REF-1252 · Lee et al., 2025 · Journal of Clinical Sleep Medicine · Paper · PubMed · Free full text

How accurate is my watch's sleep time?

On average wrist trackers read about 17 min more sleep than a lab sleep test and missed some wake (24 studies). Useful for general patterns; one night's number can be far off.

B+moderate evidenceMeasured on non-athletes and recreational people

Lee and colleagues pooled studies that compared sleep measured by consumer wrist-worn devices with polysomnography on the same night. Searches of seven databases (to March 2024) yielded 24 studies with 798 participants, most of them healthy adults, using Fitbit, Garmin, Jawbone, WHOOP, Apple Watch, Xiaomi, Huami, myCadian, Basis, Zulu, Empatica E4 and Fatigue Science devices.

Random-effects meta-analyses found significant mean differences between devices and polysomnography for total sleep time (−16.854 min, 95% CI −26.332 to −7.375; I² 67.5%), sleep efficiency (−4.691 points; I² 84.1%), sleep latency (2.574 min; I² 57.9%) and wake after sleep onset (13.255 min; I² 67.4%).

After trim-and-fill for suspected publication bias, the latency and wake differences were no longer significant. In subgroups, Fitbit's wake after sleep onset and other devices' sleep latency did not differ from polysomnography.

The authors conclude that wrist trackers are not as reliable as polysomnography for total sleep time, efficiency and latency but remain useful for general sleep patterns. The paper never says whether a difference is polysomnography minus device or the reverse.

In the forest plot, a Fitbit Flex in "sensitive" mode (which scores more movement as wake than the normal mode) gives total sleep +86.3 min and wake −74.8 min, against −45.9 and +44.0 min in normal mode; that is only consistent with polysomnography minus device, so devices over-read total sleep by about 17 min and efficiency by about 4.7 points and under-read wake by about 13 min (our reading).

That matches REF-364 and REF-1248, but contradicts the authors' wording that Fitbit "reported longer sleep latency" and other devices "reported longer wake after sleep onset" than polysomnography.

Our reading: on average a wrist device flatters the night by a quarter of an hour, mostly by counting quiet wake as sleep; how far one night can be off is not pooled, and per-study averages alone ran from about −105 to +86 min.

Protocol

Search: SCOPUS, EMBASE, Cochrane Library, PubMed (Medline), Web of Science, KoreaMed and OVID Medline, plus reference lists; English only; run 21 October 2022 and updated 15 March 2024. No protocol registration is mentioned.
Eligibility: adults (healthy or with disease), consumer wrist-worn sleep trackers, objective sleep indicators, comparison with polysomnography (including sleep tests with EEG analysis); excluded studies of non-wrist devices, unclear age or no objective indicators.
Selection: two researchers independently; a third resolved disagreements. 3,727 records after de-duplication, 3,640 excluded on title and abstract, 63 more at full text; 24 studies included, n = 798.
Devices: Fitbit (Classic, Flex, Versa, Versa 2, Charge 2, Charge 4, Alta HR, Inspire HR), Garmin (Vivosmart 3, Vivosmart 4, Fenix 5S), Jawbone (UP, UP3), WHOOP strap, E4 wristband, myCadian watch, Basis B1, Fatigue Science Readiband, Zulu Watch, Xiaomi Mi Band 5, Huami Arc, Apple Watch (Series 8).
Quality: Newcastle–Ottawa Scale converted to AHRQ categories; all studies rated good.
Analysis: mean differences with 95% CI from study means and SDs; Cochran's Q and I²; random effects because heterogeneity was high; Begg's and Egger's tests (one-sided) and Duval–Tweedie trim-and-fill; subgroups healthy vs unhealthy, Fitbit vs other devices, and studies after 2021 vs before 2020.
Funding and interests: Soonchunhyang University Research Fund; the authors report no conflicts of interest.

