Sleep stages: How the Ultrahuman Ring measures Light, Deep, REM, and Awake

What are sleep stages?

Sleep stages are repeating cycles of brain and body states that structure every night.

A healthy adult's night moves through four to six 90-minute cycles, each containing NREM (non-rapid eye movement) and REM phases. Clinicians identify them via polysomnography (PSG), which records brain-wave activity with scalp electrodes alongside eye movements, muscle tone, and other signals.

Light sleep covers entry stages N1 and N2, when you are easily roused, and brain activity begins to slow.

Deep sleep (N3, also called slow-wave sleep) is when the brain generates slow, high-amplitude delta waves, heart rate and blood pressure fall to their daily minimum, and the body carries out physical repair.

REM sleep, which concentrates in the second half of the night, is when dreaming is most vivid and when the brain consolidates procedural and emotional memories.

Brief Awake periods interrupt this pattern, becoming longer and more frequent with age.

How does Ultrahuman measure sleep stages?

The Ring reads three streams of physiological data from the finger throughout the night and uses a machine learning model to infer which sleep stage best matches each successive window.

The Ring's green-LED and infrared photoplethysmography (PPG) sensor tracks your heart rate and the beat-to-beat intervals that define heart rate variability (HRV) while you sleep.

Autonomic nervous system activity shifts markedly across sleep stages: parasympathetic (rest-and-digest) activity rises during Deep sleep, suppressing heart rate and widening beat intervals; REM brings a more variable pattern with episodic sympathetic bursts; and Awake periods show elevated heart rate and irregular variability as the body responds to activity.

A 6-axis motion sensor (3-axis accelerometer plus 3-axis gyroscope) tracks body movement: Deep and REM sleep are both low-movement states, while Awake moments produce characteristic motion bursts.

An NTC thermistor at the skin surface tracks relative skin temperature. At the finger, distal skin temperature shows a momentary rise at sleep onset as blood vessels in the extremities dilate to offload heat, which is part of what drives the core-body-temperature fall that accompanies falling asleep; across the rest of the night, the overall trend is a gradual decline.

These three streams feed a per-epoch classifier that assigns one of four labels (Awake, Light, Deep, REM) to each successive window. The classifier is an on-device machine-learning model that processes sequences of epochs together rather than each epoch in isolation, allowing it to pick up on gradual stage transitions and the 90-minute cycle structure of normal sleep. Each epoch's label then feeds the stage-duration totals and contributors that appear in the Ultrahuman app.

Heart rate, movement, and relative skin temperature are direct measurements.

Sleep stage is a statistical inference from those measurements, not a direct observation of brain activity.

A finger ring cannot carry scalp electrodes, which is the trade that lets it monitor sleep continuously every night in a device members already wear.

However, research using raw accelerometry and PPG heart rate from a consumer wearable has shown that combining these signals classifies wake, NREM, and REM sleep at about 72% accuracy against PSG, well enough to confirm that the heart-rate signal adds real information on top of movement alone, which is why the Ring samples all three streams together.

Why did Ultrahuman build it this way?

The core design decision is the sensor set: PPG plus accelerometer plus temperature, rather than EEG. This is a physical constraint, not a preference. A finger ring cannot carry scalp electrodes. The consequence is a 4-stage estimate from peripheral signals rather than the direct measurement of cortical activity that a PSG lab provides.

That tradeoff has a documented benefit. Published work on ring-form-factor staging shows that 4-stage accuracy lifts substantially when you add autonomic-nervous-system signals (derived from PPG) to accelerometry: movement alone reached 57% in that study, and adding ANS and circadian features raised it to 79%.

Each sensor stream adds independent information. Heart rate and HRV distinguish Deep from Light in a way that motion alone cannot, because both stages look similar on an accelerometer (both are low-movement). Temperature provides a slower-changing signal that helps anchor stage transitions to the night's thermal arc.

The on-device model processes sleep as a sequence, which lets it capture that deep sleep tends to concentrate in the first third of the night and REM in the last third, and that transitions between stages follow patterns rather than occurring randomly. Reading the night as a sequence classifies each window more accurately than treating it in isolation.

What does the research say?

The physiological signals the Ring reads are well-established markers of sleep stages. Overnight HRV analysis during PSG-scored sleep shows that high-frequency HRV, a measure of parasympathetic activity, differs measurably across the N1, N2, N3, and REM stages in an older-adult cohort. The autonomic signatures the Ring's classifier reads are grounded in physiology.

