Sleep Tracking: How Wearables Estimate Your Sleep and How Accurate They Really Are

Short Answer

Sleep tracking is an estimate of how long and how well you slept, built from signals your device can actually measure: movement, heart rate, skin-level circulation signals, and — on some wearables — blood oxygen, breathing, or temperature. A sleep score compresses those signals into one number so you can compare one night with another at a glance. That number can be useful, but it is not the same thing as a lab sleep study.

The strongest use case is simple: most modern wearables are better at telling sleep vs. wake than they are at naming the exact stage of sleep. In a polysomnography (PSG) validation study of Oura Ring Gen3, Fitbit Sense 2, and Apple Watch Series 8, sleep-vs-wake sensitivity was at least 95% for all three devices, while sleep-stage sensitivity was wider and less consistent — roughly 50–86% across stages and devices (Sensors, 2024). A larger multicenter validation of 11 consumer sleep trackers found the same kind of spread: epoch-by-epoch sleep-stage performance varied substantially, with macro F1 scores ranging from 0.26 to 0.69 (JMIR mHealth and uHealth, 2023). That is why a tracker can be directionally helpful and still be wrong about whether a block of the night was "deep" sleep or REM.

So read your tracker as a pattern-spotting tool, not a diagnosis. It is best for trends over weeks: whether your bedtime is drifting, whether alcohol or late stress shortens your night, whether travel shifts your rhythm, whether "8 hours in bed" is really closer to 6.5 hours asleep. It is weaker as a single-night judge of how your brain recovered. This is especially important because consumer devices infer stages from body signals, while polysomnography scores sleep using brain activity, eye movement, muscle tone, heart rhythm, and other physiological channels (Sensors, 2024).

That gap is why a 2026 living systematic review and meta-analysis of Apple Watch measurements reports that "Accuracy for sleep and step count was moderate" for the device (Apple Watch living systematic review, 2026). The takeaway is not "ignore your sleep score." It is: use the score as context. If your tracker says you slept well but you feel wired, foggy, or unrefreshed, your body is giving you data too — and sometimes that subjective signal matters more than the neat number on the screen.

Sleep tracking at a glance

Sleep tracking is not a mini sleep lab on your wrist. It's an estimate of when you slept, how long you slept, and how "smooth" the night looked, built from signals your device can actually see: movement, heart rate, blood oxygen, breathing or snoring, temperature, and sometimes phone or bedside-sensor data. That makes it useful for noticing patterns — for example, that late alcohol, stress, travel, or a shifted bedtime tends to break up your nights — but it is still consumer-grade tracking, not a medical measurement (Cleveland Clinic).

A sleep score turns that messy night into one simple number, usually 0–100 or 0–1, so you can compare one night with another. The catch is that the number is a device-made composite: it may blend duration, efficiency, sleep stages, movement, awakenings, and recovery-type metrics, but the recipe is usually proprietary. In one randomized controlled trial protocol, the primary outcome was described as "the sleep score (scale 1-100; composite device metric)" (JMIR Research Protocols, 2026). That wording matters: a higher score can mean "your device liked this night," not "your sleep was clinically normal."

Accuracy is strongest for the broad question — "was I asleep or awake?" — and weaker for the finer question — "was that REM, deep, or light sleep?" A 2026 living systematic review of Apple Watch measurements reported that "Accuracy for sleep and step count was moderate" (Apple Watch living systematic review, 2026), with good sleep-wake differentiation but poorer separation of physiologically similar sleep stages. A multicenter PSG comparison of 11 consumer sleep trackers found wide variation in sleep-stage performance, with macro F1 scores ranging from 0.26 to 0.69 (JMIR mHealth and uHealth, 2023); another validation study of Oura, Fitbit, and Apple Watch found all three had sleep-detection sensitivity of 95% or more, while concordance for deep and REM sleep was weaker (Sensors, 2024).

A tracker is also not the same thing as a sleep study. Polysomnography, or PSG, records brain waves, eye movements, muscle activity, heart rhythm, breathing, oxygen levels, and body movement so a clinician can diagnose or rule out sleep disorders. PSG is powerful, but it is also more intrusive: as one 2026 paper on non-intrusive alternatives puts it, conventional sleep studies require "multiple body-attached sensors, which are uncomfortable and impractical for routine use" (Frontiers in Network Physiology, 2026). A wearable is the opposite tradeoff: easy to repeat at home, good for trends, but limited by the signals it can collect (Cleveland Clinic).

