Apple Watch Sleep Stages vs. Polysomnography: How Accurate Is the Staging?

Apple Watch sorts each night into Awake, REM, Core and Deep sleep using its accelerometer, not brain waves. Core is light and intermediate sleep (N1 and N2), Deep is slow-wave sleep (N3). Apple's own validation against sleep-lab polysomnography found a four-stage kappa of 0.68: good for trends over weeks, not a substitute for a sleep study.

What Apple Watch actually measures

Sleep staging in a clinical sleep laboratory relies on polysomnography (PSG): a comprehensive sensor array including electroencephalography (EEG) to capture brain activity, electro-oculography (EOG) to detect eye movements, electromyography (EMG) for muscle tone, respiratory sensors, and pulse oximetry. Trained sleep technologists visually inspect these combined signals to assign a stage label to every 30-second window of the night, following the American Academy of Sleep Medicine (AASM) scoring rules.

Apple Watch uses none of those sensors for sleep staging. Instead, it relies on its 3-axis accelerometer - the same sensor that tracks steps, detects falls, and monitors activity. The algorithm reads the continuous accelerometer signal and extracts not only gross body movement but also subtle oscillations caused by breathing. These respiration-induced motion patterns carry information about physiological state that an ordinary activity tracker's step-counting algorithm would discard. Each 30-second epoch is classified into one of four outputs: Awake, REM, Core, or Deep.

This approach is not simply a limitation - it is a deliberate design choice that makes continuous multi-night tracking possible in a way that PSG never could be. Apple's white paper on sleep staging notes that the accelerometer-based method outperforms traditional wrist actigraphy, which uses binned activity counts rather than the high-frequency signal Apple's algorithm processes.

How the four stages map to PSG's five

Clinical PSG defines five stages: Awake, N1, N2, N3, and REM. Apple Watch outputs four. The mapping is:

The name "Core" - rather than "Light" - was a deliberate choice. N2 typically accounts for more than 50% of a full night's sleep. Calling it "Light" could imply it is less important or less restorative, which the sleep science does not support. N2 contains sleep spindles and K-complexes and is a normal, central feature of healthy sleep architecture. Apple chose Core to convey that this is the foundational majority of a night's sleep, not a thin or inferior layer.

The practical implication: if your Apple Watch shows a night with a large percentage of Core sleep and a small percentage of Deep, that is not necessarily cause for alarm. It may simply reflect normal sleep architecture.

What the accuracy data actually shows

Apple published a detailed white paper documenting the algorithm's development and validation - and updated it in October 2025 to include improvements made in 2024 and 2025. The numbers reported here come from Apple's own validation study, which compared Apple Watch against expert-scored PSG in an independent participant group not used in algorithm training.

Original algorithm (watchOS 9, 2022): The validation set included 166 participants and 299 recorded nights. The four-stage kappa was 0.63 (standard deviation 0.13). Binary sleep/wake classification achieved a median sensitivity of 97.9% - meaning nearly all true sleep epochs were correctly identified - and a specificity of 75.0%, meaning three quarters of true awake epochs were correctly classified.

Updated algorithm (watchOS 26 / iOS 26, 2025): Using foundation models trained on data from the Apple Heart and Movement Study, Apple improved the algorithm's handling of "quiet wake" - the state of lying still just before falling asleep or just after waking, which is particularly difficult to distinguish from light sleep. In the same validation dataset, the updated algorithm achieved a four-stage kappa of 0.68 (SD 0.11) and improved wake detection from 70% correct to 79% correct. Binary sensitivity shifted slightly to 96.8% and specificity improved to 78.9%.

In a separate cohort of clinical patients undergoing PSG as part of actual clinical care - a more challenging population that includes people with sleep disorders - the updated algorithm achieved a kappa of 0.66, binary sensitivity of 95.0%, and specificity of 86.5%.

What kappa of 0.68 actually means

Cohen's kappa corrects for the agreement that would occur by chance alone.

Apple Watch: 0.63 (original), 0.68 (updated), 0.66 (updated, clinical patients). Trained PSG technologists scoring the same recording independently: typically 0.76-0.83, depending on the study. Landis and Koch call 0.61-0.80 "substantial". The technologists' range is the ceiling Apple Watch is implicitly competing against, because PSG technologists themselves do not agree perfectly, especially at stage boundaries.

