How sleep tracking actually works

What the clinical standard actually measures

Polysomnography is the reference against which every sleep device is judged. It is not a better version of what your ring does — it is a different measurement entirely.

A sleep technician attaches electrodes and records at minimum three signals: an electroencephalogram measuring electrical activity at the scalp, an electrooculogram measuring eye movement, and an electromyogram measuring muscle tone, usually at the chin. Sleep stages are then scored in 30-second epochs against published criteria.

Those criteria are defined in terms of those signals. Deep sleep — N3 — is identified by slow delta waves occupying a threshold share of the epoch. REM is identified by a distinctive low-amplitude EEG pattern combined with rapid eye movements and near-total skeletal muscle atonia. The stages are not abstract states that happen to show up on an EEG. They are categories constructed from EEG, EOG and EMG.

That matters more than it first appears. When a wearable reports "deep sleep", it is not measuring a phenomenon that PSG also measures by another route. It is predicting what a technician scoring an EEG would have written down.

What your wearable actually collects

Strip away the app and a consumer sleep tracker has a short list of raw inputs:

• Movement, from an accelerometer. The oldest sleep-tracking signal there is. Stillness suggests sleep; movement suggests wake or a transition.
• Heart rate, from an optical sensor. Green or infrared light into the skin, measuring what scatters back off moving blood.
• Heart rate variability, derived from the same signal. Autonomic tone shifts across the night, and that shift carries real information about state.
• Respiratory rate, extracted from the modulation breathing imposes on the pulse waveform.
• Skin temperature, on devices that have a thermistor. Body temperature has a strong circadian rhythm.
• Blood oxygen estimates, on devices with a second wavelength.

Not one of these is an EEG signal. The device sees the peripheral consequences of what your brain is doing, several inferential steps removed from the thing being scored.

The inference, in the order it happens

1. Detect sleep onset. Sustained stillness plus a drop in heart rate below your daytime baseline. This is the easiest step and it is the one devices do well.
2. Segment the night. Split the record into short windows, conventionally 30 seconds to match PSG scoring epochs.
3. Extract features per window. Mean heart rate relative to your baseline, HRV, movement counts, respiratory rate, temperature trend, and how each is changing.
4. Classify. A machine-learning model trained on data from people who wore the device while undergoing PSG maps those features to a stage label.
5. Smooth. Raw classifier output is noisy, so it is constrained by expected sleep architecture — cycles of roughly 90 minutes, deep sleep concentrated early, REM concentrated late.

Step five is worth pausing on. Some of the structure you see in your hypnogram is coming from the model's prior expectation about how a night is shaped, not from anything measured about your night. That is a reasonable engineering decision and it is also why two devices on the same person can produce two plausible-looking, quite different graphs.

The physiological logic, which is genuinely sound

It would be unfair to present this as guesswork. The inference rests on real relationships.

During deep sleep, parasympathetic activity dominates: heart rate falls to its nightly minimum, HRV rises, breathing is regular, and you are close to motionless. During REM, the autonomic picture becomes far more variable — heart rate and breathing become irregular while skeletal muscles are atonic, so you have variability without movement. Light sleep sits between the two, with more micro-movements and brief arousals.

Those are genuine, well-established physiological signatures. A model that watches heart rate, HRV, respiration and movement is watching things that really do differ by stage. The question is only how reliably the peripheral signature identifies the central state, on an individual, on a given night.

How well it works, measured

Robbins et al. (2024) at Brigham and Women's Hospital ran the comparison directly: 35 healthy adults, a single night of inpatient polysomnography, wearing an Oura Ring Gen3, a Fitbit Sense 2 and an Apple Watch Series 8.

Sleep versus wake was good. 91-93% agreement across all three devices. Your total sleep time is broadly trustworthy — which makes sense, because that classification leans hardest on movement and heart-rate level, the two things a wearable genuinely measures.

Four-stage classification was substantially weaker.76.3% for Oura, 75.0% for Apple, 70.9% for Fitbit, with Cohen's kappa of 0.65, 0.60 and 0.55 respectively.

Deep sleep specifically was the weak point. Sensitivity ranged from 79.5% (Oura) to 61.7% (Fitbit) to 50.5% (Apple Watch). The Apple Watch overestimated light sleep by 45 minutes a night and underestimated deep sleep by 43.

Read those numbers as a picture of the method rather than a scoreboard. The step the devices do well is the one supported by what they actually measure. The step they do badly is the one that requires reconstructing a brain state from a pulse. The full device-by-device table is here.

Why placement changes the answer

The best-performing device in that study was a ring, and there is a mechanical reason for it. A finger is a better optical measurement site than a wrist: the tissue is well perfused, the arteries sit close to the surface, and a ring holds a fixed position rather than sliding on a moving joint.

The same principle shows up in the heart-rate literature. The 2025 Sensors placement study ran three identical sensors at wrist, forearm and upper arm simultaneously and found proximal placements beat the wrist at every intensity tested. Better raw signal produces better features, which produces better classification — the chain is the same one.

It is also why an upper-arm night band exists as a product category: it puts the sensor somewhere better and solves the other problem below.

The failure mode nobody talks about

A large share of complaints that begin "my sleep data is wrong" turn out to be "my sleep data is missing". The watch was on the charger at bedtime.

This is not a trivial point. Every one of these devices produces a baseline — your normal resting heart rate, your normal HRV, your normal sleep duration — and a baseline built from four nights a week is mostly noise. A device with a two-day battery that you charge before bed is a device that will never tell you anything reliable, regardless of how good its classifier is.

Which is why battery life is a sleep-tracking specification, not a convenience feature — and why a 14-day watch can beat a better sensor you take off every night.

What no consumer device can do

Score your stages the way a lab does. That requires electrodes on your scalp. Everything else is prediction, and the best prediction available agrees with the clinical standard about three quarters of the time on four-stage classification.

Diagnose anything. Some devices advertise sleep-apnea-related features. No consumer wearable is a diagnostic instrument, and we found no independent validation of those claims. Persistent unrefreshing sleep, loud snoring or witnessed breathing pauses are reasons to see a doctor, not reasons to buy a different tracker.

Tell you why. A device can show you that last night differed from your normal. It cannot tell you whether that was the late meal, the alcohol, the argument or the training.

How to read your own data sensibly

• Trust total sleep time. 91-93% sleep-wake agreement is genuinely good and it is the number most worth acting on.
• Treat the stage breakdown as indicative. Particularly deep sleep, where measured sensitivity ranged from 50.5% to 79.5% between devices.
• Read weeks, not nights. Every device is far more informative about trend than about any single night, and so is your own subjective impression.
• Do not switch devices mid-baseline. Different sensor, different classifier, different smoothing. The step change is the device, not you.
• Wear it every night. The most common cause of useless sleep data is missing sleep data.