Can I trust my watch's calories for a run?
Only roughly: over 2 km an Apple Watch S6 and a Garmin Fenix 6 read 17-23% high on average; by our arithmetic single runs could read 51-57% high, the funder's Huawei GT 2e ±29% either way.
Le et al. asked how well three smartwatches of early 2021 estimate the energy spent walking and running outdoors. Twenty healthy young adults wore an Apple Watch Series 6, a Garmin Fenix 6 and a Huawei Watch GT 2e at the same time (two on one wrist, one on the other, sides randomised) while a COSMED K5 portable metabolic system measured expenditure breath by breath.
Each walked 2 km at about 6 km/h and, after a 10-min seated break, ran 2 km at about 10 km/h (measured 10.94 km/h) on a 400-m track. Against about 109 kcal measured for the walk and 112 kcal for the run, the Apple watch read 20.5 and 25.7 kcal high on average and the Garmin 31.0 and 19 kcal high; the Huawei watch was within 2 kcal on average.
The mean absolute percentage error was 19.8% (Apple), 32.0% (Garmin) and 9.9% (Huawei) for walking and 24.4%, 21.8% and 11.9% for running; the authors' validity cut-off was 10%. Single bouts scattered widely around those averages: the limits of agreement span 44-150 kcal on bouts of about 110 kcal.
Our reading: a small, one-visit study that gives the base its first direct measurement of watch calories; it says that even the best of the three watches missed a single run by up to about 30%, that two popular watches read high, and nothing about cycling. The best-performing watch's maker paid for the study.
Protocol
Design: one-visit concurrent validation; approved by the Institutional Review Board for Human Research Protections of Shanghai Jiao Tong University (B2020024I).
Who: n = 20 healthy Chinese adults (10 men, 10 women), recruited on a university campus; age 23.0 ± 2.2 y, 67.1 ± 11.1 kg, BMI 22.1 ± 1.9, body fat 19.2 ± 6.9% (InBody 720); inclusion 18-30 y, BMI 18.5-25, three or more exercise sessions a week.
Devices: Apple Watch Series 6, Garmin Fenix 6, Huawei Watch GT 2e, bought commercially and described as the newest generations at January 2021; each uses wrist photoplethysmography for heart rate and GPS for distance and speed; the outdoor walk or run mode was started and stopped at the same time on all three; height, weight, sex and date of birth entered. Firmware or software versions are not reported.
Criterion: COSMED K5 breath-by-breath portable gas analysis, warmed up 15 min and calibrated with reference gases and a 3-L syringe before each test; expenditure summed per bout.
Exercise: 2 km walking at a target 6 km/h (achieved 6.15 ± 0.29 km/h) and, after 10 min seated, 2 km running at a target 10 km/h (achieved 10.94 ± 0.99 km/h), steady self-selected pace, 400-m track; 16.5 ± 2.6 °C and 13.9 ± 2.9 °C. Six hours without food or stimulants and 24 h without vigorous exercise or alcohol before testing.
Watch values: read from screenshots of each watch's app (cumulative totals); whether these were total or "active" calories is not stated. The Huawei watch gave 19 of 20 values in each bout ("technical errors"; whether the same participant both times is not stated).
Analysis: paired t-tests, mean absolute percentage error (MAPE, cut-off ≤10% for validity), ICC, Bland-Altman 95% limits of agreement (SPSS 24); post-hoc power computed per device.
Funding and interests: "supported by Huawei Terminal Co. LTD"; the authors state that the funder was not involved in the design, collection, analysis, interpretation, writing or the decision to publish, and that they "declare no other competing interests"; a separate conflict-of-interest statement says the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Key Results
Energy expenditure per 2-km bout, watch vs K5 (Table 2 and text; kcal as mean ± SD; difference = watch minus K5):
| Bout | Device | Watch (kcal) | Difference (kcal) | Bias, % of K5 | MAPE | ICC | Limits of agreement, kcal (our arithmetic) |
|---|---|---|---|---|---|---|---|
| Walk (K5 108.7 ± 17.4) | Apple Watch S6 | 129.1 ± 20.1 | +20.5 ± 11.3 (p < 0.001) | +18.9% | 19.8% | 0.821 | −1.6 to +42.6 |
| Garmin Fenix 6 | 139.6 ± 39.6 | +31.0 ± 38.3 (p = 0.002) | +28.5% | 32.0% | 0.216 | −44.1 to +106.1 | |
| Huawei GT 2e (n = 19) | 111.2 ± 18.1 | +2.0 ± 12.4 (p = 0.491) | +1.8% | 9.9% | 0.760 | −22.3 to +26.3 | |
| Run (K5 112.2 ± 21.3) | Apple Watch S6 | 137.8 ± 23.1 | +25.7 ± 16.0 (p < 0.001) | +22.9% | 24.4% | 0.741 | −5.7 to +57.1 |
| Garmin Fenix 6 | 131.1 ± 28.8 (text) | +19 ± 22.8 (p = 0.001) | +16.9% | 21.8% | 0.594 | −25.7 to +63.7 | |
| Huawei GT 2e (n = 19) | 111.8 ± 21.1 | −0.3 ± 16.7 (p = 0.946) | −0.3% | 11.9% | 0.698 | −33.0 to +32.4 |
The paper reports each device's limits of agreement as one number — 44.1, 150.1 and 48.6 kcal for walking and 62.8, 89.4 and 65.6 kcal for running.
