Sleep stages are the distinct brain states you pass through while asleep, scored in a sleep lab as wake, three stages of non-REM sleep (N1, N2 and N3) and REM sleep. Wearables such as Apple Watch estimate the same states from the wrist and report them as Core (N1 plus N2), Deep (N3) and REM.
The mix of stages shifts from night to night and across your lifetime. A watch estimates it with far more error than it estimates total sleep. Knowing what each stage is, how much of each a normal night contains and where wearable numbers go wrong lets you judge whether a low Deep figure on your report reflects a change in your sleep or a misclassified epoch. The wearable data quality guide compares stage accuracy with other wearable metrics.
01The four AASM stages
In a sleep lab, a technologist scores polysomnography in 30 second epochs using rules from the American Academy of Sleep Medicine (AASM). Each epoch gets one label, based on brain waves (EEG), eye movements (EOG) and chin muscle tone (EMG). The AASM published these rules in 2007 and revised them in 2012 (Boulos et al., 2019). Stage N3 replaced the older Rechtschaffen and Kales stages 3 and 4 (Berry et al., 2012). The manual's version 3 was released in February 2023, and accredited sleep facilities had to adopt it by the end of that year (AASM, 2023).
N1. The alpha rhythm of relaxed wakefulness gives way to low-amplitude activity, mostly between 4 and 7 Hz, often with slow rolling eye movements and vertex sharp waves (Berry et al., 2012). N1 is the doorway into sleep. It usually lasts a few minutes and makes up about 5% of sleep time (Cleveland Clinic, 2023).
N2. This stage carries two signature waveforms. Sleep spindles are bursts of 11 to 16 Hz activity lasting at least half a second. K complexes are large, sharp waves that stand out from the background (Berry et al., 2012). N2 fills more of the night than any other stage, often more than half of it (Apple, 2025).
N3. Technologists score slow-wave sleep when at least 20% of an epoch consists of 0.5 to 2 Hz waves with a peak-to-peak amplitude above 75 µV over the frontal regions, and the rule applies irrespective of age (Berry et al., 2012). Heart rate and breathing fall to their lowest levels of the night, and you are hard to wake (NINDS).
REM. Stage R requires three findings in the same epoch. They are a low-amplitude, mixed-frequency EEG, chin muscle tone at or near its lowest level of the recording, and rapid eye movements (Berry et al., 2012). Breathing turns faster and irregular, heart rate and blood pressure rise toward waking levels, and the arm and leg muscles are temporarily paralyzed. Most dreaming happens here (NINDS).
02How wearables collapse the stages
A watch or ring has no electrodes on your head, so it cannot see spindles or slow waves. It infers the stage of each epoch from motion, and in some devices from pulse signals, and reports four classes. N1 is a small share of a normal night (Apple, 2025), and Apple, Oura and WHOOP all fold it into N2 as a single lighter class.
| Lab stage | Apple Watch and Apple Health | Oura | WHOOP |
|---|---|---|---|
| Wake | Awake | Awake | Awake |
| N1 and N2 | Core | Light | Light |
| N3 | Deep | Deep | Slow wave sleep |
| REM | REM | REM | REM |
Apple chose the word Core over "light" on purpose. N2 dominates a normal night and carries the spindles and K complexes that sleep researchers treat as markers of healthy sleep (Apple, 2025). Apple's classifier reads the watch's 3-axis accelerometer at a rate high enough to register the small movements of breathing, and it labels every 30 second epoch. Oura's published model combines motion with autonomic nervous system signals measured at the finger and circadian features (Altini and Kinnunen, 2021). WHOOP reports light sleep, slow wave sleep and REM (Miller et al., 2021).
In Apple Health, each stage arrives as its own sleep sample. A tracker that does not stage sleep writes plain Asleep time instead, and that difference changes how Titan scores the night, as the last section explains. How sleep is detected covers how Titan assembles those samples into one night.
