Is Apple Watch Sleep Tracking Accurate? Good at Total Sleep, Weak on Deep Sleep and Time Awake

Short answer: accurate for whether and how long a person slept, much less so for what kind of sleep it was. In two independent lab studies of the Series 8 against polysomnography, the watch detected 96 to 97% of sleep and put total sleep time not significantly different from the lab in one study and 20 minutes high in the other. It is weaker at spotting wakefulness: in one study only 52% of awake time in bed was recognised, and time awake after falling asleep was undercounted by 10 to 21 minutes. Deep sleep read 25 to 43 minutes low, because only about half of it was labelled as deep. Stage-level agreement was well short of perfect (kappa 0.30 to 0.60, where 1 is full agreement). The sleep score has no independent validation that we could find.

What the lab studies found

The reference method for sleep is polysomnography (PSG): a night in a lab with electrodes on the scalp, beside the eyes and on the chin, scored by a technologist in 30-second blocks called epochs. Every study below compared the watch's epoch-by-epoch output with a PSG recording made on the same night.

Kappa is agreement corrected for chance: 0 is no better than guessing, 1 is perfect. On the scale Schyvens et al. apply, 0.21 to 0.40 is fair and 0.41 to 0.60 is moderate. None of the three independent papers reports which watchOS version was running, which matters because Apple has since changed the algorithm (see below).

All three independent studies recorded a single night per person. Robbins et al. tested healthy adults aged 20 to 50, was funded by Oura Ring Inc., and lost Apple Watch data for 6 of 35 participants. Schyvens et al. was publicly funded with no declared conflicts, but only 20 people wore the Apple Watch and the sample was 52 men to 10 women, with over half having some degree of sleep apnea. Lee et al. had authors with a stake in one of the phone apps tested.

Sleep versus wake: the part the watch does well, and the part it does not

Sleep trackers are usually judged on two numbers. Sensitivity is the share of true sleep epochs the device calls sleep. Specificity is the share of true wake epochs it calls wake.

Robbins et al. (n=29) reported sensitivity of 97% and did not report specificity. Schyvens et al. (n=20) reported 96.27% and 52.15%. Lee et al. (n=26) gave no two-state figure, but the watch caught 45% of wake epochs. Apple's own validation set gives a median sensitivity of 97.9% and a median specificity of 75.0%.

A night is mostly sleep, so a device that labelled every epoch as sleep would score a sensitivity of 100%. Specificity is the harder test, and here the independent figures are about half: in the Schyvens study, roughly one in two epochs the lab scored as awake was scored by the watch as asleep.

The reason is mechanical. A person lying awake and still produces very little wrist movement, which to a motion sensor looks like sleep. The result is a predictable direction of error:

Positive numbers mean the watch read higher than the lab. Robbins et al. give no Apple-specific figure in minutes for total sleep time; they state that all three devices they tested were within 10 minutes. The watch gives a slightly flattering night: a little more sleep, fewer and shorter awakenings. The distortion is small for a sound sleeper and larger for someone who lies awake for long stretches, the person most likely to be checking.

The limits of agreement matter as much as the averages. In the Schyvens data, an individual's total sleep time could be 38 minutes under or 78 minutes over the lab figure. The average error for time to fall asleep was under two minutes, but for an individual it could be off by more than 45 minutes either way, so the watch is a poor tool for answering how long it takes to fall asleep on any single night.

Sleep stages: where the accuracy drops

The watch reports four states: Awake, REM, Core and Deep. Core is Apple's name for what sleep labs call light sleep (stages N1 and N2); Deep corresponds to N3.

Deep sleep is the consistent weak point. In all three studies the watch recognised about half of true deep sleep or less. The Robbins figures show the character of the error: precision was 87.8%, meaning that when the watch did say "deep" it was usually right, but sensitivity was 50.5%, meaning it stayed silent about half the deep sleep that occurred. The missing half was filed under Core, which is why light sleep is overestimated by 45 to 59 minutes in the same studies. Apple's white paper reports the same pattern: its most common misclassification overall was true Deep sleep labelled as Core.

Nightly stage totals also tracked the lab poorly. Across the three devices Robbins et al. tested, intraclass correlations for nightly deep and REM totals were 0.13 to 0.37, against 0.56 to 0.85 for total sleep time (the Apple Watch had the 0.85).

How the Apple Watch tracks sleep, and why stages are an inference

Sleep stages are defined by brain activity, which a lab measures directly with electrodes. A watch has no electrodes on the head. It has a motion sensor and an optical pulse sensor on the wrist. Apple's white paper describes its stage classifier as "based on accelerometer signals derived from Apple Watch", classifying each 30-second epoch and developed against PSG scored by certified technologists. The document does not name heart rate as an input. Apple's separate sleep apnea paper notes that the accelerometer captures fine movements, including motion associated with breathing. From patterns in those signals the watch infers which brain state is most likely.

The relationship between body signals and brain state is statistical, and it differs between people. Apple's document reports that kappa fell with more stage transitions and with more severe sleep apnea, and was lower on clinical recordings (0.55). It also says that "Apple Watch sleep stages aren't intended for clinical use."

