Is my watch's VO2max accurate?
On average, yes after runs (bias −0.1 mL/kg/min), but for one person about ±10 off; resting-test estimates −13 to +17. Read it as a range; whether it tracks change is barely tested. 14 small studies.
The INTERLIVE network (six European universities and Huawei Technologies) pooled every study that compared a consumer wearable's VO2max estimate with a graded exercise test to exhaustion with gas analysis.
Fourteen studies (403 participants; 218 men, 185 women) met the criteria: eight tested estimates built from a resting measurement (Polar's fitness test: resting heart rate and HRV with age, sex, height, weight and activity level) and eight tested estimates built from a run (Garmin and Fitbit, using Firstbeat's heart-rate-to-speed method); two studies tested both.
Resting-based estimates overestimated VO2max on average (bias +2.17 mL/kg/min, 95% CI 0.28 to 4.07) with pooled limits of agreement of −13.07 to 17.41; exercise-based estimates had no meaningful bias (−0.09, −1.66 to 1.48) and limits of −9.92 to 9.74.
Heterogeneity was considerable, the resting result was not robust to leaving out single studies, and 11 of 14 studies were at high risk of bias, mostly in how the laboratory test was run and processed. Estimates using wrist optical heart rate had greater bias (resting) and wider limits (exercise) than those using a chest strap.
The authors judge exercise-based estimates fit for population-level use but "large" in error for an individual, and would not advise relying on resting-based estimates; they then set out best-practice validation rules in six domains. A correction (13 October 2026) reports three duplicated reference-list entries and confirms that all in-text citations, and so all numbers, stand.
Our reading: the source the base lacked for watch VO2max — a watch's number after runs is about as uncertain for one person as the power-based estimates in SYNTH-78 (about ±10 mL/kg/min), resting-test numbers more so; read it as a range, and its change over weeks as a direction, never as a measurement.
Protocol
Search: PubMed, Web of Science and Scopus to 14 January 2021, plus a Google Scholar hand search; strategy defined by the network; protocol registered in PROSPERO (CRD42021246192).
Eligibility: any population; VO2max estimated by a consumer wearable and measured by a graded exercise test to exhaustion with gas analysis (large-muscle-group mode); criterion-validity studies; research devices, grey literature and algorithms not in a commercial product excluded.
Selection and extraction: two authors independently (Covidence), a third for disagreements; 1224 records, 14 studies included, n = 403 (mean 29 per study).
Studies: resting-based — Polar A300, S410, F11, FT40, F6, RS300X and V800 (Table 1 shows one study with wrist optical heart rate and seven with chest straps; the text speaks of two optical studies), 5 min supine or seated rest in most; exercise-based — Garmin Fenix 3, Fenix 5X, Forerunner 230, Forerunner 920XT and "GF5", and Fitbit Charge 2, from 10-45 min runs (wrist optical heart rate in four studies, chest strap in three, not stated for one).Reference: indirect calorimetry in all, Parvo Medics in 10 studies, Cosmed 2, Metalyzer 2; all treadmill protocols.
Risk of bias: COSMIN-adapted, four domains (participants, index measure, reference, statistics); any "no" or more than two "unclear" = high risk.
Pooling: bias = wearable minus reference (positive = overestimate), inverse-variance random effects in RevMan 5.3; pooled limits of agreement by Tipton and Shuster, δ ± 2√(σ² + τ²), in R. For five studies the bias was derived from correlation coefficients or digitised from plots. Sensitivity analyses: leave-one-out, funnel plot and Egger test, wrist optical vs chest-strap heart rate, and studies from the last 3 years.
Funding and interests: partly funded by Huawei Technologies Oy (Finland), a member of the INTERLIVE consortium; the authors declare no conflicts of interest. No Huawei device was among the included studies (our reading of Table 1). Not declared here but printed in REF-1248's byline and Table 2: two authors (Argent, Caulfield) later co-wrote that umbrella review, which rated this review high quality.
Correction (found through the article's Crossref "updated-by" record; text from the publisher's free PDF): references 37 and 41, 15 and 52, and 38 and 51 are duplicates of each other; the authors confirm all in-text citations remain accurate. No number in this entry is affected.
