À retenir
In a sleep laboratory, 34 healthy young adults wore a subset of seven consumer sleep trackers for three nights. Against polysomnography every device detected sleep very well (sensitivity 0.93 or higher) but caught wake poorly (specificity 0.18–0.54), and five of the seven matched or beat research actigraphy (specificity 0.39). Sleep-stage estimates were inconsistent, and the participants were healthy adults aged 18 to 35 in a fixed 8-hour bed window, with each device tested in only 11 to 20 of them, so the results say little about people with sleep problems.
Study design
Validation : étude/revue
Reading basis
Full text
Reading focus
Reference & error
Cette explication éditoriale existe en anglais, chinois simplifié et chinois traditionnel. Cette interface affiche la lecture en anglais.
Question de recherche
Hundreds of millions of people now use consumer devices to track sleep, and around 25% of US adults surveyed have used a wearable or app for it at least once. Polysomnography (PSG), which records brain waves, eye movements, muscle activity and heart rhythm overnight, is the gold standard, but it is costly, needs trained staff and can itself disturb sleep. Wrist actigraphy, the research standard outside the lab, tends to overestimate sleep and underestimate wake and can only label sleep or wake, not stages. Newer consumer devices add sensors such as heart rate to their sleep algorithms and might equal or beat actigraphy, but earlier studies found performance to be quite variable. The authors tested many of the latest wearable and nonwearable trackers against PSG, alongside actigraphy, as a first controlled-laboratory step in healthy young adults.
Population et méthode
This was a controlled laboratory validation study at the Naval Health Research Center in San Diego, California. It enrolled 34 healthy young adults (12 men, 22 women; age 28.1 ± 3.9 years, mean ± SD; BMI 24.2 ± 2.1 kg/m2) aged 18 to 35 who reported average nightly sleep of 6 to 9 hours and no diagnosed sleep, mental health or significant medical disorder, no sleep medication, nicotine or shift work. For 6 days before the first visit they gave up caffeine and alcohol, and for four nights they kept a steady 8-hour time in bed at home while wearing a Philips Respironics Actiwatch 2 research actigraph and the consumer wrist devices they would be tested with. They then slept three consecutive nights in the lab with exactly 8 hours in bed at the same clock times; on one of the last two nights, randomly assigned and counterbalanced, auditory tones woke them briefly every hour to mimic disrupted sleep. Four wearables (Fatigue Science Readiband, Fitbit Alta HR, Garmin Fenix 5S, Garmin Vivosmart 3) and three nonwearables (EarlySense Live under the mattress; ResMed S+ and SleepScore Max, bedside radiofrequency devices) were tested, but not all on everyone: each person wore the Actiwatch plus a subset, giving 15 participants for the Readiband, 20 for the Fitbit Alta HR, 11 for the Fenix 5S, 15 for the Vivosmart 3, 19 for EarlySense Live, 19 for ResMed S+ and 15 for SleepScore Max.
The reference was overnight PSG (EEG at six sites, eye movements, chin muscle activity, ECG), scored by hand in 30-second epochs by two certified technologists and re-scored in 60-second epochs for devices that output 60-second data; N1 and N2 were merged as light sleep and N3 counted as deep sleep. Each device was compared with PSG epoch by epoch for sleep versus wake and for each stage (sensitivity, specificity, positive and negative predictive value, accuracy and prevalence- and bias-adjusted kappa, PABAK) and on whole-night measures: total sleep time (TST), sleep efficiency, sleep onset latency, wake after sleep onset (WASO) and REM latency. Whole-night measures used Bland–Altman bias and limits of agreement (two standard deviations around the bias), paired t-tests, Hedges' g effect sizes and an R2 test for proportional bias, with P < 0.05 as significant. Three nonwearables' epoch data came from company research portals and the Fitbit's from the third-party platform Fitabase, and device data were post-processed as needed to cover the 8-hour in-bed window.
Résultats
- 1Every device was very good at recognizing sleep and much worse at recognizing wake. Epoch-by-epoch sensitivity against PSG was 0.93 or higher for all devices (the Actiwatch 0.97), while specificity ranged only from 0.18 to 0.54, so a large share of true wake time was scored as sleep. Accuracy was similar for all, 0.88 to 0.90.