Adding pulse data, such as the beat-to-beat variation discussed in HRV by age, would not change the basic point: even a good wrist heart rate reading is not a brain recording.

Why deep sleep numbers differ between brands

Because each company trains its own model on its own data and chooses its own trade-off between over-calling and under-calling a stage. The Schyvens study, which ran six wrist devices against the same lab method, shows how far apart those choices land.

Against the same lab recordings, the Apple Watch Series 8 (n=20) read deep sleep 25.2 minutes low. The Fitbit Sense (n=37) and Fitbit Charge 5 (n=39) were within 4 minutes, differences that were not statistically significant. The Whoop 4.0 (n=40) read 31.5 minutes high and the Garmin Vivosmart 4 (n=25) 44.4 minutes high. Four-stage kappa ran from 0.21 (Garmin) through 0.37 (Whoop) and 0.41 to 0.42 (the Fitbits) to 0.53 (Apple).

Two people with identical sleep, one wearing an Apple Watch and one a Whoop, could see deep sleep totals nearly an hour apart on these averages. Neither is measuring deep sleep; each reports its model's guess.

It follows that comparing a watch's deep sleep total with published guidance on how much deep sleep a person needs is comparing two different kinds of number. The guidance figures come from lab recordings.

Apple Watch vs Oura

Two of the studies tested both on the same nights.

In Robbins et al., four-stage kappa was 0.60 for the watch and 0.65 for the Oura Ring Gen3. The ring detected 79.5% of deep sleep against the watch's 50.5%, and its nightly deep sleep total did not differ significantly from the lab, where the watch read 43 minutes low. Concordance for the sleep-or-wake call favoured the watch (ICC 0.85 against 0.74). In Lee et al., kappa was 0.30 for the watch (n=26) and 0.35 for the Oura Ring 3 (n=53); deep sleep sensitivity was 0.41 against 0.78, and wake sensitivity 0.45 against 0.38.

On these two studies the ring is somewhat better at stages, deep sleep in particular, and the watch is as good or better at the plain sleep-or-wake call. The caveats are substantial. The Robbins study was funded by Oura and its lead author sits on Oura's medical advisory board. The Lee study, funded by neither company, found a smaller gap in overall agreement. Both are single-night studies, and Apple's algorithm has since changed. More detail on the ring is in the separate review of Oura Ring accuracy.

Apple's own numbers, and what changed in 2025

Apple's white paper is the largest dataset here: a validation set of 166 people and 299 nights (mean age 47), separate from the 858 people and 1,171 nights used to design the algorithm. For the original four-stage algorithm, introduced with watchOS 9 in 2022, Apple reports median sensitivity of 97.9%, median specificity of 75.0% and a four-stage kappa of 0.63. In a separate clinical set of 236 people, kappa was 0.55 and specificity 72.5%.

The sensitivity matches the independent studies closely and the kappa is slightly above the best independent result. The specificity is clearly higher than the 52% Schyvens et al. measured in a small sample with many apnea patients.

The paper also describes an updated algorithm released with watchOS 26 in 2025. Apple reports that it raised wake accuracy from 70% to 79%, validation-set specificity from 75.0% to 78.9% and kappa from 0.63 to 0.68, and clinical-set specificity from 72.5% to 86.5%, with sensitivity slipping to 96.8% and 95.0%. If those figures hold up outside Apple, the update addresses the watch's main weakness. We did not find an independent PSG study of the updated algorithm; every independent number in this article describes the earlier one.

Sleep apnea notifications: a different feature with different evidence

The sleep apnea notification is separate from sleep staging and is the only sleep feature on the watch with regulatory clearance. The FDA cleared it on 13 September 2024 (510(k) number K240929) as an over-the-counter device to assess risk of sleep apnea. Its indication is limited to patterns suggestive of moderate-to-severe sleep apnea in adults with no prior diagnosis; mild apnea is not a target. According to Apple's validation paper, it uses the accelerometer to estimate nightly "breathing disturbances", reviews 30 days of data, requires at least 10 nights, and notifies if at least half of those nights are elevated. Apple's September 2024 document lists Series 9 or later and Ultra 2 as supported.

A notification is very likely to be real: nobody without apnea was flagged. Silence means much less: more than half of people with moderate sleep apnea received no notification, and about one in ten with severe apnea. Apple's paper says as much, noting that the feature does not consider symptoms and that the absence of a notification does not rule the condition out.

The overall sensitivity and specificity also appear in the FDA's 510(k) summary, but they come from Apple's study; we did not find an independent replication. The reference test was a home sleep apnea test over at least two nights, not in-lab PSG.

The sleep score

Recent versions of watchOS show a morning score from 0 to 100. Apple's user guide gives the formula in outline: 50 points for sleep duration, 30 for bedtime consistency over the last 13 nights, and 20 for interruptions, meaning how often and how long the wearer was awake. Bands run from Very Low (0 to 40) through Low (41 to 60), OK (61 to 80) and High (81 to 95) to Very High (96 and above).