Key Results
| Method | Studies (devices) | Pooled bias, mL/kg/min (95% CI) | Pooled limits of agreement (95% CI of the limits) | I² |
|---|---|---|---|---|
| Resting-based | 8 (Polar) | +2.17 (0.28 to 4.07), P = 0.020 | −13.07 to 17.41 (−22.18 to 26.53) | 77% |
| Exercise-based | 8 (Garmin, Fitbit) | −0.09 (−1.66 to 1.48), P = 0.910 | −9.92 to 9.74 (−16.79 to 16.61) | 88% |
Robustness: the resting result changed significantly when any of five individual studies was removed; the exercise result was robust. No publication bias was found by funnel plot or Egger test.
Wrist optical vs chest strap: in resting tests, studies with wrist optical heart rate had significantly greater bias (only two such studies); in exercise tests the limits-of-agreement span was ±23.03 with optical vs ±17.97 mL/kg/min with a strap.
Recent studies only (resting): excluding five older articles gave a bias of 1.66 (95% CI −0.58 to 3.90), P = 0.150.
Against the makers' claims: Firstbeat's white paper states 5% MAPE for running; the four included studies of Garmin and Fitbit that reported MAPE found 8 to 10.2%. Polar states a mean error of 8% (about 3.7 mL/kg/min) to 15%.
For planning studies: expected mean absolute difference 2.30 mL/kg/min (SD 7.20) for resting and 1.32 (SD 4.03) for exercise-based estimates.
Reliability: only three studies tested it (r and ICC above 0.90), none with limits of agreement.
What counts as meaningful (authors' context): increases of 1.75-3.5 mL/kg/min have been associated with lower mortality and cardiovascular risk, so errors beyond 3.5 "will be missing clinically relevant changes".
Numbers that do not add up (our arithmetic):
The "±" given for single studies (resting ±17.75 to ±38.97, exercise ±11.18 to ±23.53) and for the optical-vs-strap split (±23.03 vs ±17.97) cannot be half-widths: every one is wider than the pooled ±15.24 or ±9.83, and a Tipton–Shuster pooled limit cannot be narrower than all its studies. Read as full widths (upper minus lower) they fit the pooled full widths of 30.48 and 19.66, and the Discussion itself writes "± 30.48" for the resting span.
"A significant reduction in the estimation errors" after excluding older resting studies: 2.17 to 1.66, with no test of the change and a CI that includes zero; the Discussion then reads the 0.51 difference as support that newer devices are more accurate.
"Five studies did not apply an appropriate analysis" cites six references; "five studies [29–31]" cites three. One included study is listed as 65 participants, "men (27) and women (33)" (60).
The umbrella review REF-1248 restates these limits in its abstract as VO2max "overestimated … by ± 15.24% … and ± 9.83%"; they are half-widths in mL/kg/min.
Mechanisms
Measured: estimates from running agreed better with the laboratory than estimates from rest, and chest-strap heart rate gave narrower agreement than wrist optical heart rate.
Authors' interpretation: Polar's resting test feeds resting heart rate, HRV, sex, age, height, weight and self-reported activity into a neural network, and Firstbeat's method relates heart rate to running speed over the most reliable segments of a run; they call the chest strap's advantage expected because its error against ECG is minimal.That the resting method says little about one person's aerobic capacity, and the running method rests on measured exercise physiology, is our inference, not the authors' wording.
Inference: an exercise-based estimate is a heart-rate-to-speed relation, so anything that changes heart rate at a given pace (heat, hills, wind, fatigue, caffeine, dehydration, wrist-sensor error — REF-1249) or the measured speed (GPS) moves the number without any change in fitness. Run on the same watch in similar conditions, part of the error repeats from week to week and a direction becomes readable; the random part does not.
Inference: the 95% CI of the exercise limits (−16.79 to 16.61) means a new device study could plausibly find individual errors up to about ±17 mL/kg/min.
