A wearable reports a heart rate, a heart rate variability, a night of sleep split into stages and an estimate of aerobic fitness, and it reports all of them in the same confident typeface. The 2025 validation literature says they deserve very different levels of trust. Overnight resting heart rate from a good ring can be within two percent of an ECG. Wake time from a wrist device can be wrong more often than it is right. A VO2max estimate can be off by several points of fitness in a predictable direction. The useful response is to grade each reading separately, and to show the grade next to the number.
Overnight heart rate and HRV: close, for some devices
The most careful 2025 study of overnight cardiac metrics is by Dial and colleagues in Physiological Reports. Thirteen healthy adults slept with a Polar H10 ECG chest strap as the reference and several consumer devices at once, for a total of 536 nights.
For resting heart rate, the Oura Gen 4 ring reached a concordance correlation of 0.98 with a mean absolute percentage error of 1.94%, the Oura Gen 3 0.97 and 1.67%, the WHOOP 4.0 0.91 and 3.00%, and the Polar Grit X Pro 0.86 and 2.71%. The Garmin Fenix 6 could not be included, because the watch does not say when its 30-minute resting window falls, so its readings could not be lined up with the reference.
Heart rate variability spread the field further. Oura Gen 4 reached a concordance of 0.99 with a 5.96% error, Oura Gen 3 0.97 and 7.15%, WHOOP 0.94 and 8.17%, the Garmin Fenix 6 0.87 and 10.52%, and the Polar 0.82 with 16.32%. The standard deviations matter as much as the means: WHOOP's HRV error varied by more than its average, at plus or minus 10.49 percentage points, and Polar's by plus or minus 24.39. The authors' conclusion is the right one: devices vary, and validation has to be continuous, because firmware and hardware change.
Heart rate in motion: fine at steady state, slow on the ramp
Optical heart rate behaves differently when the body is changing pace. A 2025 study in Frontiers in Digital Health compared a Garmin Forerunner 45 with an ECG chest strap in 29 adults across rest, ramps and steady exercise, and across skin tones on the Fitzpatrick scale. At rest the mean difference was about one beat per minute; in steady exercise it was close to zero. During the first ramp-up in intensity the wrist read on average 14 beats per minute below the ECG, and 21 of 29 participants differed by more than five beats. Skin tone had no significant overall effect in this sample, with one moderate correlation during the second ramp.
The lesson is about timing as much as devices. A sensor that agrees at rest and at steady state can still lag badly on exactly the transitions a training session is built from.
Sleep: good at sleep, weak at wake
Sleep staging is where consumer claims and measured performance sit furthest apart. Schyvens and colleagues, in SLEEP Advances in 2025, put six wrist devices against polysomnography in 62 adults: the Fitbit Charge 5, Fitbit Sense, Withings ScanWatch, Garmin Vivosmart 4, WHOOP 4.0 and Apple Watch Series 8. All of them detected more than 90% of sleep epochs. Their specificity, the share of wake epochs correctly called wake, ranged from 29.39% to 52.15%. Agreement across stages, measured by Cohen's kappa, ran from 0.21 to 0.53, fair to moderate.
In plain terms, the devices are good at noticing that someone is asleep and poor at noticing that they are lying awake. Since total sleep time is sleep minus wake, that weakness flows straight into the headline number. The authors judge that the better devices can track prolonged and significant changes in sleep architecture, and that wearables do not stand in for polysomnography in diagnosis.
VO2max: a consistent underestimate
Aerobic fitness estimates rest on more modelling than any other wearable metric, because the watch never measures oxygen. Lambe and colleagues, in PLOS ONE in 2025, compared the Apple Watch's VO2max with laboratory testing and found that it underestimated by 6.07 mL/kg/min on average, with a mean absolute percentage error of 13.31% and limits of agreement from -6.11 to 18.26 mL/kg/min. They conclude the estimate needs refinement before clinical use, while recognising its practical value.
With limits of agreement more than 24 mL/kg/min wide, a single reading cannot separate a real change in fitness from the instrument.
Why every measurement needs its own grade
Put the studies side by side and a pattern appears. Error depends on the metric, the device, the moment and the person. Overnight resting heart rate on a good ring is close to the ECG. HRV on the same ring is good, and on other devices noticeably worse. Heart rate on the ramp of an interval session lags. Wake detection was weak on every wrist device tested. VO2max is biased in a known direction.
A single readiness score built on top of all of these hides that structure. It combines a strong signal with weak ones, takes a population model the wearer was never fitted to, and hands back one number with no error bar. Two consequences follow. A change in the score can be noise in the weakest input. And a person trying one change, a later training session, less alcohol, a different bedtime, cannot tell whether it worked.
The alternative is to grade each reading by what is known about that metric on that device, and then to measure the person against their own normal range rather than a population. A reading with a tight grade can support a decision. A reading with a loose one can show a trend and nothing more. And when the evidence is not strong enough for an answer, the honest output is to say so.
Where Livgens fits
Livgens represents the thing exactly, in sound, light, matter and motion, and Living Worlds runs it forward through its lifecycle; a body, measured night after night, is one of those things. Livgens Health is live at app.livgens.com. It connects the ring or watch a person already wears, with no new hardware and nothing to log. After eight nights it knows that person's own normal range, and when they change one thing, some days on and some off, it reports the size of the effect on them and how sure it is, or says the effect is too small to bother with. Its example reads: evening training is costing 3.9 bpm, between 2.9 and 5.0, on that person's own nights at their own measurement error.
Grades are shown, not hidden. Device errors are taken against an ECG criterion. Gait speed and heart-rate variability are shown as a trend, never as a verdict. Per-food verdicts from one continuous glucose monitor pass are refused, because around 90% of individual food effects fall inside the within-person noise of a single pass. Where a reference instrument exists, Health holds to it: VO2max by true CPET against roughly plus or minus 12 mL/kg/min for a wearable estimate, and a countermovement jump on a force plate at 2.79 cm against 8.63 cm for a wearable sensor.
Tested on real people across 1,623 nights of measurement, all 13 ended up with a threshold around five times tighter than the device figure: 2.5 beats against 13. Readings are opened with a passkey and forecast on the person's own phone or laptop, and the analysis was pre-registered, with every threshold and decision rule fixed in advance.
Sources
- Dial et al., Validation of nocturnal resting heart rate and heart rate variability in consumer wearables (Physiological Reports, 2025)
- Schyvens et al., A performance validation of six commercial wrist-worn wearable sleep-tracking devices for sleep stage scoring compared to polysomnography (SLEEP Advances, 2025)
- Lambe et al., Investigating the accuracy of Apple Watch VO2 max measurements: A validation study (PLOS ONE, 2025)
- Icenhower et al., Investigating the accuracy of Garmin PPG sensors on differing skin types based on the Fitzpatrick scale (Frontiers in Digital Health, 2025)
