Sleep stages explained: light, deep, REM, and what wearables guess

If you have worn a sleep tracker for more than a few weeks, you have seen the morning chart: a stacked band of colors labelled awake, light, deep, REM, and a single number out of 100 that is supposed to summarize the night. The chart looks confident. The science underneath it is less confident than the chart.

Here is what the stages actually are, what they appear to be for, how a sleep lab scores them, and what a consumer wearable is doing when it gives you the same labels without ever touching your scalp.

Where the labels come from

The labels you see on your tracker — light, deep, REM, awake — come from a clinical scoring system maintained by the American Academy of Sleep Medicine (AASM). The current rulebook, the AASM Manual for the Scoring of Sleep and Associated Events, was most recently revised in 2023. The original system was Rechtschaffen and Kales, 1968, which split non-REM sleep into stages 1 through 4; the AASM consolidated stages 3 and 4 into a single N3 in 2007. The modern scheme:

• N1: the lightest non-REM stage; the drift from wake into sleep.
• N2: still light non-REM, but unambiguously asleep; the dominant stage of a typical night.
• N3: deep, slow-wave non-REM sleep.
• REM: rapid eye movement sleep; the vivid-dreaming stage.
• Wake: scored separately whenever the criteria for sleep aren’t met.

These stages are defined by what the brain is doing, not by what the heart or the limbs are doing. The clinical rules score every 30-second window of a night based on the EEG pattern, with additional input from eye movement (electrooculography, EOG) and chin muscle tone (electromyography, EMG). That triad of EEG, EOG and EMG is the gold standard, and it is what a polysomnography (PSG) study in a sleep lab actually measures, alongside breathing, oxygen saturation, and ECG.

No consumer wearable measures any of this. They measure things that correlate with what the brain is doing (heart rate, heart-rate variability, motion, and on newer devices skin temperature and breathing rate) and they fit a model that predicts stages from those correlates. Some of the correlations are quite tight (REM has a recognizable autonomic fingerprint). Some are loose.

What happens in your body during each stage

A short tour of the physiology, because the stage names by themselves don’t mean much without the body signs.

N1, the doze

N1 is the transition stage. It typically lasts 1 to 7 minutes at sleep onset and accounts for under 5 percent of total sleep in healthy adults. EEG shows the alpha rhythm of relaxed wakefulness fading and slower theta waves appearing. Eye movements become slow and rolling. Muscle tone drops a little. If you wake someone out of N1 they will often insist they were still awake. This is the “I wasn’t asleep, I was just resting my eyes” stage.

Heart rate drifts downward. HRV begins to rise (we have a longer treatment of what HRV is and how it changes with state). Breathing slows. There is sometimes a hypnic jerk, the whole-body twitch most people have experienced at sleep onset.

N2, the bulk of your night

N2 is where you spend the majority of your sleep, somewhere around 45-55 percent of total time in a typical adult. EEG shows two characteristic features: sleep spindles, short bursts of 11-16 Hz activity lasting half a second to two seconds, and K-complexes, large sharp deflections that can be triggered by external sounds. The current best guess is that spindles are involved in memory consolidation and in keeping you asleep through minor disturbances; the K-complex is a “stay asleep” response.

Heart rate is lower, HRV is higher, breathing is regular. Eye movements are essentially absent. Core temperature is dropping; skin temperature is rising as peripheral vasodilation sheds heat.

N3, deep slow-wave sleep

N3 is what your tracker calls “deep sleep.” The EEG signature is unmistakable: large, slow delta waves at 0.5-4 Hz dominating the trace, with amplitudes greater than 75 microvolts. This is the stage that is hardest to wake out of. Someone shaken out of N3 will be confused for a minute or two, the phenomenon called sleep inertia.

In the body: heart rate is at its slowest of the night, HRV is at its highest, breathing is deep and very regular. Skin temperature is at its peak; core temperature is at its nightly low. Growth hormone is released in pulses, most strongly during the first deep period of the night. Blood pressure dips by roughly 10-20 percent in healthy adults, the so-called “nocturnal dip” that clinicians watch for in hypertension work.

The hot topic in N3 research is glymphatic clearance, the idea (first proposed by Maiken Nedergaard’s group at Rochester in Science, 2013) that the brain uses sleep, and deep sleep in particular, to flush metabolic waste products through cerebrospinal-fluid channels around the blood vessels. The mouse work is convincing. The human work is still arriving and the picture is messier than the popular-science version of the story suggests. We treat it as plausible and developing, not settled.