On staging accuracy, the published literature is consistent for 4-class classification. A study using PPG from a wrist-worn device with a deep learning model reached 78.7% overall 4-class accuracy (Cohen's kappa 0.68) against PSG-scored labels. Sequence-to-sequence modeling over actigraphy and coarse heart rate, tested across the MESA (n=808) and MrOS (n=817) epidemiological cohorts, reached 70-72%, and showed that deep sleep duration can be estimated reliably even though it is the least represented stage in a typical overnight recording.

Core body temperature drops in the hour before sleep onset, and the rate of that drop correlates with how well nocturnal parasympathetic activity (measured via HRV) develops across the night. That study measured core temperature; the Ring measures relative skin temperature at the finger, which is a peripheral proxy for the same thermoregulatory process, since the distal vasodilation that warms the extremities is what drives the core-temperature fall. That is why the temperature sensor earns its place in staging: the skin temperature arc across the night reflects the same thermoregulatory processes that accompany the autonomic changes between stages.

Across the literature, the consistent picture is that 4-stage classification from wearable peripheral sensors is a meaningful estimate of sleep architecture, particularly for tracking direction and magnitude of change across nights, while carrying more uncertainty on any single epoch than a PSG recording would.

What does this look like across Ultrahuman Ring members?

Across 351,106 Ultrahuman Ring members aged 20 to 69 (last six months of data, valid sleep sessions only, internal team members excluded), both deep sleep and REM sleep decline gradually with age, while time spent awake at night rises.

In the 20-29 age band (n=137,581 members), deep sleep averaged 77.8 minutes per night (16.8% of total sleep time) and REM averaged 112.4 minutes (23.7%).

In the 50 and above band (n=47,608 members), deep sleep averaged 67.6 minutes (14.8%) and REM averaged 97.0 minutes (20.9%). Awake time rose from 39.3 minutes per night in the youngest band to 54.2 minutes in the oldest. Total sleep duration was broadly stable across age bands, ranging from 465 to 482 minutes, suggesting the shift is in composition rather than in overall time in bed.

These numbers come from the on-device classifier's stage assignments. They track how stage distribution changes across nights and across the membership, and they read as physiological estimates rather than lab-scored measurements. That is what the Ring's sensors deliver: a consistent, repeatable picture of sleep architecture every night, in a device worn at home.

How does it show up in the Ultrahuman app?

Sleep stage data appears in the Sleep tab of the Ultrahuman app each morning after a Ring sync. The main view shows a hypnogram: a timeline graphic that plots Awake, Light, Deep, and REM across the night, so you can see when each stage occurred, how long it lasted, and how it fits into the overall sleep cycle structure. Above the hypnogram, the app shows duration totals for each stage.

Deep sleep and REM sleep each contribute to the Sleep Index through their own contributor inputs alongside sleep efficiency, timing, and total duration. Unusually short deep or REM windows typically shift the Sleep Index lower; unusually long stretches of either tend to lift it. Awake time contributes through the awake-time contributor and restfulness factor. For the full picture of how these stage inputs roll up, see how the Ring builds your Sleep Index and how the same stage totals feed how deep sleep drives restorative sleep.

Stage distribution sits alongside how total sleep duration is measured across the night, so a short night and a stage-poor night are tracked as separate signals. Members who want to understand how sleep stages connect to recovery can cross-reference nightly stage data with their Recovery Score, which draws on how overnight HRV is measured from the Ring separately from the staging algorithm. A night that shows low deep sleep in the Sleep tab will often correspond to a muted overnight HRV rise in the Recovery view.

For members interested in how specific lifestyle choices affect stage distribution, the Ultrahuman app's smart-tags feature lets you tag events (alcohol, caffeine, exercise, stress) against your nightly data, so you can observe how your own stage pattern shifts on nights when you logged those events. If you want to act on what the stages show, there are practical ways to increase your deep sleep, and you can read about how to time naps around your sleep stages without disrupting your overnight architecture, drawing on Ultrahuman's own research on naps and non-sleep deep rest. The night-to-night consistency of your sleep timing is itself a tracked signal, building on the sleep-regularity research in the literature; you can browse the full body of validation work in the Ultrahuman Studies hub.