- 2Five of the seven consumer devices matched or beat research actigraphy at catching wake, and the two Garmins did far worse. Specificity was 0.54 for the Fitbit Alta HR, 0.51 for ResMed S+, 0.50 for SleepScore Max, 0.47 for EarlySense Live and 0.45 for the Fatigue Science Readiband, against 0.39 for the Actiwatch; the Garmin Fenix 5S and Vivosmart 3 scored 0.18 and 0.19.
- 3Average errors in total sleep time and wake time were small for some devices and large for others. The Fitbit Alta HR overestimated TST by 2.6 min (P = 0.421) and underestimated WASO by 2.1 min (P = 0.472), and ResMed S+ underestimated TST by 0.3 min (P = 0.953). Both Garmins overestimated TST by 43.7 and 46.8 min (P < 0.001 for each) and reported WASO of 7.7 and 8.0 min against 57.2 and 55.6 min on PSG, while the Actiwatch overestimated TST by 23.9 min and underestimated WASO by 16.6 min (both P < 0.001).
- 4A small average bias hides large errors on individual nights, and errors grew on poorer nights. For the Fitbit Alta HR the TST difference from PSG on single nights ranged from 42.0 min below to 47.1 min above PSG, and for ResMed S+ from 70.7 min below to 70.2 min above. Bias was smallest when people slept longer and more efficiently and larger on nights with less sleep, longer time to fall asleep or more wake, with significant proportional bias in TST for the Actiwatch, Fitbit Alta HR, both Garmins and SleepScore Max (for example R2 0.69, P < 0.001 for the Vivosmart 3).
- 5Sleep-stage estimates were mixed and often poor. Five of six stage-reporting devices overestimated light sleep by 17.1 to 34.7 min (EarlySense Live underestimated it by 22.3 min, P = 0.006); EarlySense Live, ResMed S+ and SleepScore Max overestimated deep sleep (for example by 35.9 min for EarlySense Live, P < 0.001); Fitbit Alta HR, ResMed S+ and SleepScore Max underestimated REM sleep (by 11.4 min, P = 0.002; by 31.9 and 35.8 min, both P < 0.001). Epoch-by-epoch sensitivity was only 0.53 to 0.68 for deep and 0.49 to 0.69 for REM sleep, and wrongly scored epochs were usually called light sleep: 0.35 to 0.54 of PSG wake epochs, 0.27 to 0.43 of deep and 0.24 to 0.46 of REM. The Fitbit Alta HR had the highest values on most light and REM measures.
- 6Falling-asleep times were close, but REM timing was not. Sleep onset latency differed from PSG by under 5 min for every device (Fitbit Alta HR 3.1 min shorter, P = 0.005; ResMed S+ 4.0 min longer, P = 0.049; SleepScore Max 4.4 min longer, P = 0.017), and no consumer device differed on both latency measures. Five of the six stage-reporting devices significantly overestimated REM latency.
Limites et incertitude
- 1The authors note that not all seven devices could be tested at once or on all participants, because of comfort, availability and possible interference between devices; each device was therefore tested in only 11 to 20 of the 34 people, on different subsets, and occasional data were missing or partly lost.
- 2The authors note that the sample was healthy sleepers with no reported sleep problems and a fixed 8-hour bed window, so the results best apply to healthy sleepers on a typical night, that the lab and PSG electrodes can themselves affect sleep, and that they did not collect race or ethnicity, which matters for photoplethysmography heart rate sensors. They also note the results correspond best to the device versions used and to the medium Actiwatch setting.
- 3Our observation: device data were post-processed to the 8-hour in-bed window, so the study does not test how well a device finds your bedtime and wake time on its own, a step the authors flag as a source of error in real use. Three nonwearables' epoch data also came from company research portals rather than the standard app.
- 4Our observation: the disrupted-sleep night is described as a condition, but the results text pools all nights and reports poorer-night effects only as proportional bias, so the size of the effect on disrupted nights is not shown. Many devices, measures and P values were compared without a stated correction for multiple testing.