REM, paradoxical sleep

REM is the strangest stage. EEG looks almost identical to wake: low-voltage, fast, mixed-frequency activity. The eyes are moving rapidly under closed lids — that is what gives the stage its name. Yet skeletal muscle tone is essentially abolished from the neck down. This is REM atonia, a protective mechanism; we have it because if we did not, we would act out our dreams. People with REM sleep behavior disorder lack the atonia, often as an early sign of Parkinson’s or related synucleinopathies.

In the body during REM: heart rate is variable and often elevated, sometimes near daytime values, breathing is irregular, blood pressure is irregular, and thermoregulation is partly suspended. HRV behavior in REM is distinctive enough that it is the single most useful signal a consumer wearable has to work with.

The functional role of REM is the contested part. Memory consolidation, particularly of procedural and emotional memory, is the most-supported hypothesis (Robert Stickgold at Harvard has been the most prominent advocate). Matthew Walker at Berkeley has argued REM is also important for emotional regulation. There are people with very little REM (some antidepressants suppress REM almost entirely) who appear cognitively intact. We know REM matters; we are still learning exactly for what.

A typical night, end to end

Sleep cycles run roughly every 90 minutes in adults, with real variation (the range is more like 70 to 120 minutes person to person). A typical 7.5-8 hour night gives you four to six cycles. The architecture is asymmetric in two important ways:

• Deep sleep front-loads. The first cycle has the largest N3 block, often 30-60 minutes. By cycle three N3 is much shorter; by cycles four and five it may be absent. This is why early-evening alcohol or an inconsistent bedtime hits your deep sleep specifically.
• REM back-loads. The first REM period is short, often under 10 minutes, and may not appear until 70-90 minutes after sleep onset. REM periods get longer through the night, and the last cycle before waking can have a REM block of 30-45 minutes. Cutting your sleep short by an hour disproportionately costs you REM.

Brief awakenings between cycles, around four to six per night for a typical adult, are completely normal. You almost never remember them. They get scored as “wake after sleep onset” (WASO) on a PSG report.

Age changes this picture a lot. Newborns spend close to 50 percent of sleep in REM; adults spend around 20-25 percent. Deep sleep declines steadily from the 20s onward, and by the 60s and 70s many otherwise healthy adults get very little N3 at all. The “your deep sleep is 8 percent” alert on a 65-year-old’s tracker may be telling them only that they are 65.

How a sleep lab actually measures all this

The gold standard is in-lab polysomnography. A technician glues electrodes to your scalp (typically 6 EEG channels following the international 10-20 system), to the outer canthi of your eyes (EOG), under your chin (submental EMG), and to your legs (tibial EMG). A nasal cannula and thermistor measure airflow. Belts around your chest and abdomen measure respiratory effort. A finger pulse oximeter measures oxygen saturation. ECG leads measure heart activity. A technician watches the data on monitors next door.

In the morning the data is scored by a registered sleep technologist in 30-second windows (“epochs”) against the AASM rules. The whole thing is expensive (often $1,500-$3,000 in the US, though insurance covers most clinical indications) and disruptive, which is why home sleep tests have largely replaced PSG for uncomplicated sleep apnea work-ups. Home tests give you the apnea-hypopnea index but they don’t score full sleep stages; they typically don’t have EEG.

No consumer device replicates this. The Muse S headband has EEG and produces stage estimates that are arguably the most defensible in the consumer category, but it is awkward to sleep in. Everything else is inferring stages from peripheral signals.

What consumer wearables actually do

Consumer sleep trackers measure proxy signals very well and then use a model to guess at stages. The proxies vary by device, but the menu is roughly:

• Motion, from a 3-axis accelerometer. The original sleep-tracking signal, pre-2014. Still useful for wake-versus-asleep and sleep-onset latency. Useless on its own for distinguishing non-wake stages.
• Heart rate, from a green-LED photoplethysmography (PPG) sensor on the skin.
• Heart-rate variability, derived from beat-to-beat intervals in the PPG signal. The single most informative signal for stage classification, because REM has a distinctive autonomic pattern.
• Skin temperature, from a thermistor in contact with the wrist or finger. Useful for tracking circadian phase and for catching anomalies. Our primer on skin-temperature wearables goes into where the signal does and does not help.
• Respiratory rate, derived from modulation of the PPG signal by breathing.
• SpO2, blood oxygen saturation, from a red-and-infrared LED. Mostly used for breathing-disturbance flags, less for staging.