Practical Application
Is my watch's VO2max accurate? On average, yes after runs (bias −0.1 mL/kg/min), but for one person about ±10 off; resting-test estimates −13 to +17. Read it as a range; whether it tracks change is barely tested. 14 small studies.
Quote a watch VO2max with its band: ±10 mL/kg/min for an estimate from runs, about −13 to +17 for a resting test (A-, healthy young adults, older devices). This is the source for SYNTH-78's line "The number on a watch or head unit is an estimate of the same kind and deserves the same caution", and the band matches the power-based estimates there (REF-1016, REF-1017).
Do not compare two athletes by watch VO2max, call a change of a few mL/kg/min progress, or set zones from it — the same three rules SYNTH-78 sets for power-based estimates; the authors put meaningful change at 1.75-3.5 mL/kg/min, below the individual error, and test–retest reliability is barely studied.
If the number matters, prefer the exercise-based estimate, from steady outdoor runs, with a chest strap (the narrower limits in this review are for runs and straps; steady conditions are the base's inference), and send anyone who needs the real value to a laboratory test (SYNTH-78; even the laboratory reference carries error, REF-169).
A cycling athlete's watch VO2max is untested here: every exercise-based estimate in this review came from running.
Limitations
Small, biased studies: 14 studies of 29 participants on average; 11 at high risk of bias and 3 with some concerns, mostly in the laboratory protocol and data processing; for five studies the bias was reconstructed from correlations or plots.
Heterogeneity considerable (I² 77% and 88%); the resting result was not robust.
Method confounded with brand: resting = Polar, exercise = Garmin and Fitbit; "exercise beats rest" cannot be separated from "these brands beat that one".
Old devices, one activity: studies 2004-2020, all estimates from running, none from cycling; firmware not reported in the review; healthy, active adults aged about 25.
Reliability almost untested, so whether a watch tracks one person's change is unknown.
Reporting: "±" used for both half-widths and full widths; a "significant reduction" without a test; citation counts that do not match.
Interests: partly funded by Huawei, a member of the consortium that wrote the review; no Huawei device was evaluated and no conflicts were declared. The 2026 correction concerns duplicated references only.
How this entry was checked and changed
Why this level
a meta-analysis of 14 small validation studies (29 participants on average) with considerable heterogeneity (I² 77% resting, 88% exercise) and 11 of 14 studies at high risk of bias, read from the full text (open access, CC BY 4.0) and the correction notice. The pooled limits of agreement, not the near-zero bias, are the finding an individual can use
Full text (checked 2026-10-27)
14 articles included; a total sample size of 403 participants (218 men and 185 women), a mean sample per article of 29 participants; pooled age of 24.6 ± 5.7 years (full text)
Resting test: bias = 2.17 ml·kg−1·min−1; 95% CI 0.28–4.07; P = 0.020; limits of agreement − 13.07 to 17.41 (95% CI − 22.18 to 26.53) (full text)
Exercise tests: bias = − 0.09 ml·kg−1·min−1; 95% CI − 1.66 to 1.48; P = 0.910; limits − 9.92 to 9.74 (95% CI − 16.79 to 16.61) (full text)
Considerable heterogeneity: Higgins I2 = 77% and 88% in resting and exercise test, respectively (full text)
Overall risk of bias "some concerns" for three (21%) and "high" for 11 (79%) of the 14 studies (full text)
Excluding five articles from more than 3 years ago in the resting conditions: bias = 1.66 ml·kg−1·min−1; 95% CI − 0.58 to 3.90; P = 0.150 (full text)
PPG vs chest strap in the exercise tests: limits of agreement span ± 23.03 vs ± 17.97 ml·kg−1·min−1 (full text)
Firstbeat white paper: 5% MAPE for running; four included studies of Fitbit and Garmin found values ranging from 8 to 10.2% (full text)
Increases in VO2max of 1.75–3.5 ml·kg−1·min−1 are associated with a lower risk of all-cause mortality (full text)
This research was partly funded by Huawei Technologies Oy (Finland) Co. Ltd.; none of the authors has any conflict of interest to declare (full text)
Correction: references 37 and 41, 15 and 52, and 38 and 51 refer to the same articles; all in-text citations remain accurate (full text)