REF-1249 · Zhang et al., 2020 · Journal of Sports Sciences · Paper · PubMed

Is my watch's wrist heart rate accurate on the bike?

On average it was off by about 4.5 bpm cycling and 7 bpm lifting (probably low), but close at rest, in sleep and on the treadmill (44 studies). One session's error is not given.

B+moderate evidenceMeasured on non-athletes and recreational people

Zhang and colleagues pooled studies that compared heart rate from wrist-worn optical (photoplethysmography) devices with a criterion of ECG or a chest strap. From five databases they included 44 articles with 738 effect sizes across 15 brands and analysed them with multilevel random-effects models by activity.

The mean difference was small during sleep (−0.40 bpm, 95% CI −1.64 to 0.83), at rest (−0.01, −0.02 to 0.00) and on the treadmill from walking to running (−0.51, −1.60 to 0.58), larger in cycling (−4.55, −7.24 to −1.87) and largest in resistance training (−7.26, −10.46 to −4.07), where the mean difference "increased by 3 bpm (2.5 to 3.5)" per 10 bpm rise in heart rate (whether the gap widened or narrowed is not stated).

The authors conclude that wrist devices show acceptable validity for most common activities. Our reading: on average a wrist sensor is close in steady, rhythmic or still conditions and reads low when the hands grip or the wrist flexes; the abstract reports group means only, so it says nothing about how far a single reading may stray — the number an athlete actually sees.

Protocol

Design (from the abstract): systematic review and multilevel random-effects meta-analysis; searches of PubMed/Medline, Web of Science, EBSCOhost, PsycINFO and EMBASE.
Included: 44 articles, 738 effect sizes, 15 brands of wrist-worn optical devices; criterion ECG or chest strap.
Who: not described in the abstract; REF-1248 lists the review as covering healthy adults, n = 1738 (its Table 1; not from this abstract).
Analysis: mean difference in bpm with 95% CI by activity (sleep, rest, treadmill walking to running, cycling, resistance training); a moderator analysis of heart-rate level in resistance training.
Funding and interests: not reported in the abstract or in the PubMed record; the full text was not available to us, so they could not be checked.

Key Results

Mean difference of wrist optical heart rate from the criterion (bpm, 95% CI; from the abstract):

ConditionMean difference
Rest−0.01 (−0.02 to 0.00)
Sleep−0.40 (−1.64 to 0.83)
Treadmill, walking to running−0.51 (−1.60 to 0.58)
Cycling−4.55 (−7.24 to −1.87)
Resistance training−7.26 (−10.46 to −4.07)
Resistance training, per 10 bpm higher heart ratethe mean difference "increased by 3 bpm (2.5 to 3.5)"
Sign convention: not stated in the abstract. We read negative as the wrist reading below the criterion, which matches the under-reading in cycling that the Fitbit meta-analysis in REF-1248 reports; the abstract's "increased" for resistance training is ambiguous: under that convention a rising mean difference moves toward zero (a smaller gap at higher heart rates), unless the authors meant the absolute gap.
Not in the abstract: limits of agreement, heterogeneity, risk of bias of the included studies, participant numbers, device models, skin tone, exercise intensity other than the resistance-training moderator.

Mechanisms

Measured: the gap from the criterion depended on the activity and, in resistance training, on how high the heart rate was.
Authors' interpretation: none beyond "acceptable validity compared to a criterion measure of HR for most common activities" is given in the abstract.
Inference: gripping a handlebar or a bar and flexing the wrist change blood flow and sensor contact under an optical sensor, while running and sleep keep the wrist relatively still or rhythmic; this fits the pattern but was not tested here.
Inference: a mean difference near zero for running can hide errors in both directions; REF-1248's narrative estimate of about ±3% and the Fitbit limits of agreement there are a closer guide to one session than this abstract's means.

Practical Application

Is my watch's wrist heart rate accurate on the bike? On average it was off by about 4.5 bpm cycling and 7 bpm lifting (probably low), but close at rest, in sleep and on the treadmill (44 studies). One session's error is not given.
For steady runs, resting heart rate and sleep, a wrist reading is close on average (mean differences −0.01 to −0.51 bpm; B+, from the abstract); expect individual sessions to vary more than the mean.
On the bike, read heart-rate zones, decoupling and the last minute of a fatigue test from a chest strap (Not from this paper; the base's own practice, motivated by the −4.55 bpm cycling gap here and the near-ECG agreement of a Polar H10 in REF-1250); this applies to SYNTH-41's heart rate over the last minute of a 3-min effort and to the heart-rate rows of SYNTH-38 and SYNTH-40.
In strength sessions, ignore wrist heart rate as a load measure: it read about 7 bpm off on average (probably low), and the gap changed with heart rate in a direction the abstract does not state.
A wrist-derived threshold or VO2max may inherit this error (inference): REF-1232 found watch threshold heart rate about 9-11 bpm off a lab value, and REF-1253 found wider VO2max limits of agreement with wrist optical than with chest-strap heart rate.

Limitations

Abstract only. No full text was available through Europe PMC, the publisher's open-access pages or an OpenAlex repository copy, so heterogeneity, risk of bias, participant numbers, device models, limits of agreement, funding and conflicts of interest could not be checked; nothing beyond the abstract is claimed for this paper.
Group means only. Mean differences with confidence intervals describe the average device in the average study; an athlete's single reading lives inside the limits of agreement, which the abstract does not give.
Sign convention and wording. The direction of the differences and the meaning of "increased by 3 bpm" are our reading.
Old devices. Studies published before 2020, devices without model names (REF-1248); accuracy of one model is not the brand's, and firmware has changed since.
Conditions pooled. "Treadmill activities (walking to running)" and "cycling" each pool intensities; intervals, transitions, outdoor riding on rough surfaces and skin tone are not separated in the abstract.
How this entry was checked and changed
Why this level

(from the abstract) — a systematic review and meta-analysis of 44 validation studies, held below the rubric's A- because the umbrella review in this batch rated it low quality on the Drucker tool (no registered protocol, no grey-literature search, precision and consistency not discussed; REF-1248, Table 2), and because heterogeneity, risk of bias, limits of agreement, participant numbers and funding are not in the abstract, the sign convention of the mean differences is not stated there, and validation studies of this kind test one device generation in a single session (no full text: not in Europe PMC, no open-access or repository copy listed by OpenAlex)