A sleep tracker cannot diagnose a sleep disorder by itself. It can flag patterns worth checking — repeated oxygen dips, high estimated breathing disturbances, long awakenings, or a mismatch between "good" sleep scores and feeling unrefreshed — but diagnosis still belongs with a clinician and, when needed, formal testing. Even newer screening tools are validated "while highlighting the need for further optimization to accurately detect mild cases" (Real-World Smartwatch OSA Validation, 2026). That's especially important if you snore loudly, stop breathing in sleep, wake gasping, feel very sleepy during the day, or have heart, blood pressure, or neurologic symptoms.

So the best use is simple: don't judge your body by one night's score. Watch the pattern over weeks. If your tracker says "great" but you feel wrecked, believe that mismatch and look for context — stress, illness, caffeine, alcohol, menstrual cycle changes, late workouts, pain, bedtime drift, or possible sleep apnea. The number is a clue. Your lived sleep is data too.

How wearables estimate sleep

A consumer sleep tracker does not "see" sleep the way a lab does. In polysomnography — the gold-standard sleep study — sleep stages are scored from brain activity, eye movements, muscle activity, heart rhythm, breathing, oxygen levels, and other signals; each 30-second epoch is assigned a stage such as wake, N1, N2, N3, or REM. A wrist or ring sensor has a much narrower view of your body, so it has to infer what your brain is probably doing from signals it can measure at the skin (StatPearls / NCBI Bookshelf).

The first signal is movement. This is the actigraphy part: an accelerometer tracks how still or active your body is across the night. Long, quiet stretches tend to look like sleep; repeated movement tends to look like wakefulness or restless sleep. That is why movement is usually strongest for the basic sleep-versus-wake question, not for the fine details of sleep architecture. Clinical actigraphy is built around this same idea — recording limb movement over time and applying algorithms to estimate sleep and wake — but it generally cannot stage NREM and REM sleep on its own because that requires EEG, EOG, and EMG information (PMC6040804).

The second signal is your cardiovascular pattern. Most consumer wearables use optical photoplethysmography, or PPG, to estimate pulse through the skin, then derive heart rate and sometimes heart-rate variability. Your autonomic nervous system shifts across the night: heart rate and variability patterns are different in deeper NREM sleep than they are in REM or wake. A model can use those patterns, especially when paired with movement, to make a more educated guess about whether you are asleep and which broad stage you may be in (PMC7956647).

Some devices add oxygen and breathing-related signals. An SpO₂ sensor estimates blood oxygen, and the device may also estimate breathing rate or overnight breathing disturbances. These features are useful context for disrupted sleep and sleep-breathing patterns, but they still do not turn a wearable into a diagnostic sleep study. In practice, they are usually part of the device's "breathing disturbance" or recovery picture rather than the whole foundation of the sleep/wake call. See also measuring blood oxygen.

The algorithm then fuses these inputs window by window — commonly aligned to 30-second sleep-scoring epochs — and labels each slice of the night as awake, asleep, or a probable stage. More signals usually give the model more context: movement helps separate stillness from activity, heart-rate patterns add autonomic information, and breathing signals can flag physiology that movement alone may miss. That is why multi-signal systems tend to perform better than movement-only systems, especially when the goal is sleep staging rather than just sleep duration (StatPearls / NCBI Bookshelf).

But the number of stages matters. Collapsing sleep into fewer buckets is easier than reproducing the full lab-style breakdown. In one 2026 contactless model that combined heart and breathing signals, the authors reported "an accuracy of 80.51% (Cohen's κ = 0.65) for wake/non-REM/REM sleep stages compared with PSG under leave-one-subject-out cross-validation" (Frontiers in Network Physiology, 2026). In the same study, performance was lower when the model tried to classify all stages separately: 63.24% accuracy for wake, N1, N2, N3, and REM.

You see the same pattern in wearable research more broadly: devices are often good at telling "mostly asleep" from "probably awake," but less precise when they split the night into light, deep, and REM sleep. A healthy-adult validation study of Oura Ring Gen3, Fitbit Sense 2, and Apple Watch Series 8 found high epoch-by-epoch agreement with PSG for sleep versus wake, while four-stage agreement was lower and device-dependent, at roughly 70–79% across the three devices (Sensors, 2024).