03How a night is structured
Sleep begins in N1, deepens through N2 into N3, and the first REM period arrives about 90 minutes after you fall asleep (NINDS). One pass through non-REM and REM sleep is a sleep cycle. A cycle usually runs 90 to 120 minutes, and an eight hour night holds four or five of them (Cleveland Clinic, 2023).
The cycles change as the night goes on. N3 comes in its longest periods during the first half of the night (NINDS). The first REM period is the shortest, around 10 minutes, and each one after it runs longer, up to about an hour (Cleveland Clinic, 2023). So the timing of a short night decides which stage it costs. Going to bed late and waking on time trims a little of every stage. An alarm 90 minutes early removes a disproportionate share of REM, because the longest REM periods sit in the final cycles.
Time awake shapes the night too. The homeostatic sleep drive builds with every hour awake and makes the next sleep longer and deeper (NINDS). Your circadian rhythm sets the timing window that drive works within.
04Typical proportions
Stage percentages depend on the population, the scoring rules and even the denominator a study uses. Compare values within one study, and treat any single number as a rough reference.
| Source | Population | N3 or Deep | REM |
|---|---|---|---|
| Apple (2025) validation set | 166 adults, 299 nights, mean age 47 | 12.8% | 21.4% |
| Van Cauter et al. (2000) | Healthy men aged 16 to 25 | 18.9% | Not stated |
| Van Cauter et al. (2000) | Healthy men aged 36 to 50 | 3.4% | Not stated |
| Mitterling et al. (2015), median | Screened healthy sleepers aged 30 or under | 20.7% | 15.5% |
| Mitterling et al. (2015), median | Screened healthy sleepers over 60 | 14.9% | 10.3% |
| Cleveland Clinic (2023) | General adult guidance | About 25% | About 25% |
Apple's validation set is the closest match to a current watch owner. Its nights were scored against polysomnography, averaged 378 minutes of sleep, and split 65.9% Core, 21.4% REM and 12.8% Deep (Apple, 2025). That works out to about 81 minutes of REM and 48 minutes of Deep. Mild sleep apnea appeared on 11.8% of those nights, so the figures describe ordinary adult sleep. The 25% deep sleep figure common on health websites sits well above what lab studies record in middle-aged adults.
05How stages shift with age
Ohayon and colleagues (2004) pooled 65 studies covering 3,577 people aged 5 to 102. In adults, the share of slow-wave sleep, the share of REM sleep, total sleep time and sleep efficiency all fell with age, while the shares of stage 1 and stage 2 rose. Only sleep efficiency kept declining after 60. The same analysis found that effect sizes shrank when studies screened participants carefully for sleep disorders, medical and psychiatric illness, and drug or alcohol use.
The deep sleep decline in men is steep and early. In Van Cauter's pooled data from 149 healthy men, slow-wave sleep fell from 18.9% of the night between 16 and 25 to 3.4% between 36 and 50, replaced by lighter sleep, with no significant loss of REM over that span (Van Cauter et al., 2000). REM fell later, by 10 minutes per decade from midlife to late life, while time awake rose by 28 minutes per decade. Mitterling and colleagues (2015) found the same direction in 100 screened healthy sleepers aged 19 to 77. N3 and REM shares fell with age, and N1 and N2 shares showed no age difference.
The largest pooled analysis of nights scored under current AASM rules found a smaller age effect on stages. Across 169 studies and 5,273 healthy adults, each decade of age cost 10.1 minutes of total sleep, added 9.7 minutes of wake after sleep onset and added 0.5 percentage points of N1 (Boulos et al., 2019). The per-decade changes in N2, N3 and REM share were not statistically significant. Every one of these studies agrees on more wakefulness and less total sleep with age. How much deep sleep you lose depends on the study, your sex and how healthy the sample was. A 45 year old with 10% Deep is within what the lab literature records. Your own baseline is a better reference than any population table.