These equal the full width of the band (2 × 1.96 × SD of the differences), so the limits in the last column are the mean difference ± 1.96 SD (our arithmetic); the bias percentages are the mean difference over the K5 mean (our arithmetic). Values inside the limits: 100%, 90% and 95% of walks and 95%, 95% and 100% of runs (Apple, Garmin, Huawei).
Direction: Apple and Garmin over-read in both bouts (significant); Huawei did not differ on average.
Agreement by the authors' categories: good for Apple (ICC 0.821) and Huawei (0.760) in walking, moderate for Apple (0.741) and Huawei (0.698) in running, low for Garmin (0.216 walking, 0.594 running).
Only one MAPE met the 10% cut-off: Huawei walking, 9.9%.
Per run in percent (our arithmetic): a single run's reading could be about 5% low to 51% high on the Apple watch, 23% low to 57% high on the Garmin and 29% either way on the Huawei.
Power: the authors' post-hoc power for the Huawei comparisons was 0.105 (walking) and 0.051 (running), against 0.93-1.00 for Apple and Garmin, so "no significant difference" for Huawei is not evidence of agreement; its MAPE and limits of agreement are.
Mechanisms
Measured: none of the three watches measures energy; each estimates it, and the estimate missed in a model-specific way — two read high on average, one did not, and all three scattered widely from bout to bout.
Authors' interpretation: the algorithms are not published but probably combine accelerometry, wrist heart rate and GPS speed; error can come from the heart-rate input (motion artefact), from the algorithm, or both, and the study cannot tell which.They write that a maker based in China and a Chinese sample likely explain the Huawei watch's accuracy, citing a study in which activity-monitor accuracy differed by ethnicity; that its algorithm was built on similar people is our gloss, not theirs.
Inference: a watch's calorie number rests on the profile entered (weight, age, sex) and on heart rate, so anything that moves heart rate without moving energy cost (heat, dehydration, caffeine, drift late in a long session) or a wrist heart-rate error moves the calories with it; in cycling, where wrist heart rate under-reads (REF-1249), the bias may differ from the running bias here. Not tested in this study.
Inference: the sign of the error is not general. The umbrella review REF-1248 reports energy expenditure under-read on average (about 3 kcal/min in a Fitbit meta-analysis), while here two watches read high in steady walking and running; the direction belongs to the device, the activity and the algorithm version, so neither "watches over-count" nor "watches under-count" holds as a rule.
Inference: the bouts were about 11-20 min of steady exercise. Whether the percentage error holds, grows or shrinks over a 2-h session is not known from this study; if the running bias were proportional, a +17-23% over-read would add roughly 120-160 kcal to a 700-kcal run (our arithmetic, not measured).
Practical Application
Can I trust my watch's calories for a run? Only roughly: over 2 km an Apple Watch S6 and a Garmin Fenix 6 read 17-23% high on average; by our arithmetic single runs could read 51-57% high, the funder's Huawei GT 2e ±29% either way.
Do not fuel or diet by the number on the watch (B-, 20 young adults, 2-km bouts): two of three watches over-read, and the best one still missed single runs by up to about 30%. In a deficit, an over-read means eating back calories that were never spent.This closes, for running and walking only, the gap SYNTH-87 states (no test of watch estimates of exercise expenditure); the rule there — judge intake by the body-mass trend, performance and recovery, not by a calorie target — stands.
