG. Deep sleep (N3) shows the lowest heart rate, highest HRV (RMSSD), and most prominent respiratory sinus arrhythmia. REM sleep increases heart rate variability with sympathetic bursts and irregular respiratory patterns. Light sleep (N1/N2) shows intermediate values. Wake periods exhibit higher heart rate, reduced HRV, and increased movement.
Machine learning classifiers extract features from 30-second epochs matching PSG scoring convention. Features include: mean HR, RMSSD, LF/HF ratio, respiratory rate, respiratory amplitude variability, body movement count, and movement intensity from accelerometer. Random forest classifiers achieve 70–75% four-stage accuracy, while deep learning (LSTM, CNN-LSTM) achieves 75–80% — approaching the 82–85% inter-rater agreement between human PSG scorers.
Oura Ring achieves the highest published consumer wearable sleep staging accuracy (79% four-stage), attributed to finger PPG's superior signal quality during sleep. Wrist-based devices (Apple Watch, Garmin, Fitbit) achieve 65–74% accuracy. Deep sleep detection is the most challenging stage with only 50–65% sensitivity across devices, while REM detection achieves 70–80% sensitivity. Sleep onset and wake detection are most reliable (85–95% sensitivity).
Frequently Asked Questions
How accurate is wearable sleep staging compared to polysomnography?
Consumer wearables achieve 65–80% epoch-by-epoch agreement with PSG for four-stage classification. This compares to 82–85% inter-rater agreement between trained PSG technicians. The largest errors are in deep sleep detection and N1/N2 differentiation.
Why is deep sleep hardest to detect from PPG?
Deep sleep occupies only 15–20% of total sleep time, creating class imbalance. Autonomic features during deep sleep overlap with relaxed wakefulness. Movement-based features cannot differentiate deep sleep from quiet wake or light sleep.
Can PPG-based sleep staging detect sleep disorders?
PPG sleep staging can flag abnormal sleep architecture (reduced deep sleep, fragmented REM) that may indicate sleep disorders. However, it cannot diagnose conditions like narcolepsy or REM behavior disorder that require EEG-based staging.