06How much deep and REM sleep adults need
No clinical guideline sets a minimum number of deep or REM minutes. The AASM and the Sleep Research Society recommend that adults sleep 7 or more hours per night on a regular basis (Watson et al., 2015), and NINDS puts most adults at 7 to 9 hours. Stage targets in sleep apps are reference points chosen by each app maker.
Applying the Apple validation set's average proportions to 7 hours of sleep gives about 54 minutes of Deep and 90 minutes of REM. Treat those figures as descriptive averages. The practical route to more of both stages is more total sleep at a regular time. N3 is concentrated early, so a short night costs relatively little of it. REM is concentrated late, so REM is the stage an early alarm removes. The sleep for athletes guide covers what each stage contributes to training adaptation, and sleep hygiene covers the habits that protect total sleep.
07What training and alcohol do to your stages
Exercise changes sleep architecture only a little. A meta-analysis of 66 studies found that a single bout of exercise had small beneficial effects on total sleep time, sleep onset latency, sleep efficiency, stage 1 sleep and slow-wave sleep, a moderate beneficial effect on wake after sleep onset, and a small effect on REM sleep (Kredlow et al., 2015). For evening sessions specifically, a meta-analysis of 23 studies found that exercise raised slow-wave sleep by 1.3 percentage points, cut stage 1 by 0.9 points and delayed the first REM period by 7.7 minutes (Stutz et al., 2019). Vigorous exercise ending within an hour of bedtime may lengthen the time to fall asleep and reduce total sleep and sleep efficiency. On a 7 hour night, 1.3 percentage points is about 5 minutes of deep sleep. That is far smaller than a watch's error for Deep, so a hard training day will rarely show up as a visible jump in your Deep minutes.
Alcohol changes the night more, and in a recognizable pattern. At every dose it shortens the time to fall asleep, consolidates the first half of the night and disrupts the second half. It delays the first REM period at every dose, reduces total REM at moderate and high doses, and increases slow-wave sleep in the first half (Ebrahim et al., 2013). A 2025 meta-analysis of 27 studies found REM loss beginning at a low dose, 0.50 g/kg or less, which is about two standard drinks, and growing with each higher dose (Gardiner et al., 2025). A faster sleep onset and earlier deep sleep appeared only at high doses of 0.85 g/kg or more, about five drinks. On a sleep report, a night after drinking tends to look deep-heavy early, short on REM and restless toward morning.
08How accurate Apple Watch, Oura and WHOOP are
Every consumer tracker detects sleep well and separates the stages less well. Researchers usually report stage agreement as Cohen's kappa, which corrects for agreement expected by chance. Lee and colleagues (2023) grade 0.2 to 0.4 as fair agreement and 0.4 to 0.6 as moderate. Per-stage accuracy scores each stage against all the others, so it credits correct rejections and runs much higher than kappa. The two statistics cannot be compared directly.
| Device and algorithm | Study | Sample | Four-stage agreement | Deep sleep finding |
|---|---|---|---|---|
| Apple Watch, watchOS 9 model | Apple (2025), manufacturer | 166 adults, 299 nights | Kappa 0.63 | Most common error was Deep called Core |
| Apple Watch, watchOS 26 model | Apple (2025), manufacturer | Same validation set | Kappa 0.68 | Not reported in text |
| Apple Watch Series 8 | Schyvens et al. (2025) | 20 adults | Kappa 0.53 | 50.7% of N3 found, 25.2 min underestimate |
| Apple Watch 8 | Lee et al. (2023) | 26 adults | Kappa 0.30 | Not reported as minutes |
| Oura Ring Gen3, OSSA 2.0 | Svensson et al. (2024) | 96 adults, up to 3 nights | Per-stage accuracy 75.5% to 90.6% | Deep time not significantly different |
| Oura Ring, research model | Altini and Kinnunen (2021), Oura-affiliated | 106 people, 440 nights | 79% four-stage accuracy | Not reported as minutes |
| Oura Ring 3 | Lee et al. (2023) | 53 adults | Kappa 0.35 | Not reported as minutes |
| WHOOP 4.0 | Schyvens et al. (2025) | 40 adults | Kappa 0.37 | 69.6% of N3 found, 31.5 min overestimate |
| WHOOP | Miller et al. (2021) | 6 adults, 54 nights | 63% four-stage agreement, auto-detect | 15.5 min underestimate of slow wave sleep |
Deep sleep is where every device struggles. N2 and N3 both involve stillness and slow, regular breathing, and the difference between them lies mostly in brain wave amplitude, which a wrist sensor cannot see. Deep is also a small share of the night, so an uncertain classifier tends to default to the majority class. Apple reports that its most common error overall was true Deep sleep labeled Core, and that every stage, including wake, was most often misread as Core (Apple, 2025). In the Schyvens lab, the Series 8 labeled 47.7% of true N3 epochs as light sleep and underestimated REM by 13.4 minutes. WHOOP 4.0 in the same lab overestimated both Deep and REM.