So your tracker's sleep score is best read as a body-signal estimate, not a direct readout of your brain. It can be very useful for spotting patterns — later bedtimes, shorter sleep, more restless nights, possible breathing disruptions — especially across weeks. It is not the same thing as PSG, and it should not be used to diagnose insomnia, sleep apnea, REM behavior disorder, or any other sleep disorder on its own.

What a "sleep score" actually is

A sleep score is not a lab result. It is a device-made summary of your night: one number that tries to compress several signals into something you can read quickly. Depending on the tracker, that number is often shown on a 0–100 scale, while Welltory uses an internal 0–1 scale. Under the hood, the score may pull from how long you slept, how much of your time in bed was actually spent asleep, how your sleep stages were distributed, and how fragmented the night looked — the small awakenings, movement, and restlessness that make sleep feel less continuous. In a 2026 protocol for a pilot randomized controlled trial using Samsung Galaxy Watch 4 data, the primary outcome was "the sleep score (scale 1-100; composite device metric)," and the authors describe it as a proprietary score that combines sleep dimensions such as duration, architecture, efficiency, and fragmentation (JMIR Research Protocols, 2026).

That "composite" part matters. Your watch is not measuring sleep quality the way a sleep lab would. It is taking sensor signals, running them through a brand-specific algorithm, and producing a score. The exact weighting is usually not public. That means an 82 on one device is not the same thing as an 82 on another device, and neither should be treated as a medical scale. The Samsung trial protocol says this plainly: the composite score is proprietary, the exact calculation is not reproducible outside the manufacturer's algorithm, and the composite score itself has not been validated against polysomnography in the same way individual sleep parameters may be (JMIR Research Protocols, 2026).

So the best way to use a sleep score is personally, not competitively. Don't ask, "Is my number normal?" Ask, "Is this normal for me?" Welltory's sleep score is built on its own 0–1 internal scale, so it is most useful when you compare tonight with your own last few weeks — not with your partner, your friend, or another app.

A single night can be noisy. You may get a decent score after a night that felt awful, or a low score after a night when you feel surprisingly fine. The useful signal is the pattern: what happens to your score when you drink alcohol, train late, go to bed at irregular times, travel, get sick, or move through a stressful week. Over time, the score becomes less of a verdict and more of a feedback loop — a way to see which habits tend to make your sleep more stable, and which ones quietly break it up.

How accurate are sleep trackers? (the honest version)

Here's the nuance most headlines skip: accuracy depends entirely on what you're asking the tracker to do. A wearable is not reading your brain the way polysomnography does; it is inferring sleep from signals like movement, heart rate, photoplethysmography, temperature, and sometimes oxygen saturation. That works best when the question is broad: "Was my body probably asleep or awake?" It gets shakier when the question becomes microscopic: "Was this exact 30-second slice light sleep, deep sleep, or REM?"

For sleep vs. wake, modern wearables can be genuinely useful. In a 2024 validation study comparing Apple Watch Series 8, Fitbit Sense 2, and Oura Ring Gen3 with polysomnography, all three reached at least 95% sensitivity for detecting sleep (Sensors, 2024). That means if you were asleep, the device usually caught it. The catch is wake: lying still in bed can look a lot like sleep to a wrist sensor, because your body is quiet even when your brain is not.

For sleep stages, the honest answer is "moderate, and very device-dependent." Splitting sleep into light, deep, and REM is harder because these stages are defined by brain, eye, and muscle signals — the things PSG measures directly and most consumer wearables do not. A 2026 living systematic review and meta-analysis of Apple Watch accuracy concluded that "Accuracy for sleep and step count was moderate," and that "Measurement accuracy varied by metric, measurement conditions, and individual physiology" (Apple Watch living systematic review, 2026). In a multicenter validation of 11 consumer sleep trackers, epoch-by-epoch stage performance varied widely across devices, with macro F1 scores ranging from 0.26 to 0.69 (JMIR mHealth and uHealth, 2023). Other validation work has found that some consumer devices consistently overestimate deep sleep and underestimate REM, so the stage graph is better read as a rough pattern than a precise breakdown (systematic review and meta-analysis, 2026).