Key Results

OutcomeStudiesPooled mean difference (95% CI)I²Direction by our reading of the forest plot
Total sleep time22−16.854 min (−26.332 to −7.375), P < .00167.5%device reads more sleep
Sleep efficiency18−4.691 points (−7.079 to −2.302), P < .00184.1%device reads higher efficiency
Sleep latency182.574 min (0.606 to 4.542), P = .010; trim-and-fill 0.611 (−1.577 to 2.788)57.9%device reads shorter latency
Wake after sleep onset2013.255 min (4.522 to 21.988), P = .003; trim-and-fill 0.171 (−0.122 to 0.464)67.4%device reads less wake

Subgroups (mean differences, 95% CI):

SubgroupTotal sleep timeSleep efficiencySleep latencyWake after sleep onset
Fitbit−6.564 (−18.890 to 5.763)−3.626 (−7.639 to 0.387)5.602 (2.105 to 9.098)0.533 (−12.517 to 13.583)
Other devices−26.674 (−39.051 to −14.297)−5.498 (−8.050 to −2.946)−0.122 (−1.493 to 1.249)24.140 (15.735 to 32.545)
Healthy−12.504 (−21.420 to −3.587)−6.493 (−8.989 to −3.996)2.005 (−9.452 to 13.452)13.910 (3.036 to 24.782)
Unhealthy−26.188 (−50.158 to −2.217)−2.159 (−6.731 to 2.413)3.120 (−2.231 to 8.470)11.955 (−4.192 to 26.102)
After 2021−20.544 (−34.832; 6.255 as printed)−5.830 (−9.502 to −2.158)1.782 (−0.452 to 4.016)14.929 (0.404 to 29.453)
Before 2020−13.174 (−25.443 to −0.905)−3.744 (−7.057 to −0.431)3.406 (0.102 to 6.710)12.634 (2.798 to 22.469)
Per-study spread (Figure 1, read from the image): study-level total-sleep differences ran from about −105 min (an E4 wristband) to +86 min (a Fitbit Flex in sensitive mode); a Garmin Fenix 5S row shows about −41 min of total sleep and +50 min of wake.
Publication bias: Begg and Egger tests were non-significant for total sleep time and efficiency; Egger's P = .017 for latency.
Numbers that do not add up (our reading):
The sign convention is not stated. The Fitbit Flex sensitive-mode rows only make sense as polysomnography minus device; under that convention the authors' sentences that Fitbit "reported longer sleep latency than polysomnography" and that other devices "reported longer wake after sleep onset" describe the opposite of their numbers.
Trim-and-fill: the text says latency and wake "significantly different" with confidence intervals that include zero, then that they "may not demonstrate significant differences"; the wake estimate after trim-and-fill (0.171) appears to be on a different scale from the 13.255-min estimate.
Egger's P = .017 for latency is first reported as confirming "no publication bias", then as the reason bias was suspected.
Recent studies, total sleep time: "−20.544; 95% CI, [−34.832; 6.255]; P = .005" — a CI that crosses zero cannot give P = .005; −20.544 is the midpoint of −34.832 and −6.255, so the upper limit has lost its minus sign and recent studies did differ (our arithmetic), contrary to the next sentence saying they showed no significant difference in total sleep time.
Healthy group, latency: a CI of −9.452 to 13.452 around 2.005 implies a P near 0.7, not the .054 printed (our arithmetic).
Statements that Fitbit and other devices differ significantly are not backed by a reported test between subgroups.

Mechanisms

Measured: pooled across devices, wrist trackers and polysomnography differed on all four sleep measures, with high heterogeneity; differences in latency and wake depended on the device group.
Authors' interpretation: polysomnography scores sleep by hand from EEG, eye and chin-muscle signals, while wrist devices infer it from movement, heart rate and HRV with proprietary algorithms that differ by maker and model; populations with sleep disorders add variability.
Inference: a wrist device cannot see the brain, so lying still awake looks like sleep to it — REF-364 found wake detection of 26-57% across six devices — which is why the errors run toward more sleep, higher efficiency and less wake, and why the error grows on nights with long quiet wakefulness.
Inference: a pooled mean of about 17 min is a group average; with study averages alone spread from about −105 to +86 min, a single athlete's night can be off by much more than the headline figure.

