Wearable Sleep Staging: Why Deep Sleep Is Hard to Measure

Your wearable can usually tell whether you slept. It has a much harder time telling whether you spent 83 minutes in deep sleep.

Across recent studies that compared wearable signals with polysomnography (PSG), the laboratory standard for sleep measurement, four-stage algorithms landed in roughly the same range: 75.3% to 80.8% accuracy and Cohen's kappa values of 0.6151 to 0.70. That is useful agreement, but it falls short of the agreement expected between trained human PSG scorers. The difficult boundary is N2 versus N3, or slow-wave sleep—the stage consumer apps label "Deep Sleep."

The practical conclusion is less dramatic than the marketing. Use a wearable to follow broad sleep patterns over weeks. Treat a single night's deep-sleep total as a noisy estimate, not a measurement precise enough to steer tomorrow's training, alcohol experiment, or supplement stack.

Different sensors, similar four-stage limits

Three PSG-concurrent validation studies reached similar results despite using different devices, models, and patient populations.

Zhang et al. (2026) tested a Mamba-based deep-learning model in 357 sleep-clinic participants. The ANNE One wearable combined chest ECG, finger PPG, accelerometry, and skin temperature. Against PSG, its four-class balanced accuracy was 75.30% and its kappa was 0.6151. This was the most sensor-rich setup in the group, yet it did not outperform the PPG-only studies.

Constantin et al. (2025) used reflective PPG from the wrist during 68 PSG-concurrent overnight recordings. Their deep-learning model reached 78.7% four-class accuracy and kappa of 0.68 against PSG. In a separate long-term ambulatory analysis, sleep/wake agreement reached kappa of 0.80 against actigraphy. That higher figure cannot be compared with PSG results: actigraphy is a movement-based reference, not a direct measurement of sleep stages.

Aguet et al. (2025) evaluated PPG-based deep-learning models during clinical PSG in 134 people with suspected sleep apnea. Wrist PPG achieved 80.8% four-class accuracy and kappa of 0.70. The cohort matters. Sleep apnea can fragment heart-rate variability and sleep architecture, so these results do not establish the same performance in healthy adults.

These are cross-study comparisons, not a controlled contest between sensors. The studies differed in model architecture, training data, placement, and population. Still, a chest ECG-plus-finger-PPG system did not clearly exceed wrist PPG alone. More sensors may help particular algorithms, but current evidence does not show that sensor count reliably predicts better four-stage sleep scoring.

The missing signal is the N2-to-N3 boundary

The most informative result comes from simplifying the classification task. Zhang and colleagues improved from kappa 0.6151 for four classes to 0.7289 for three classes: wake, light NREM sleep, and REM sleep. Removing the N2/N3 boundary produced a gain of about 0.11 kappa points.

That pattern fits sleep physiology. N3 is defined in PSG by high-amplitude delta activity on EEG, typically in the 0.5–4 Hz range. ECG, heart-rate variability, and PPG capture cardiac and peripheral vascular signals; they do not directly record that brain activity. A model can learn correlations between autonomic patterns and sleep stages, but correlation is not a substitute for the signal used to define N3 in the first place.

This does not make non-EEG wearables worthless. Sleep/wake and broad wake-NREM-REM classification have more support. Constantin et al. reported short-term sleep/wake kappa of 0.71 against PSG, while Zhang's three-class result was 0.7289. The evidence supports broad categories better than a minute-by-minute claim about slow-wave sleep.

More training data have not solved the problem

Pini et al. (2023) developed a heart-rate-based staging algorithm using 12,404 PSG recordings from clinics on three continents, then prospectively validated it with a wearable ECG device. The training cohort had a mean age of 42.3 years, with a standard deviation of 16.8 years, and 59% had sleep-disordered breathing. The available report does not provide a four-class kappa figure for direct comparison with the studies above.

Its value is scale, not a new numerical benchmark. A much larger and geographically broader ECG dataset established that heart-rate-derived staging is feasible. It did not supply evidence that ECG-derived signals can cleanly resolve N3 at the level of EEG-based scoring. The clinical mix also leaves unanswered how well such a model transfers to healthy athletes, older adults without sleep-disordered breathing, or adolescents.

Placement can matter, but the model decides how much

Aguet et al. found that moving the PPG sensor from the wrist to the upper arm reduced accuracy by 6.2 percentage points and reduced kappa by about 10% in their standard deep-learning model. Yet an inter-beat-interval-only model eliminated that placement penalty. The result does not prove that the wrist always beats the upper arm; it shows that placement effects depend on the algorithm as well as the signal.

Ring devices add another unresolved question. Finger PPG may provide a strong optical signal, but the primary studies above did not run a controlled finger-versus-wrist comparison using the same people and model. Svensson et al. (2024) independently evaluated the Oura Ring Gen3 with its OSSA 2.0 algorithm against multi-night ambulatory PSG in 96 participants across 421,045 scored epochs. That is valuable commercial validation, but it does not establish that finger placement is superior to wrist placement.

Why published research cannot certify your app

The leading wearable brands use proprietary algorithms, undisclosed training data, and firmware updates that can change performance without notice. Research models and commercial products therefore belong to related but separate evidence streams.

Commercial PSG validation exists. Schyvens et al. (2025) compared six commercial wrist-worn devices with PSG in 62 participants. The same research group also published a 2024 systematic review of Fitbit Charge 4, Garmin Vivosmart 4, and WHOOP against PSG. Lee et al. (2023) prospectively evaluated 11 wearable, nearable, and airable consumer trackers against PSG.

Those studies show that commercial devices can be tested against PSG. They do not create a clean head-to-head comparison with Zhang's, Constantin's, or Aguet's research algorithms. Per-stage N3 sensitivity is also not reported consistently across commercial studies, so no published figure can tell you exactly how much trust to place in one brand's deep-sleep graph.

How to use the data without fooling yourself

A wearable's sleep stages are best treated as estimates with different error bars, even when the app displays them as exact minutes.

• Use total sleep time and broad sleep/wake patterns as your most dependable outputs.
• Review 7- to 14-night rolling averages before interpreting a change in sleep-stage trends.
• Do not change a training plan or supplement protocol because one night produced an unusually low or high deep-sleep total.
• Prefer peer-reviewed, PSG-benchmarked validation that reports stage-specific outcomes over manufacturer accuracy claims.

The evidence could change. A PSG study in healthy adults showing non-EEG N3 discrimination above kappa 0.80, or a controlled comparison showing a commercial algorithm clearly exceeding published research systems, would force a reassessment. Neither result appears in the current evidence. For now, the honest reading is simple: wearables can help you track sleep behavior, but they cannot see the EEG signal that defines deep sleep.