The simpler the category, the better the numbers tend to look. In the Apple/Fitbit/Oura validation study, sleep–wake classification was strong, but four-stage agreement dropped to roughly 70–79% across the three devices (Sensors, 2024). That is the core pattern: when the tracker only has to separate "sleep" from "wake," it performs well; when it has to label the architecture of sleep, the uncertainty grows.

Trackers can also over-count sleep on messy nights. If you have insomnia, wake up often, or spend long stretches lying still, the device may score quiet wakefulness as sleep. In a study of people with insomnia, both a research actigraph and Fitbit Alta HR overestimated total sleep time and sleep efficiency while underestimating sleep latency and wake after sleep onset (PMID 31626361). A broader review found the same direction of bias: actigraphy tends to overestimate total sleep time and sleep efficiency compared with PSG.

Sleep apnea adds another problem: the issue may be brief, repeated breathing disruption rather than long obvious awakenings. A 2026 real-world validation of a smartwatch apnea estimate found that "Bland-Altman analysis revealed systematic underestimation of actual AHI by the smartwatch, particularly in mild OSA" (Real-World Smartwatch OSA Validation, 2026). The authors supported its use as a screening aid, while "highlighting the need for further optimization to accurately detect mild cases" (Real-World Smartwatch OSA Validation, 2026). In plain English: a wearable may flag a strong pattern, but it can miss subtler disease — especially when the night looks "normal enough" from the outside.

So the takeaway is not "sleep trackers are wrong." It is more precise than that: they are useful for big-picture trends — did you sleep, roughly how long, how consistent was your schedule, did something change — and much fuzzier for fine-grained claims like your exact REM minutes or whether a mild sleep disorder is present. Read the trend. Don't treat the score like a diagnosis.

Tracker vs. sleep study (polysomnography)

A clinical sleep study — polysomnography (PSG) — is the reference test for answering medical sleep questions. It records several body systems at the same time: brain waves (EEG), eye movement, muscle activity, breathing, heart rhythm, airflow, and blood oxygen. That mix matters. Your brain waves and eye movements tell whether you're in REM, light sleep, or deep sleep; your breathing and oxygen show whether your airway or breathing control is failing; muscle signals can reveal movements or behaviors that a wrist sensor may never "see" (Johns Hopkins Medicine).

It's also, by design, a lot. As one 2026 paper on non-intrusive alternatives puts it, conventional sleep studies require "multiple body-attached sensors, which are uncomfortable and impractical for routine use" (Frontiers in Network Physiology, 2026). You usually do PSG once, in a lab or sometimes with a home sleep apnea test, when there's a specific question to answer — loud snoring and pauses in breathing, severe daytime sleepiness, unusual movements, suspected narcolepsy, or another disorder your clinician wants to confirm (MedlinePlus).

A wearable makes the opposite trade-off. It is less precise, but it is easy. It sits on your wrist or finger every night, for weeks or months, and turns movement, heart rate, and sometimes oxygen or breathing-related signals into estimates. That makes it useful for patterns: your bedtime drifted later, your sleep became more fragmented, your resting heart rate stayed higher after alcohol, travel, stress, or illness. But it cannot see your sleeping brain directly, so it should not be treated like a diagnosis. Recent validation and review data keep landing in the same place: wearables can give reasonable estimates for broad sleep metrics and long-term self-tracking, but their device-to-device variability limits them as stand-alone tools for detailed sleep-stage scoring or clinical diagnosis (JMIR mHealth and uHealth, 2023).

Think of it this way: a wearable is like a bathroom scale you step on every day; PSG is like a DEXA scan you get when something specific needs measuring. One gives you a trend you can live with. The other gives your clinician a much more exact answer.

When the score and the feeling disagree

Trackers give you a number. Your body gives you a feeling. The useful part starts when those two do not line up.

A sleep score is a compression. A wearable can summarize signals it can measure — movement, heart-rate patterns, sometimes oxygen-related signals, and device-estimated sleep stages — but it cannot fully see how your nervous system, stress load, illness, timing, medications, alcohol, pain, breathing, or repeated tiny awakenings felt from the inside. Research reviews of wearable sleep technology describe real limitations, including misclassification of wakefulness during the sleep period and uncertainty in people with comorbidities or sleep disorders (PMID 38149978). That is one reason a "good-looking" night on the app can still feel unrefreshing.