The guide makes no claim about validation, and we found no published study, from Apple or anyone else, testing the score against PSG or against any health outcome.

The score inherits the errors above. Half of it rests on duration, the watch's strongest measurement. Twenty points rest on interruptions, the measurement the independent studies found weakest, with the error running in the flattering direction. Sleep stages do not appear in the published formula.

What is not worth doing

Acting on the nightly deep sleep figure. It is the least accurate number on the screen, biased low by 25 to 43 minutes on average in independent tests, with individual errors far wider.

Comparing stage totals across brands, or with someone else's watch. The cross-device spread in the Schyvens data was more than an hour between the lowest- and highest-reading devices.

Treating a low score as the explanation for a bad day. Baron and colleagues coined "orthosomnia" in 2017 for patients who sought treatment for sleep problems they had diagnosed themselves from tracker data, a perfectionistic pursuit of ideal sleep. The paper is a case series of three patients, not a prevalence study, and it does not show that trackers cause insomnia. It is a caution that a noisy number, checked every morning, can become its own source of worry.

Treating no apnea notification as an all-clear. On Apple's own figures the feature misses most moderate cases.

Upgrading the watch for better staging. Apple's paper reports that watch model had no effect on agreement. The changes that mattered were software.

What to actually do

Use the watch for the things it measures well. Total sleep time, and the timing of bed and wake, are within minutes to about 20 minutes of a lab on average. Those are also the sleep variables a person can most directly change.

Read weekly averages, not single nights. Individual nights carried errors of 40 to 80 minutes in total sleep time at the extremes. None of the studies followed the same people over many nights, so how well averaging cancels the error has not been tested directly.

Mentally add some wake. The true time awake is more likely higher than the watch's figure, particularly for restless sleepers.

Treat stages as a rough sketch. REM and Core figures are approximate; Deep is a floor more than an estimate.

Let symptoms outrank the screen. Loud snoring, witnessed pauses in breathing, or persistent daytime sleepiness are reasons to speak to a clinician whether or not the watch has said anything. A good score does not cancel feeling unrested, and a poor one does not outweigh feeling fine. The same applies to the watch's VO2 max estimate: useful for direction, unreliable as a verdict.

Questions people ask

How accurate is Apple Watch sleep tracking? For sleep versus wake, good: two independent lab studies found 96 to 97% of sleep detected and total sleep time within about 20 minutes of polysomnography on average. For sleep stages, much weaker, with four-stage kappa between 0.30 and 0.60 across three studies of the Series 8.

Is Apple Watch deep sleep accurate? It is the weakest of the watch's sleep figures. Three studies found it recognised 41 to 51% of lab-scored deep sleep, and two measured an average underestimate of 25 and 43 minutes. When the watch does label a period as deep it is usually right; it simply misses about half.

How accurate are Apple Watch sleep stages? Independent studies put epoch-by-epoch agreement with a sleep lab at 56 to 75%, with kappa from 0.30 to 0.60. Apple reports 0.63 for its original algorithm and 0.68 after a 2025 update. Light sleep and REM are identified more reliably than deep sleep or wake.

Why does my Apple Watch say I got no deep sleep? Most likely because the algorithm is conservative. It assigns Deep only when confident and otherwise defaults to Core, so deep sleep that occurred is often shown as Core. The watch cannot observe brain waves.

Apple Watch vs Oura: which is more accurate for sleep? In two same-night comparisons the Oura Ring Gen3 was better at deep sleep (78 to 80% detected against 41 to 51%) and slightly ahead on stage agreement overall. The watch matched or beat it on the basic sleep-or-wake call. One of those studies was funded by Oura.

How does Apple Watch track sleep? Apple describes its stage classifier as based on accelerometer signals, analysed in 30-second windows and trained against lab recordings. The motion sensor picks up body movement and the small wrist motion of breathing. The watch infers sleep stages from those patterns; it does not measure brain activity.

Is the Apple Watch sleep score accurate? Nobody has shown that publicly. The score combines duration (50 points), bedtime consistency (30) and interruptions (20). We found no independent study validating it against a sleep lab or health outcomes. Its duration component rests on a sound measurement; its interruptions component rests on the watch's weakest one.

How accurate is Apple Watch sleep apnea detection? In Apple's own validation study (1,278 people analysed), the notification, which targets moderate-to-severe apnea only, caught 66.3% of moderate-to-severe cases overall: 89.1% of severe and 43.4% of moderate. Specificity was 98.5%. A notification is therefore worth following up, but no notification does not rule sleep apnea out.

Can Apple Watch detect when you fall asleep? On average, yes: the mean error in time to fall asleep was under 2 minutes in the one independent study that reported it for the watch. For an individual night it can be off by more than 45 minutes in either direction, because lying still while awake looks like sleep to a wrist sensor.

This article summarises published validation studies and manufacturer documentation for general information and is not medical advice. A consumer sleep tracker is not a diagnostic test; anyone with persistent poor sleep, loud snoring, pauses in breathing or daytime sleepiness should speak to a clinician and should not rely on a watch reading or the absence of a notification.