Practical Application

How accurate is my watch's sleep time? On average wrist trackers read about 17 min more sleep than a lab sleep test and missed some wake (24 studies). Useful for general patterns; one night's number can be far off.
Use the tracker for bedtime, wake time and the trend in total sleep — consistent with SYNTH-07's row "use the device for bedtime, wake time and total sleep only" — but expect the total to run about a quarter of an hour high on average (B+), and more on restless nights (inference).
Do not judge wake time in the night or how fast the athlete fell asleep from a wrist device: the pooled latency and wake differences depended on the brand and were unstable under publication-bias correction, single studies' wake differences ran from about −75 to +76 min (forest plot), and REF-1248 reports latency and wake errors of 12-180%.
A short night on the tracker is more likely real than an artefact (inference from the direction): devices over-read rather than under-read sleep on average; exceptions exist (one Fitbit setting under-read total sleep by more than an hour in the forest plot).
Believe the athlete over the device when they disagree, as SYNTH-07 already says; this paper did not analyse sleep stages at all.

Limitations

Direction is inferred. The paper does not define its mean differences; the direction above is read from the Fitbit Flex normal- and sensitive-mode rows of the forest plot image and agrees with REF-364 and REF-1248, but contradicts the authors' wording.
No individual error. Only mean differences were pooled; limits of agreement, which describe a single night, were not.
Heterogeneity and quality. I² 57.9-84.1%; all studies rated "good" on the Newcastle–Ottawa Scale, which is not designed for diagnostic-accuracy studies; no registered protocol is mentioned.
Statistical reporting contains contradictory confidence intervals, p-values and conclusions in the trim-and-fill and subgroup analyses (Key Results).
Devices and people. Mostly discontinued models, Fitbit-heavy; healthy adults and patients, no athletes; the number of nights per study is not summarised.
No sleep stages. Staging accuracy was not analysed (the authors' limitation).
Interests. University research fund; no conflicts reported.
How this entry was checked and changed
Why this level

a meta-analysis of small single-night validation studies with substantial heterogeneity would sit at A-; held at B+ because the paper never states which way its mean differences run and its own text describes latency and wake after sleep onset in the direction opposite to its forest plot, several subgroup confidence intervals and p-values contradict each other, limits of agreement were not pooled, no protocol registration is mentioned, and all 24 studies were rated "good" on a scale built for observational studies. Read from the full text (PubMed Central repository copy, located through OpenAlex) and the forest-plot image

Full text (checked 2026-10-27)
From 24 studies, data of 798 patient using Fitbit, Jawbone, myCadian watch, WHOOP strap, Garmin, Basis B1, Zulu Watch, Huami Arc, E4 wristband, Fatigue Science Readiband, Apple Watch, or Xiaomi Mi Band 5 were analyzed (full text)
Total sleep time, 22 studies: I2 = 67.5%; mean difference −16.854; 95% CI [−26.332; −7.375]; P < .001 (full text)
Sleep efficiency, 18 studies: I2 = 84.1%; mean difference −4.691; 95% CI [−7.079; −2.302] (full text)
Sleep latency, 18 studies: I2 = 57.9%; mean difference 2.574; 95% CI [0.606; 4.542]; P = .010; trim-and-fill 0.611 [−1.577; 2.788] (full text)
Wake after sleep onset, 20 studies: I2 = 67.4%; mean difference 13.255; 95% CI [4.522; 21.988]; P = .003; trim-and-fill 0.171 [−0.122; 0.464] (full text)
Fitbit vs other devices: total sleep time −6.564 vs −26.674; sleep latency 5.602 vs −0.122; wake after sleep onset 0.533 vs 24.140 (full text)
Recent studies: total sleep time MD −20.544; 95% CI [−34.832; 6.255]; P = .005 (full text)
Searches run on October 21, 2022, updated March 15, 2024; 3,727 studies retrieved after removing duplicates (full text)
This study was supported by the Soonchunhyang University Research Fund; the authors report no conflicts of interest (full text)