That does not make the score meaningless. It means the score answers a narrower question than "how rested am I?" Your lived experience of sleep can change before, after, or in a different direction than the metric you expected to explain it. The mismatch itself is information.

For everyday tracker use, the lesson is simple: if your score says "great" but you feel wrecked, do not argue with your body. Look wider. Check bedtime consistency, wake time, stress, late meals, alcohol, illness, pain, room temperature, and whether your sleep felt broken even if the app counted it as sleep. And if the pattern keeps repeating — especially with loud snoring, gasping, extreme daytime sleepiness, morning headaches, or safety problems like drowsy driving — a wearable is not the right tool to rule things out. Sleep studies are the tests clinicians use to help diagnose sleep disorders, including sleep apnea, narcolepsy, movement-related sleep disorders, and other causes of severe daytime tiredness (NHLBI). Consumer sleep technology is useful for patterns, but it is not a diagnostic device or a treatment plan (PMID 29734997).

Treat any device-estimated deep-sleep or REM percentage as a rough, device-dependent estimate, not a target, grade, or clinical norm. Welltory tracks and records your sleep data so you can watch your own pattern over time; it does not diagnose insomnia, sleep apnea, or any other sleep disorder. For related background, see sleep basics.

How to use a sleep tracker well

A tracker pays off when you use it for what it can actually see: repeated patterns in your nights, not the full truth of how restored you are. Treat the score like a weather report, not a verdict. One rough night means little. A steady drift — later bedtimes, lower efficiency, more wake-ups, shorter sleep after alcohol, stress, late caffeine, or late screens — is the signal worth noticing.

Pair the number with a quick "how do I feel?" check in the morning. That matters because subjective sleep and measured sleep can split apart in real people. In a 2025 study of nurses using PSQI questionnaires and Fitbit Charge 3 tracking, the researchers wrote that they "aimed to compare subjectively and objectively measured sleep quality in nurses" (Medicina, 2025). That is exactly the point for everyday tracking: your wearable shows one layer; your body reports another. Use both.

Use the tracker to run small experiments, not to grade yourself. If you stop caffeine earlier, keep the same wake time, dim screens, move your workout away from bedtime, or make the room cooler, watch what changes over the next stretch of nights. If the trend improves and you feel better, that is useful. If the score improves but you still wake up crushed, that is useful too — it tells you the tracker may be missing something your nervous system, mood, breathing, pain, schedule, or recovery load is still carrying.

When your score and your body disagree, do not automatically assume the device is right. The same nurse study found that "Most nurses subjectively rated their sleep as good or very good; however, according to the PSQI questionnaire results, all were classified as poor sleepers" (Medicina, 2025). Objective and subjective sleep can diverge because they measure different parts of the experience: movement, heart signals, timing, perceived rest, awakenings you remember, and next-day function do not always line up neatly.

Be especially careful with sleep-stage graphs. "Low deep sleep" or "not enough REM" on a wrist device is not a medical finding. Consumer trackers can be helpful for broad sleep/wake patterns, but validation studies consistently show weaker performance for wake detection and sleep staging than for basic sleep estimates; some reviews also emphasize that consumer sleep technologies are not a substitute for clinical diagnosis or polysomnography (PMC6908975). Use a stage pattern as a better question for a clinician, not as the answer.

Escalate the right signals. Persistent loud snoring, breathing pauses, gasping, choking, repeated very low overnight oxygen readings, insomnia that interferes with daily life, or crushing daytime sleepiness are reasons to talk to a healthcare professional — not reasons to buy a more expensive watch. Sleep apnea can involve repeated breathing pauses and loud snoring, and low oxygen readings can be inaccurate on consumer devices but still deserve context from a clinician, especially when they repeat or come with symptoms (NHLBI).

Pairing sleep data with daytime stress and recovery often tells a richer story than sleep alone: a "good" sleep score with low recovery, high stress, or poor HRV may explain why you still feel off. For the recovery side, see HRV. For what overnight oxygen readings can and cannot tell you, see measuring blood oxygen.

How we made it

Made with AI tools, then edited, fact-checked, and medically reviewed by the Welltory team.

Discounts for blog readers: up to 36% off

See what affects your energy, stress, sleep, and daily state with Welltory

Explore your data