How accurate is wrist heart rate compared with an ECG chest strap?

This report does not have a run number

Method clause M-07 covers heart rate against a 3-lead ECG harness. The harness exists and passed its competency check on 2026-04-21. The clause has said drafted ever since, because no heart-rate run has published on this desk.

So nothing below is our measurement. This is other people’s data, read under clause M-09: every figure names whoever produced it, the reference it was measured against, and the sample size, in the same row as the number. Sources checked 2026-08-10.

We are publishing it because the published record already answers this question more clearly than most buying advice admits, and because the answer does not depend on any figure we would have to generate ourselves.

What the numbers look like over a mixed protocol

The largest single comparison we found put ten wearables through the same protocol against a chest-worn ECG reference, with 45 participants aged 21 to 68.

Gielen et al., JMIR Formative Research, 2026;10:e85186. n = 45. Reference: chest-worn ECG-derived heart rate. The study also covered three devices outside this report’s scope, including a ring; the full table is in the paper.

Two things to take from that. The spread between the best and worst wrist device is three times the size of any difference between brands people usually argue about. And even the best of them, at 4.5 bpm mean absolute error, is not a device you would use to hold a target heart rate to within five beats.

The failure is rate of change, not intensity

This is the part most buying guides get backwards. Wrist optical does not degrade smoothly as you work harder. It degrades when your heart rate moves quickly, and it recovers when the rate settles, however high it has settled.

A 28-participant study measured that directly by putting the same devices through rest, steady cycling, a graded treadmill protocol, and 30 seconds of burpees followed by a minute of recovery, all against a Polar H10 reference.

Moghaddam et al., Sensors, 2025;26(1):176. n = 28. Reference: Polar H10. Values are MAPE ± standard deviation.

The graded treadmill column is the interesting one. It is the hardest work in the protocol and the error is close to the resting figure, because a treadmill ramp changes heart rate slowly and holds the arm in a repeatable position. The burpee column is much easier work and the error is five to twelve times higher, because heart rate is climbing and falling fast and the wrist is being thrown around.

A caution on that last column, from the desk’s own standing rule under M-03: the standard deviations there are larger than the means. Read literally, those five devices are not separable from each other during burpees. What the column supports is that every one of them got much worse; it does not support ranking them against each other, and anyone quoting a burpee figure as a device comparison is over-reading it.

A separate 24-participant study using a 12-lead ECG reference found the same direction of travel across a 20-minute protocol with five deliberate intensity transitions: every wrist device performed worse during the transitions than during steady states.

Van Oost et al., Sensors, 2025;25(20):6319. n = 24. Reference: 12-lead ECG.

That study reported the steady-versus-transition drop as statistically significant for three devices and not for the other three. It is worth being careful about what that means. Significance there partly tracks how tight a device’s error distribution already was, not how good the device is: the tightest performer in the set is the one where a small degradation was easiest to detect, while the noisiest devices had error distributions wide enough to absorb the change without reaching significance. A non-significant result in that table is not a clean bill of health.

Where the wrist is worst is not where you would guess

Two findings in this literature cut against intuition, and both are consistent across studies.

Low heart rate is hard, not easy. One wrist watch measured against a Polar H10 across nine activities showed within-subject coefficients of variation of 23.0% to 26.4% while the participant was lying down — its worst condition in the entire protocol, worse than jogging or high-intensity interval work. A low, weak pulse gives the sensor less signal to work with.

Walking is hard too. In the ten-device study above, intermittent walking produced the largest errors in a subset of the devices. Cycling, by contrast, was consistently among the best cases in every study we read, for the obvious reason that the hand is clamped to a bar and barely moves.

The variable is not effort. It is signal amplitude and motion artifact.

The cold-skin claim, which we could not support

We expected to write that cold skin degrades wrist heart rate, since the mechanism is textbook: cooling causes peripheral vasoconstriction, which reduces the optical pulse amplitude the sensor depends on. That part is established.

The device-level claim is not. The ten-device study tested all of them at 10 degrees C, 23 degrees C and 36 degrees C, and found no statistically significant difference in accuracy between climate conditions after correction for multiple comparisons. The authors offer two candidate explanations: 10 degrees C in a short indoor protocol may not be a hard enough cold exposure, or the devices’ signal processing may already compensate.

Either way, we are not in a position to tell a reader that their watch will read badly in the cold, and neither is anyone else who is working from published data. It is a plausible mechanism with a negative controlled result attached to it. It stays on the list of things clause M-07 should measure rather than in the recommendations below.

Why the strap is the reference

A chest strap is not merely a better version of a wrist sensor. It measures a different thing. It picks up the electrical depolarisation that drives the contraction, which is a large, sharply timed signal, instead of inferring the beat from a small optical fluctuation several centimetres from any significant artery.

The Polar H10 in particular is what most of the studies cited on this page used as their criterion device. Measured against a clinical Holter recorder, its RR-interval signal quality was 99.6% overall and held at 99.4% during high-intensity activity, while the Holter itself fell from 94.6% to 89.8% (Gilgen-Ammann et al., European Journal of Applied Physiology, 2019;119:1525-1532). Correlation between the two systems was r = 0.997.

That is the actual reason the strap wins: its error does not care what your heart rate is doing.

What to buy

If you own a watch and mostly want trends, keep it. If you are training to heart-rate zones and the session has repeated hard efforts, the watch will under-read your starts and over-read your recoveries, and a strap is the fix. Position matters more than brand: moving the same optical sensor from wrist to upper arm improved accuracy more than switching between wrist brands did.

What would change our mind

The honest gap is that we are reporting rather than measuring, and every limitation of these studies is inherited. Sample sizes run from 6 to 62 people. Skin tone, tattoo coverage and wrist circumference all affect optical sensing, and none of the studies here was powered to characterise that. Several use different references and different protocols, which is exactly why we have not built a single league table across them.

Three results would move us. A cold-exposure protocol harder than 10 degrees C, which is the open question this page could not close. A transition-specific dataset with per-participant residuals published, so the burpee-style figures could be re-derived rather than taken on trust. And our own run under M-07, which is the only one of the three we control.

We have also not price-checked any of these devices. The first sweep under M-08 covered a different category, and quoting a figure off a marketing page is exactly what that clause forbids. When we check sensor prices at the point of purchase, they will appear here with a date.