Your Ring Says 47 Minutes of Deep Sleep. The Lab Says Maybe.

The full story

Somewhere on your phone right now is a number telling you how many minutes you spent in deep sleep last night. It is rendered in the same crisp typeface as your step count, on the same dashboard, with the same air of having been measured. It was not measured. It was inferred, from pulse and motion, by a model trying to reconstruct a classification scheme that was originally defined on brain electrical activity.

The good news is that these devices have been subjected to unusually thorough scrutiny, and the resulting literature is large, consistent and blunt. Its verdict comes in three parts. Your ring or watch is genuinely good at knowing whether you are asleep. It is mediocre at knowing when you are awake. And it is worst at the one number the marketing leans on hardest.

The asymmetry nobody prints on the box

Validation studies report two figures that sound like siblings and behave like opposites. Sensitivity for sleep is the share of genuinely-asleep moments a device correctly calls sleep. Specificity for wake is the share of genuinely-awake moments it correctly calls awake. Consumer trackers post excellent numbers on the first and poor ones on the second, and the size of that gap is the whole story.

A 2024 validation study in Sleep Medicine put the Oura Ring Gen3 against multi-night ambulatory polysomnography in 96 adults and scored 421,045 thirty-second epochs. Sensitivity for sleep came in around 94 percent. Specificity for wake was 73.0 to 74.6 percent, and when the ring did announce you were awake it was right only about two-thirds of the time. A 2025 validation study in Sleep Advances was harsher. Across six wrist devices in 62 adults, covering the Fitbit Charge 5 and Sense, Withings Scanwatch, Garmin Vivosmart 4, Whoop 4.0 and Apple Watch Series 8, every one detected more than 90 percent of sleep epochs while wake specificity ranged from 29.39 to 52.15 percent. A 2018 validation study in Chronobiology International found the same shape a hardware generation earlier for the Fitbit Charge 2: 0.96 sensitivity for sleep, 0.61 specificity for wake.

Read that asymmetry as a bias. A device that resolves ambiguity toward asleep will tend to be generous about your night. In aggregate the arithmetic partly cancels out: a 2025 meta-analysis in the Journal of Clinical Sleep Medicine pooled 24 studies and 798 participants and found discrepancies against polysomnography that were statistically significant but modest, at about 17 minutes on total sleep time, 4.7 percentage points on sleep efficiency, under 3 minutes on sleep latency and about 13 minutes on wake after sleep onset. Those are the well-behaved numbers. They are also the ones nobody stares at.

Deep sleep is the least trustworthy figure on the dashboard

Ask a device which stage you were in and agreement collapses. A 2024 study in Sensors ran the Oura Ring Gen3, Fitbit Sense 2 and Apple Watch Series 8 through a single inpatient night in 35 adults aged 20 to 50. All three cleared 95 percent sensitivity for sleep versus wake. Asked to discriminate individual stages, sensitivity fell to between 50 and 86 percent, and the Apple Watch underestimated deep sleep by 43 minutes while overestimating light sleep by 45 minutes. Not a rounding error: a whole sleep cycle misfiled.

The pattern repeats. A 2022 validation study in Sensors tested six devices against simultaneous polysomnography in 53 adults and found 86 to 89 percent agreement on the two-state sleep-or-wake question but only 50 to 65 percent when the devices had to name a stage, with chance-corrected agreement (Cohen's kappa) of 0.20 to 0.52. The 2025 Sleep Advances validation landed in the same band and described its stage-level agreement as fair to moderate, kappa 0.21 to 0.53. A 2025 study in Scientific Reports, run in a sleep-laboratory patient cohort, found the Oura ring's group-average total sleep time came within 12 minutes of polysomnography while individual-night errors stayed large; sleep-versus-wake accuracy was about 85 percent for two of the rings tested, but four-stage accuracy was 53.18, 50.48 and 35.06 percent across the three devices, with per-stage sensitivity as low as 0.14. And a 2023 multicentre validation study in JMIR mHealth and uHealth ran 11 consumer trackers against polysomnography in 75 participants over 349,114 epochs, reporting epoch-by-epoch staging performance from a macro F1 of 0.69 for the best device down to 0.26 for the worst. The distance between best and worst is wider than most product comparisons will admit exists.

The errors are biggest in the people most likely to be worried

A 2026 study in Sleep Advances tested consumer wearables and bedside sensors in 19 older adults aged 56 to 80. The devices underestimated total sleep time badly, the Fitbit Sense 2 by 74.5 minutes and Oura by 75.5 minutes, while overestimating deep sleep by 71.5 minutes for Oura, 88.8 minutes for SleepScore Max and 97.4 minutes for the Withings Sleep Mat. Limits of agreement were wider in the older cohort than the younger one, and across the board the devices performed worst at identifying deep sleep. Which is to say: the group most likely to buy a tracker because their sleep has changed is the group in which it is least reliable.

What the evidence does not show

It does not show that these devices can find a disease. The American Academy of Sleep Medicine's 2018 position statement is explicit that, given the lack of validation against polysomnography and the absence of FDA clearance, consumer sleep technologies cannot be used to diagnose or treat sleep disorders, and that patient-generated sleep data should not replace validated diagnostic testing. The academy's more constructive point is that such data can usefully enter a conversation with a clinician rather than substitute for one.

Nor does anything in this evidence set establish that your stage percentages predict how you will feel or function, or that nudging your deep-sleep number upward changes any outcome. Those questions were not asked here, and the honest answer is that they remain open. Note also what this literature conspicuously fails to deliver: a stable ranking. Device generation, algorithm version, cohort age and study design all move the numbers, which is why the same manufacturer looks competent in one study and mediocre in the next.

What to actually do with the thing on your finger

Buy for the part that validates. The sleep-versus-wake call is good enough to make a tracker an excellent ledger of behaviour: when you got into bed, when you got up, and how much those times wandered across a fortnight. That is a genuine measurement, and it happens to describe the lever most people can actually move.

• Look for published epoch-by-epoch polysomnography validation that reports specificity for wake, not just overall accuracy or sensitivity. Accuracy alone hides the bias.
• Prefer multi-night ambulatory validation over a single laboratory night. One night wired up in a lab is not a week in your own bed.
• Check the validation cohort's age range against your own. Error widened measurably in adults aged 56 to 80.
• Treat the stage breakdown as a low-resolution estimate. If a device's premium tier is priced on stage detail, you are paying for the weakest thing it does.

The interesting failure here is not technical, it is typographic. Four-stage accuracy of 53 percent is not scandalous for an optical sensor strapped to a finger; for the physics involved it is arguably impressive. What is indefensible is rendering that inference as a crisp "47 minutes" of deep sleep, a two-digit precision implying a measurement nobody made. These are honest instruments wrapped in dishonest design. Until the design catches up, read your own dashboard the way a clinician reads a screening test: note the direction, distrust the decimal, and remember that the only figure on the page with real evidence behind it is the boring one at the top.