Sleep tracker accuracy is not one number. A wrist device makes at least three different claims each night: when you fell asleep and woke up, how much you slept, and which stages you passed through. Each claim has a different evidence base, and the third is the weakest. Before you trust a result, check which claim you are leaning on, which model produced it, and what it was compared against.
This article is about auditing the evidence. It does not interpret sleep physiology, set training around recovery scores, or diagnose anything. Splitting "accuracy" into separate metrics is the general first step for any fitness tracker; this piece applies that idea to sleep.
Separate the three sleep outputs
| Output | The question | What a validation must show |
|---|---|---|
| Sleep start and end | Did the device place the night's boundaries correctly? | Time-aligned onset and offset against a reference, with bias and spread |
| Total sleep time | Is the amount of sleep close? | Minutes of difference per night, not just a high correlation |
| Sleep stages | Were light, deep and REM labeled correctly? | Epoch-by-epoch comparison against polysomnography (PSG) |
Boundaries and total time are the easier targets. Stages are a classification problem made from movement and heart-rate signals, so errors can hide inside a plausible-looking nightly total.
One 2021 naturalistic study shows the gap. It compared a consumer wrist device against portable home PSG in 59 first-responder police officers and paramedics doing shift work. It found unbiased average estimates of sleep onset and offset, total sleep time, REM and N3 duration, but wide limits of agreement. The authors concluded the wide limits hamper precision for individual sleep episodes. The same study found that the device overestimated WASO (wakefulness after sleep onset) by 37.1 minutes. Its epoch-level accuracy was 0.86 for REM but 0.55 for N1+N2.
"Unbiased on average" means errors roughly cancel across many nights. It does not mean your Tuesday is right.
Check the reference and the validation population
A validation is only as useful as its details. When you see a claim like "this tracker is 90% accurate," look for these items:
- Exact model and generation. The shift-worker results above belong to one older wrist model. A newer or different model, or different firmware, needs its own evidence. A systematic review of three wrist-worn consumer devices, based on 8 included articles, covered only those specific devices.
- Reference method. PSG is the standard. Studies against other reference devices are not equivalent, and that review notes that one of its comparisons involved PSG together with a different headband-style sleep monitor.
- Population and sample. The shift-worker study also cites research in obstructive sleep apnea patients in which WASO was underestimated by two models, suggesting performance can differ by clinical population. A healthy-sleeper result is not automatically yours.
- Conditions. A lab night is not a home night. Naturalistic studies add realism but also noise, including time alignment between device and PSG, which the authors identify as a validation challenge.
- Metric. Correlation says two series move together; it does not say they match. Look for mean bias, limits of agreement, sensitivity, specificity, or Cohen's kappa.
- Direction of the error. Sensitivity and specificity can diverge sharply.
A 2025 laboratory study of six commercial wrist-worn devices against PSG included 62 adults and shows why direction matters. All devices showed sensitivity above 90% for sleep epochs, but specificity for wake ranged only from 29.39% to 52.15%, so wake was often scored as sleep. Kappa ranged from 0.21 to 0.53, fair to moderate agreement. Most devices differed significantly from PSG on total sleep time, sleep efficiency, WASO and light sleep. These are results for those tested devices in that laboratory setting, not benchmarks for current models.
Why one device can look good on one output and weak on another
The systematic review above found that devices could show small average differences from PSG for total sleep time and some stages while disagreeing more on another stage, such as REM. It also found moderate accuracy for stages and total time overall, and warned that research on these wearables is limited and that all devices could improve at specific stages.
Treat that as a pattern, not a ranking: a good total-sleep figure says little about stage labels, and a result for one model does not transfer to another model, firmware or maker.
A worked audit: wear time and missing data
Before judging any night, confirm the device actually recorded it. The numbers below are illustrative, not from a real test.
| Step | Input (illustrative) | Judgment | Output |
|---|---|---|---|
| Log lights-out and out-of-bed time by hand | In bed 11:00 pm, up 6:30 am | Time in bed is context, not time asleep | 7 h 30 min in bed |
| Read the device's sleep window | 11:20 pm to 6:25 am | Differs from the diary by 20 min at the start and 5 min at the end; people take time to fall asleep and may lie awake after waking, so this is a mismatch to inspect, not proof of error | 7 h 5 min window to inspect |
| Note wear gaps | Band removed to charge 5:50–6:10 am | Gap overlaps the end of the night | Mark the night as incomplete |
| Check the stage summary | Deep sleep looks low | Stages are the weakest output, and the night is incomplete | Do not act on it |
A diary records when you got into bed and got up. It is not a sleep-stage reference, and it does not define sleep onset the way PSG does. A close match with your diary therefore does not prove the device's boundaries or stages are accurate; only a PSG comparison can do that. The diary's job is to flag nights that look odd.
The result is a rule: a night with a charging gap, loose strap or late boundary is excluded from comparison rather than explained. If boundaries disagree with your log by a large margin across several nights, treat the sleep window as questionable, and give stage details inside it little weight.
If you see recurring gaps, fix wear habits first: charge at a consistent daytime time, and keep the band snug and in the same position.
Limit how you interpret personal results
The shift-worker study found wide individual limits of agreement even where averages were fine. That supports a plain stance: use the device for trends in boundaries and total time, and treat stage labels as rough.
Acceptance checklist:
- [ ] I know the exact device model and which validation studies exist for it.
- [ ] The study used PSG as its reference, and I know the sample and conditions.
- [ ] It reported agreement (bias and limits, or kappa), not only correlation.
- [ ] The population resembles me, or I accept that the result may not transfer.
- [ ] The night has no wear gaps, and device boundaries are not wildly off my diary (a flag, not validation).
- [ ] I am looking at a pattern over many nights, not one night.
- [ ] I am not treating the stage chart as a diagnosis.
If persistent loud snoring, breathing pauses, or daytime sleepiness worry you, a consumer interface cannot answer that. The studies above compared devices against PSG, which a clinician or sleep lab can arrange. Sleep volume, interruptions and routines are a separate topic, best handled in recovery coverage rather than inferred from stage bars.
The bottom line
Trust sleep start, end and total time more than stages, and trust none of them from a single night's screen. Match the claim to the model, reference, population and metric before it changes what you believe. When a night looks alarming, check wear time and your own log first, and let repeated patterns, not one chart, guide any question for a professional.