On the bike, prefer the power meter's work (Not from this paper; the base's own practice): a power meter measures the work done in kJ, and at a gross efficiency of about 24% 1 kJ of work costs about 1 kcal (arithmetic; gross efficiency itself varies between people, equations and intensities, REF-489). The study tested no cycling.
If the watch is all there is, use it within itself: the same watch on the same kind of session gives a consistent direction for comparing one week with another (inference: reliability was not tested in this study). Do not compare calories across brands — here the gap between brands on the same run (Apple 137.8 vs Huawei 111.8 kcal on average) was larger than the average error of the best watch.
Say which watch and which model when citing this: the numbers belong to an Apple Watch Series 6, a Garmin Fenix 6 and a Huawei Watch GT 2e on early-2021 software; newer models and firmware were not tested.
Limitations
Small and short. n = 20, one visit, one 2-km walk and one 2-km run at a steady pace; no repeat visit, so the watches' reliability is unknown (the authors' limitation).
One population and one kind of exercise. Young, normal-weight Chinese adults; steady walking and running only; cool weather. The authors say the results may not transfer to other activities or groups.
What the watch number was is not stated. Values were read from app screenshots; whether they were total or "active" calories, and which firmware each watch ran, is not reported. The K5 measures total (gross) expenditure.
Reporting errors. Table 2 prints the Garmin running mean as 137.8 ± 23.1 kcal, identical to the Apple row; the text gives 131.1 ± 28.8, and the difference column (19 ± 22.8) fits the text. The limits of agreement are reported as a single width rather than as lower and upper limits.The Discussion calls the MAPE values (19.8%, 32.0%, 24.4%, 21.8%) the size of the over-estimation; the mean over-reads are somewhat smaller (about 17-29%, our arithmetic). Two sentences in the methods and limitations are garbled (the ICC thresholds; the body-side effect).
Underpowered for the device that won. By the authors' own post-hoc calculation, power for the Huawei comparisons was 0.105 and 0.051.
Funding and interests. Funded by Huawei Terminal Co. LTD, maker of the best-performing watch; the authors state the funder had no role, but the conflict-of-interest statement ("absence of any commercial or financial relationships") does not acknowledge the funding, although the funding statement itself ends "All authors declare no other competing interests." The watches were bought commercially.
How this entry was checked and changed
Why this level
single-lab validation study, n = 20, one 2-km walk and one 2-km run, read from the full text (open access, CC BY 4.0). The rubric places a validation study at B; held at B- because the study was funded by the maker of the watch that came out best (Huawei) while the conflict-of-interest statement declares no commercial or financial relationship, and because the Garmin running row of the results table duplicates the Apple row
Full text (checked 2026-10-27)
Twenty healthy Chinese subjects (10 males and 10 females); inclusion 18–30 years, body mass index 18.5–25 kg/m2, ≥3 weekly exercise sessions (full text)
Apple Watch Series 6 (AW), Garmin Fenix 6 (GF), and Huawei GT 2e (HW), "all of which were the newest generations introduced in January 2021"; "All watches used in the study were bought commercially" (full text)
Walking speed 6.15 ± 0.29 km/h and running speed 10.94 ± 0.99 km/h over 2 km each; 16.5 ± 2.6°C and 13.9 ± 2.9°C (full text)
Walking: K5 108.7 ± 17.4 kcal; AW 129.1 ± 20.1, GF 139.6 ± 39.6, HW 111.2 ± 18.1 kcal; MAPE 19.8%, 32.0% and 9.9%; ICC 0.821, 0.216 and 0.760 (full text)
Running: K5 112.2 ± 21.3 kcal; AW 137.8 ± 23.1, GF 131.1 ± 28.8, HW 111.8 ± 21.1 kcal; MAPE 24.4%, 21.8% and 11.9%; ICC 0.741, 0.594 and 0.698 (full text)
Table 2 differences: walking 20.5 ± 11.3, 31.0 ± 38.3 and 2.0 ± 12.4 kcal; running 25.7 ± 16.0, 19 ± 22.8 and −0.3 ± 16.7 kcal; Table 2 prints the Garmin running mean as 137.8 ± 23.1 (full text)
Limits of agreement 44.1, 150.1 and 48.6 kcal for walking and 62.8, 89.4 and 65.6 kcal for running (full text)
EE estimates "obtained directly from screenshots of the respective applications"; "we only obtained 19 EE data from HW" (full text)
Statistical power for HW 0.105 (walking) and 0.051 (running) (full text)
"This research was supported by Huawei Terminal Co. LTD." and "the research was conducted in the absence of any commercial or financial relationships" (full text)