The device’s algorithm takes a sliding window of these signals and assigns a probable stage to each minute or 30-second epoch. Algorithms are proprietary, sometimes ML-based, sometimes rule-based. Manufacturers tune them against PSG in development but few publish detailed validation. The ones that do (Fitbit, Oura, and Whoop have all funded peer-reviewed work) show meaningful improvements over motion-only methods, but they are not close to PSG accuracy on stage classification.

The cleanest independent benchmark is Chinoy et al., “Performance of seven consumer sleep-tracking devices compared with polysomnography,” Sleep, 2021 (Volume 44, Issue 5). The authors ran seven popular consumer devices against simultaneous in-lab PSG in healthy adults. Total sleep time agreement was reasonable across most devices, typically within 10-15 minutes of the PSG measurement. Wake detection was acceptable but not great. Stage classification was poor, particularly for REM and deep sleep specifically. Device-by-device sensitivity for detecting deep sleep ranged from roughly 35 to 65 percent. REM was in a similar range. That is the empirical floor we are operating against.

Two notes on the table. Under-mattress and smart-pillow products (Withings Sleep Analyzer, Eight Sleep Pod) are interesting because they ask nothing of you; you just sleep. But their stage outputs are the least validated in the category. Withings Sleep Analyzer has the most clinical-adjacent backing, mostly for its breathing-disturbance index, which has CE-mark approval for sleep apnea detection in Europe. None of them are scoring REM or N3 anywhere close to a PSG.

Ring versus wrist also matters for underlying signal quality, not just comfort. The finger has a denser bed of arterioles than the wrist, which gives the PPG sensor a cleaner pulse waveform at rest, and a ring is less likely to be displaced under a pillow or compressed against the mattress. That is a real signal-quality advantage, treated more fully in our wearable sensor accuracy primer and in Oura Ring vs Whoop.

What the numbers on your dashboard are actually good for

Here is the calibrated way to read your tracker in the morning. Trust:

• Total sleep time and bedtime / wake time. The most reliable outputs of any modern consumer wearable. If your tracker says you slept 6 hours 42 minutes, the real number is very likely within 10-15 minutes of that. Enough to manage a sleep schedule by.
• Resting heart rate and HRV trends over weeks. A single night is noisy; a four-week trend is real signal. A sustained rise in resting heart rate, or a drop in HRV, is one of the most useful nonspecific signs of accumulating stress, training load, or oncoming illness (detail in our resting heart rate explainer).
• Breathing rate trends. Same logic: single nights noisy, multi-week trends meaningful.
• The anomaly flag. If your tracker says last night was unusually bad and you know why, the flag is doing its job. If it says last night was unusually bad and you don’t know why, that’s worth a moment of attention.

Distrust, or heavily discount:

• The single-night stage breakdown. The part of the dashboard with the weakest empirical backing. If your wearable says you got 22 minutes of deep sleep, the PSG-equivalent number could plausibly be 10, or 50.
• The sleep score. Every manufacturer’s sleep score is a proprietary composite that bundles total sleep time (reliable) with stage estimates (not reliable) into a number out of 100.
• Day-to-day deep-sleep comparisons. The night-to-night noise on deep-sleep estimation often exceeds the difference between any two given nights.

A practical test: ignore the dashboard for a week. Don’t open the app. Keep wearing the device. At the end of the week, ask yourself what you would have done differently if you had been checking the numbers daily. For most people the answer is “nothing.” The recommendations a sleep tracker generates — go to bed earlier, drink less, exercise earlier — are things you already know. The actionable information is in the long trends, not in this morning’s score.

One historical footnote

The first consumer wrist device to try multi-signal sleep staging, using heart rate, skin temperature, perspiration, and motion in parallel rather than motion alone, was the Basis Peak in late 2014. Intel bought Basis Science in March 2014, recalled the Peak in June 2016 over a battery-overheating issue (formalized as CPSC notice 16-235 that August), and shut the Basis cloud down at the end of that year. The Peak’s working sleep model is gone; you cannot revive a unit. But the architecture it shipped is now the category norm. Oura, Whoop, Fitbit’s Sleep Profile, Apple’s watchOS sleep staging, and the Pixel Watch sleep model are all doing some version of what Basis did first. We covered the Peak’s specific approach to sleep tracking separately.

The gap between what these devices measure and what they claim to know has been the central tension of the category for over a decade. Reading the stage chart in the morning with that tension in mind makes the dashboard more useful, not less. You stop expecting it to be a polysomnography. You start using it for the things it is actually good at.