How accurate are consumer sleep wearables? Oura, Apple Watch and Fitbit vs polysomnography
What published comparisons found for Oura, Apple Watch, and Fitbit against overnight sleep recordings. Not a recommendation.
Consumer rings and watches now print a sleep score in the morning. The useful question is narrower than the marketing: when a laboratory recording, or a careful home recording, says a person was asleep, awake, or in a particular stage, how often does the device agree?
This note reads five published comparisons. It is not a buying guide, and it is not a comment on anyone’s sleep. The studies enrolled healthy volunteers, often for one night or one week. They do not show how a device behaves across years, in shift work, or in people whose sleep is fragmented for medical reasons.
Grade B. Small, pilot or observational human study
These are agreement studies, not randomized trials of a health outcome. The grade names that design. It is not a score for a brand.
What the reference recording is
Polysomnography is an overnight recording used as the reference in four of the five papers [1]Citation 1[2]Citation 2[3]Citation 3[4]Citation 4. Technicians score the night in short epochs, usually 30 seconds, from brain waves, eye movements, muscle tone, and the other signals on the montage. From those epochs they build total sleep time, time awake after sleep onset, and the stages: light sleep, deep sleep, and REM.
A wearable does not record that montage. A ring or watch uses motion and pulse signals, then a proprietary algorithm. The test is whether those outputs land on the same epochs and the same summary minutes.
Two words in the papers are easy to mix up. Sensitivity for sleep is the share of reference sleep epochs the device also called sleep. Specificity for wake is the share of reference wake epochs the device also called wake. A device can score very high on the first and only middling on the second: it rarely misses sleep, and it often files quiet wake under sleep. Predictive value asks a third question: when the device says “wake,” how often the reference agrees.
The fifth paper did not use in-laboratory polysomnography. At home, Chinoy and colleagues compared wearables with a mobile electroencephalography headband, the Dreem 2, plus diaries and a research actigraph [5]Citation 5. That is still a physiological recording. It is not the same setup as a hospital night, and this note does not pretend that it is.
One night, three devices, healthy sleep
Robbins and colleagues enrolled 35 adults aged 20 to 50 years who did not have a sleep disorder. Each person spent a single night in hospital wearing an Oura Ring Generation 3, a Fitbit Sense 2, and an Apple Watch Series 8, with polysomnography as the reference [1]Citation 1.
For sleep versus wake, sensitivity was at least 95 percent for all three devices. That is the “good at detecting sleep” result. It was measured in people whose sleep was already healthy, on one night, in a laboratory bed.
Sleep stages were less settled. Sensitivity for stages ranged from 50 to 86 percent. For the Oura ring, stage sensitivity was 76.0 to 79.5 percent and precision was 77.0 to 79.5 percent. For the Fitbit, sensitivity was 61.7 to 78.0 percent and precision was 72.8 to 73.2 percent. For the Apple Watch, sensitivity was 50.5 to 86.1 percent and precision was 72.7 to 87.8 percent [1]Citation 1.
On summary minutes, the three devices were not interchangeable. The Oura ring was not different from polysomnography for wake, light sleep, deep sleep, or REM. The Fitbit overestimated light sleep by 18 minutes and underestimated deep sleep by 15 minutes. The Apple Watch underestimated wake by 7 minutes and deep sleep by 43 minutes, and overestimated light sleep by 45 minutes. Sleep duration, as opposed to the split among stages, was similar to polysomnography for the devices in this group [1]Citation 1.
Forty-three minutes of deep sleep is a large slice of a night. The stage chart can shift by that much even when total hours asleep look right. Robbins and colleagues also report moderate to substantial agreement on stages. Agreement and a bias in minutes can both be true.
Robbins reports membership of the Oura Ring Medical Advisory Board and consulting fees from Oura Ring Inc. Other authors list industry relationships that the paper discloses [1]Citation 1. The conflict belongs next to the measurements.