The studies disagree with one another as much as the devices do. Apple Watch ranked first of six devices in one lab (Schyvens et al., 2025) and lowest of five wearables in another (Lee et al., 2023), which recruited its 75 participants from a tertiary hospital and a sleep clinic. Apple states in its paper that watch sleep stages are not intended for clinical use. Schyvens and colleagues concluded that the better devices can track prolonged, substantial changes in sleep architecture, which is the use the evidence supports. The full comparison, including the watchOS 26 update and the studies commonly miscited as Apple Watch evidence, is in how accurate Apple Watch sleep tracking is. For the general mechanics of consumer sleep tracking, see the glossary entry.
In practice, total sleep time is the most reliable number on your report and single-night stage minutes are the least reliable. A deep sleep figure 25 minutes below your usual on one night sits inside the Apple Watch's measured error. The same drop held for two weeks on the same device is worth attention, because comparing a device with itself cancels whatever part of its error repeats every night.
09How Titan uses sleep stages
Titan's Sleep score uses stages in one of its three parts. The score combines duration against your Sleep Goal (weight 0.30), restorative sleep (0.24) and awake time as a share of time in bed (0.14). Titan divides the weighted sum by 0.68, so duration carries 44% of the score, restorative sleep 35% and awake time 21%. It then raises the result to the power 0.9.
Restorative sleep is Deep plus REM as a share of time asleep, and it has two pieces. The larger piece, 82% of the part, compares that share with a 40% reference. It raises the ratio to the power 0.9 and gives full credit at 40% or more. The remaining 18% rewards an even split between Deep and REM. Equal minutes of each earn full balance credit, and a night with REM and no Deep earns none. The Sleep screen shows the same Deep plus REM share as your restorative sleep percentage.
A worked example shows how much stage error can move the score. Take an 8 hour Sleep Goal and a night with 7 hours asleep, 7.5 hours in bed and 30 minutes awake. With Apple's validation-set average of 12.8% Deep and 21.4% REM, restorative sleep earns about 0.85 and the night scores 86. Move 25 minutes of that Deep into Core, the average deep sleep shortfall Schyvens measured for the Series 8, and restorative sleep falls to about 0.69. The score drops to 81. The whole 5 point difference comes from classification error on a night that never changed.
Restorative sleep comes from stage labels, so it depends on a source that records Core, Deep and REM. The Sleep Tracker setting sets which source Titan builds each night from, as choosing your sleep source describes.
Titan's Recovery score does not use stage minutes. It is built on overnight HRV and sleeping heart rate, so a night of low recorded Deep sleep lowers your Sleep score without changing Recovery directly. The sleep debt guide covers how Titan tracks the hours side of the same nights.