- 5Our observation: beating actigraphy is a low bar for wake detection. Even the best specificity, 0.54 for the Fitbit Alta HR, means roughly half of true wake epochs were missed, and the study cannot say whether any device is reliable enough for diagnosis or for people with insomnia.
Mise en perspective
If you wear a tracker, trust it most for whether you slept and for total sleep time, which the better devices here tracked closely on average, and trust it least for wake episodes and for deep, REM and light sleep minutes. Expect it to under-report restless wake time, especially on broken nights, and note that the two Garmin models tested here reported almost no wake after sleep onset (7.7 and 8.0 min against 57.2 and 55.6 min on PSG).
Do not conclude that stage percentages are accurate, that any device can diagnose a sleep disorder, or that these results apply to people with insomnia, older adults, children or newer device versions. A tracker showing good sleep does not rule out a problem, and persistent poor sleep or daytime sleepiness belongs with a clinician. The evidence is moderate for sleep versus wake in healthy young adults and weak for stages.
Vos données
Use total sleep time and the regularity of your bed and wake times from the same device over weeks rather than any single night, and keep a short sleep diary so you can spot the nights the tracker missed wake periods. Treat nightly deep, REM and light minutes as rough, since they were often off in this study, and do not compare stage numbers between brands. If your app lets you correct bedtime or wake time, fix wrong entries, because the study's analysis relied on the in-bed window being right.
Comprendre ce type de preuve
Agreement is specific to the device, firmware, reference, population and setting. Correlation is not agreement, and validating a metric is not proof of diagnostic or treatment benefit.
Sources et attribution
Performance of seven consumer sleep-tracking devices compared with polysomnography
Sleep · 2021-05-14
DOI: 10.1093/sleep/zsaa291 ↗PubMed · 33378539 ↗Interprétation éditoriale MoveVitals · contenu éducatif · aucune évaluation clinique individuelle enregistrée
Sources vérifiées 2026-11-01
Guides pratiques
À lire ensemble
Activity–stroke and heart attack curves differ by accelerometer processing method
Among 90,237 UK Biobank adults free of cardiovascular disease at baseline, followed until December 2022 for 1,298 strokes and 2,031 myocardial infarctions, the shape of the link between moderate-vigorous physical activity (MVPA) and each outcome depended on how the accelerometer data were processed. For stroke, hazard fell in a straight line only with the machine-learning metric; for myocardial infarction, two metrics gave a curved decline and activity counts a straight one. The abstract gives no hazard ratios, so it shows that methods disagree, not which one is right, and the findings are observational associations.
Preventive Medicine
Smartwatch data as forensic evidence: heart rate is best validated, the rest uneven
A systematic review and meta-analysis of studies on smartwatches from four major brands, framed around digital forensics, found that heart rate was both the most studied and the most accurate measure. Step count, distance, energy expenditure, activity classification and sleep were less consistently validated, and results varied by brand, protocol and reporting. The abstract gives no pooled numbers, and most studies were run in healthy people in laboratories, so it cannot say how a particular watch performs for a particular person in daily life.
Forensic Science International
Activity trackers: about 1,800 more steps a day, but small effects on other outcomes
An umbrella review of 39 systematic reviews and meta-analyses, covering 163,992 participants, found that activity trackers were associated with more physical activity (SMD 0.3–0.6), equating to about 1,800 extra steps and 40 minutes more walking per day, and with a bodyweight reduction of about 1 kg. Effects on blood pressure, cholesterol, glycosylated haemoglobin, quality of life and pain were typically small and often non-significant. The abstract does not say what else came with the trackers in these programs or how much the reviews overlap, so the gains cannot be credited to the device alone.
The Lancet Digital Health
Seven wrist devices: heart rate mostly within 5% error, calorie estimates far off
Sixty adults wore seven wrist-worn devices while sitting, walking, running and cycling in a laboratory. Median heart rate error against an electrocardiogram (ECG) ran from 1.8% on the bike to 5.5% walking, but no device achieved a calorie error below 20 percent (device medians from 27.4% to 92.6%). These were consumer models on sale at the time, tested in a lab, so the results do not rank today's watches or show how they perform in daily life.
Journal of Personalized Medicine