Oura Generation 3 across more than one night
Svensson and colleagues tested the Oura Ring Generation 3 with the company’s sleep staging algorithm 2.0 against multi-night ambulatory polysomnography. Ninety-six generally healthy adults in Japan, aged 20 to 70 years, contributed up to three nights. The analysis used 421,045 epochs of 30 seconds, scored to American Academy of Sleep Medicine criteria. Rings were worn on the dominant and the non-dominant hand [2]Citation 2.
On several summary measures the ring did not differ significantly from polysomnography: time in bed, total sleep time, sleep onset latency, sleep period time, wake after sleep onset, light sleep, and deep sleep. It underestimated sleep efficiency by 1.1 to 1.5 percent and REM sleep by 4.1 to 5.6 minutes [2]Citation 2.
Epoch by epoch, sensitivity for sleep was 94.4 to 94.5 percent. Specificity for wake was 73.0 to 74.6 percent. When the ring called an epoch sleep, the predictive value was 95.9 to 96.1 percent. When it called an epoch wake, the predictive value was 66.6 to 67.0 percent. Overall accuracy was 91.7 to 91.8 percent. A prevalence-adjusted kappa was 0.83 to 0.84. Sleep-stage accuracy ran from 75.5 percent for light sleep to 90.6 percent for REM [2]Citation 2.
When the ring called an epoch wake, it was wrong about one time in three. The authors’ conclusion is agreement on global sleep measures and on time in light and deep sleep, not a claim that every stage label matches a technician’s score [2]Citation 2.
Funding is part of the record. Thomas Svensson, Ung-il Chung, and Akiko Kishi Svensson report financial support from Oura Health Oy [2]Citation 2.
An earlier ring, in younger sleepers
de Zambotti and colleagues compared an earlier Oura sleep tracker with a single laboratory night of polysomnography in 41 healthy adolescents and young adults (13 females; mean age 17.2 years, standard deviation 2.4) [3]Citation 3. The journal issue is 2019; the paper was published online in 2017. It is not a test of the Generation 3 algorithm used by Svensson or of the Generation 3 ring in the Robbins night.
Summary sleep onset latency, total sleep time, and wake after sleep onset were not different from polysomnography. The gap in wake after sleep onset grew as people had more polysomnographic wake. The ring underestimated N3 sleep by about 20 minutes and overestimated REM by about 17 minutes. For 87.8 percent of the sample, the total-sleep-time gap sat inside a 30-minute band the authors had set in advance. For wake after sleep onset, 85.4 percent sat inside that band [3]Citation 3.
Epoch by epoch, sensitivity for sleep was 96 percent. Agreement was 65 percent for light sleep (N1 in their grouping), 51 percent for deep sleep (they pooled N2 and N3), and 61 percent for REM. Specificity for wake was 48 percent. The ring placed 90.9 percent, 81.3 percent, and 92.9 percent of people into the same total-sleep-time bins as polysomnography: under 6 hours, 6 to 7 hours, and over 7 hours [3]Citation 3.
The authors called the result promising and said more validation was needed. A mean age of 17 is not a result about older adults, and this ring is not the Generation 3 algorithm tested later. The finding that still travels is the wake problem: when the reference night contained more wake, the disagreement was larger [3]Citation 3.
A disrupted night, and a Fitbit that is not the Sense 2
Chinoy and colleagues, in 2021, tested seven consumer devices against polysomnography in 34 healthy young adults (22 women; mean age 28.1 years, standard deviation 3.9) across three consecutive laboratory nights, including a night with disrupted sleep. They also wore a research actigraph, the Philips Respironics Actiwatch 2 [4]Citation 4.
The wearables in that study were the Fatigue Science Readiband, Fitbit Alta HR, Garmin Fenix 5S, and Garmin Vivosmart 3, plus three bedside devices (EarlySense Live, ResMed S+, and SleepScore Max). Oura and Apple Watch were not in this protocol. Any sentence about “the Apple Watch on a broken night” that cites only this paper is ahead of the evidence.