Apple's white paper notes explicitly that the most common Apple Watch misclassification - Deep sleep being scored as Core - mirrors the most common disagreement between human PSG raters. The errors are not random; they are concentrated at the boundaries between physiologically adjacent stages, which is exactly where even expert humans disagree most.

This does not make the errors irrelevant. A four-stage kappa of 0.68 means that a meaningful minority of epochs are misclassified. If you are trying to measure precise deep sleep minutes on a specific night, the number Apple Watch gives you is an estimate with non-trivial uncertainty. But if you are tracking whether your deep sleep percentage has consistently declined over the past month - the kind of trend an app like Sam can surface from your Apple Health data - the directional signal is meaningful and more reliable than any single night's snapshot.

A note on the different accuracy numbers you may see quoted elsewhere: kappa, per-epoch sensitivity, and total time-per-stage error all measure slightly different things. The kappa above is a chance-corrected agreement statistic across the four-stage classification. Other studies report binary sleep/wake sensitivity (typically ~95-98%) and total per-stage duration error (an independent 2024 study found ~43 minutes mean underestimate of deep sleep on Series 8) - a pattern of consistent underestimation similar to what shows up when Apple Watch VO2 max is compared against lab tests. Those figures look different because they are different metrics, not because they contradict each other - context on what they each measure and where they come from is covered in sleep duration is a misleading health metric and the comparison summary in the best sleep tracking apps for Apple Watch in 2026.

What it is genuinely useful for

The practical strength of Apple Watch sleep staging is not that any single night is precisely measured. It is that you accumulate months of consecutive nights in a way that no clinical sleep study could provide.

PSG in a sleep laboratory typically captures one to three nights. It is expensive, disruptive (sensors glued to your scalp, cables, unfamiliar environment), and often affected by the "first-night effect" - the documented phenomenon of sleeping worse in a lab than at home. Apple Watch captures your actual sleep in your own bed, every night, over months.

Across that longer window, the signal-to-noise ratio improves. If your REM percentage drops consistently in weeks when work stress is high, if Deep sleep declines after multiple nights of alcohol, if your total time asleep shifts with season or schedule - these trends emerge clearly from imperfect but consistent data.

The Apple Support page also notes that the Health app offers a Sleep Score (0-100) combining sleep duration, bedtime consistency, and interruptions. That score is distinct from stage quality - it does not reward deep sleep over core sleep, but tracks the structural consistency and duration of your sleep across nights.

What it cannot tell you

Apple Watch sleep staging is not a diagnostic tool for sleep disorders. Sleep apnea, insomnia, restless legs syndrome, and narcolepsy all require clinical evaluation with validated tools. Apple's white paper explicitly states that sleep stages are "not intended for clinical use" - the same wellness-only framing that applies to Apple Watch blood oxygen readings. The feature's own validation showed lower kappa in clinical populations - patients with sleep disorders are harder to classify correctly, and the Apple Watch algorithm has lower accuracy in those groups.

If you have symptoms of a sleep disorder - excessive daytime sleepiness, witnessed pauses in breathing, restless legs, or persistent difficulty initiating or maintaining sleep - the appropriate step is a referral to a sleep medicine specialist, not closer inspection of your Apple Watch sleep chart.

What Apple Watch can do is provide a prompt: if your data consistently shows fragmented sleep, frequent awake epochs, or unusually low REM over weeks, that pattern is a reasonable basis for a conversation with a healthcare professional. It is a wellness signal, not a diagnosis.

In Sam

See your Apple Watch sleep stages in Sam and take your month to the appointment

You see each night's stages and how your sleep timing shifts over weeks, and you bring the month along as a PDF.

• Instead of Apple Health alone: Sam shows each night as a timeline of Awake, REM, Core and Deep sleep. Week and month views plot every night by clock time, so shifts in timing and regularity stand out.
• Instead of a single night's number: heart rate variability, resting heart rate and respiratory rate each have their own trend, next to your personal baseline.
• Instead of memory: you create the monthly report yourself, and it includes a sleep page. In the symptom diary you can log complaints such as difficulty falling asleep.

Sam shows what the watch records, so the accuracy limits above apply to Sam too.

Try Sam for free

Sam is a wellness companion, not a medical device. No diagnosis, no replacement for medical advice.

Related reading

• HRV
• complete Apple Watch sensor breakdown for 2026