10References
- AASM (2023). AASM releases updated version of scoring manual. American Academy of Sleep Medicine. https://aasm.org/aasm-releases-updated-version-scoring-manual/
- Altini M, Kinnunen H (2021). The promise of sleep: a multi-sensor approach for accurate sleep stage detection using the Oura ring. Sensors 21(13):4302. https://doi.org/10.3390/s21134302
- Apple Inc. (2025). Estimating sleep stages from Apple Watch, updated October 2025. Apple technical paper. https://www.apple.com/health/pdf/EstimatingSleepStagesfromAppleWatchOct_2025.pdf
- Berry RB et al. (2012). The AASM manual for the scoring of sleep and associated events: rules, terminology and technical specifications, version 2.0. American Academy of Sleep Medicine, Darien, Illinois.
- Boulos MI et al. (2019). Normal polysomnography parameters in healthy adults: a systematic review and meta-analysis. Lancet Respiratory Medicine 7(6):533-543. https://doi.org/10.1016/S2213-2600(19)30057-8
- Cleveland Clinic (2023). Sleep. Cleveland Clinic Health Library. https://my.clevelandclinic.org/health/articles/12148-sleep-basics
- Ebrahim IO et al. (2013). Alcohol and sleep I: effects on normal sleep. Alcoholism: Clinical and Experimental Research 37:539-549. https://doi.org/10.1111/acer.12006
- Gardiner C et al. (2025). The effect of alcohol on subsequent sleep in healthy adults: a systematic review and meta-analysis. Sleep Medicine Reviews 80:102030. https://doi.org/10.1016/j.smrv.2024.102030
- Kredlow MA et al. (2015). The effects of physical activity on sleep: a meta-analytic review. Journal of Behavioral Medicine 38:427-449. https://doi.org/10.1007/s10865-015-9617-6
- Lee T et al. (2023). Accuracy of 11 wearable, nearable, and airable consumer sleep trackers: prospective multicenter validation study. JMIR mHealth and uHealth 11:e50983. https://doi.org/10.2196/50983
- Miller DJ et al. (2021). A validation study of a commercial wearable device to automatically detect and estimate sleep. Biosensors 11(6):185. https://doi.org/10.3390/bios11060185
- Mitterling T et al. (2015). Sleep and respiration in 100 healthy Caucasian sleepers, a polysomnographic study according to American Academy of Sleep Medicine standards. Sleep 38(6):867-875. https://doi.org/10.5665/sleep.4730
- NINDS. Brain basics: understanding sleep. National Institute of Neurological Disorders and Stroke. https://www.ninds.nih.gov/health-information/public-education/brain-basics/brain-basics-understanding-sleep
- Ohayon MM et al. (2004). Meta-analysis of quantitative sleep parameters from childhood to old age in healthy individuals: developing normative sleep values across the human lifespan. Sleep 27(7):1255-1273. https://doi.org/10.1093/sleep/27.7.1255
- Schyvens AM et al. (2025). A performance validation of six commercial wrist-worn wearable sleep-tracking devices for sleep stage scoring compared to polysomnography. Sleep Advances 6(2):zpaf021. https://doi.org/10.1093/sleepadvances/zpaf021
- Stutz J et al. (2019). Effects of evening exercise on sleep in healthy participants: a systematic review and meta-analysis. Sports Medicine 49(2):269-287. https://doi.org/10.1007/s40279-018-1015-0
- Svensson T et al. (2024). Validity and reliability of the Oura Ring Generation 3 (Gen3) with Oura sleep staging algorithm 2.0 (OSSA 2.0) when compared to multi-night ambulatory polysomnography: a validation study of 96 participants and 421,045 epochs. Sleep Medicine 115:251-263. https://doi.org/10.1016/j.sleep.2024.01.020
- Van Cauter E et al. (2000). Age-related changes in slow wave sleep and REM sleep and relationship with growth hormone and cortisol levels in healthy men. JAMA 284(7):861-868. https://doi.org/10.1001/jama.284.7.861
- Watson NF et al. (2015). Recommended amount of sleep for a healthy adult: a joint consensus statement of the American Academy of Sleep Medicine and Sleep Research Society. Journal of Clinical Sleep Medicine 11(6):591-592. https://doi.org/10.5664/jcsm.4758