Most of the devices performed as well as or better than actigraphy on sleep-versus-wake measures. The Garmin watches performed worse. Epoch-by-epoch sensitivity for sleep was high, at least 0.93 for every device. Specificity for wake was low to medium, from 0.18 to 0.54. Sleep-stage comparisons were mixed. Devices tended to perform worse on the nights with poorer or disrupted sleep. The volunteers were still healthy young adults, and “disrupted” means the laboratory condition the investigators imposed [4]Citation 4.
A week at home, against a headband
The same group then left the laboratory. In 2022 they followed 21 healthy young adults (12 women, 9 men; mean age 29.0 years, standard deviation 5.0) for one week of unrestricted sleep at home. Devices were the Fatigue Science Readiband, Fitbit Inspire HR, an Oura ring, and the Polar Vantage V Titan, plus the Actiwatch 2. The reference was the Dreem 2 headband, a mobile electroencephalography device, with daily sleep diaries alongside [5]Citation 5.
Summary sleep-wake metrics (time in bed, total sleep time, sleep efficiency, sleep latency, and wake after sleep onset) were tracked accurately on most nights, and performance was best on nights with higher sleep efficiency. Epoch-by-epoch sensitivity and specificity were 0.95 and 0.35 for the Actiwatch, 0.94 and 0.40 for Fatigue Science, 0.93 and 0.45 for Fitbit, 0.94 and 0.41 for Oura, and 0.96 and 0.35 for Polar. Stage tracking was mixed and highly variable [5]Citation 5.
Specificity near 0.4 means that, of the epochs the headband called wake, the device called wake less than half the time. The authors’ conclusion matches the laboratory papers: better at sleep than at wake, best when sleep is consolidated, and not yet reliable for stages. A favorable comparison with actigraphy is not a pass against polysomnography. The authors report no relationship with the device companies. Chinoy, Cuellar, and Jameson are employees of Leidos, Inc. [5]Citation 5.
What a shopper can and cannot take from this
The devices are not one product. “Oura” covers an early ring in adolescents and young adults [3]Citation 3, Generation 3 on one hospital night [1]Citation 1, Generation 3 with algorithm 2.0 across ambulatory nights [2]Citation 2, and a ring at home against a headband [5]Citation 5. “Fitbit” covers the Sense 2 [1]Citation 1, the Alta HR [4]Citation 4, and the Inspire HR [5]Citation 5. “Apple Watch” covers Series 8 on one night [1]Citation 1. A firmware update can move the numbers. These papers cannot score a model that was not on the wrist.
Sample sizes are modest: 35, 96, 41, 34, and 21 people [1]Citation 1[2]Citation 2[3]Citation 3[4]Citation 4[5]Citation 5. Several tests are a single night. The volunteers were healthy, and several samples were young. These papers do not estimate error outside that kind of volunteer.
If the question is “did I sleep, and for about how long?”, these devices often land close to the reference in healthy sleepers, with the wake errors described above. If the question is “how many minutes of deep sleep did I get, to the precision of a sleep study?”, the stage gaps, especially on the Apple Watch in Robbins and on the earlier Oura ring in de Zambotti, are too large to treat the chart as a hypnogram [1]Citation 1[3]Citation 3. Broken nights are where the tools are least convincing [3]Citation 3[4]Citation 4[5]Citation 5.
This note does not rank a purchase. It also does not say that a wearable can rule a sleep problem in or out. The studies cited here were built to test agreement in volunteers, not to screen a clinic.
References
- Robbins et al. (2024). Accuracy of three commercial wearable devices for sleep tracking in healthy adults. Sensors. Back to text
- Svensson et al. (2024). Validity and reliability of the Oura Ring Generation 3 with Oura sleep staging algorithm 2.0 compared with multi-night ambulatory polysomnography. Sleep Medicine. Back to text
- de Zambotti et al. (2019). The sleep of the ring: comparison of the Oura sleep tracker against polysomnography. Behavioral Sleep Medicine. Back to text
- Chinoy et al. (2021). Performance of seven consumer sleep-tracking devices compared with polysomnography. Sleep. Back to text
- Chinoy et al. (2022). Performance of four commercial wearable sleep-tracking devices tested under unrestricted conditions at home in healthy young adults. Nature and Science of Sleep. Back to text