Can I trust my watch's lactate-threshold heart rate?
Only as an estimate: in 100 runners watch LTHR missed the lab value by 9-11 bpm on average and tracked it weakly (r = 0.13-0.67). Label it estimated.
Lu and colleagues compared the lactate-threshold heart rate (LT HR) and pace estimated by three smartwatch tests with laboratory values.
One hundred healthy recreational runners did a graded treadmill test (5-min stages, 1.2 km/h steps, fingertip lactate) analysed by the modified Dmax method, then, at least 48 h later and on a 400-m track, each watch's guided threshold test: all 100 with a Huawei GT Runner, 23 with a Garmin Forerunner 265 and 17 with a COROS Pace 3, by the runner's choice.
A single watch test produced a value in 78% (Huawei), 65.22% (Garmin) and 47.06% (COROS) of runners.
Watch LT HR did not differ significantly from the laboratory value on average for any device, with mean absolute errors of 8.93 to 11.44 bpm (5.95-7.15%), but tracked it poorly between runners (r = 0.36 Huawei, 0.67 Garmin, 0.13 COROS), compressing values toward the middle; by the two one-sided tests none was equivalent within d = ±0.5.
LT pace was overestimated by every watch (mean absolute error 1.22-2.17 km/h, 12.70-25.78%). The authors conclude that watches can estimate LT HR and pace but that accuracy needs to improve, and advise reading a watch LT HR with perceived exertion.
Our reading: a watch LT HR misses by about as much as an age formula misses HRmax (REF-3494), so it belongs with estimated anchors; the Garmin and COROS results rest on 15 and 8 successful tests, and the numbers for each device are reported inconsistently between the abstract, the tables and the text.
Protocol
Participants: n = 100 healthy recreational runners, 18-65 years, without heart disease, hypertension, musculoskeletal disorders or recent lower-limb injury; Huawei group 61 men and 39 women, 28.0 ± 9.6 y; Garmin 13 and 10, 29.13 ± 9.93 y; COROS 12 and 5, 25.88 ± 7.11 y. Ethics: Beijing Sport University (2024136H). Of all participants 71 tested one watch, 17 two and 11 three (watches chosen by preference and time).
Laboratory test: resting heart rate and lactate, 10-min warm-up, treadmill at 1% incline from 5 km/h (men) or 4 km/h (women), 5-min stages with a 30-s pause for fingertip lactate (Biosen C-line) and RPE, +1.2 km/h per stage to exhaustion; heart rate by Polar H10.LT by modified Dmax: cubic-spline interpolation, third-order polynomial, maximal distance from the line between the first point where lactate rose by more than 0.4 mmol/L and the last point; LT HR and pace read at that point.
Watch tests: outdoors on a 400-m track in non-extreme weather, morning or early afternoon, at least 48 h after the lab test and between watch tests; watch on the non-dominant wrist, a Polar H10 worn by all; 10-min warm-up, then 20-60 min following each watch's prompts.Huawei: HRmax entered from the lab peak (otherwise 220 − age), 8-min warm-up, six stages raising heart rate every 3-5 min, at least four completed. Garmin (Firstbeat): needs a prior outdoor VO2max estimate and a chest strap to start, 10-min warm-up, six stages of 3-4 min by heart-rate zone.COROS: needs a 10-km time, 5-min warm-up, 25 min at marathon pace, 3 min at 10-km pace, 3 min at 5-km pace, up to two extra sets if no threshold was found.
Heart-rate source: the Methods say the analysed heart rate came from the watches' optical sensors; the Limitations say Garmin used the chest strap. Which source fed each watch's algorithm is therefore unclear for Garmin.
Analysis: success rates; Welch's t tests and Mann-Whitney U; MAE and MAPE; linear regression (R², Pearson r); Bland-Altman plots; two one-sided tests with equivalence bounds of d = ±0.5 (pooled SD).
Funding and interests: National Natural Science Foundation of China (72071018) and Fundamental Research Funds for the Central Universities (2023GCZX003, 2025025); the authors declare no commercial or financial conflict, while the Acknowledgments thank three Huawei Device Co. staff for "constructive feedback" and for "supplying the equipment".
Key Results
Table 4 (successful tests only; LT HR in bpm, LT pace in km/h; lab mean ± SD vs watch mean ± SD):
| Device (successful / tested) | Lab LT HR | Watch LT HR | MAE (MAPE) | p | r (R²) | Lab LT pace | Watch LT pace | MAE (MAPE) | p | r (R²) |
|---|---|---|---|---|---|---|---|---|---|---|
| Huawei (78 / 100) | 169.20 ± 15.12 | 167.60 ± 9.93 | 10.13 (6.63%) | 0.24 | 0.36 (0.13) | 9.76 ± 2.28 | 10.71 ± 2.19 | 1.22 (12.70%) | 0.01 | 0.88 (0.78) |
| Garmin (15 / 23) | 162.14 ± 14.81 | 168.67 ± 8.25 | 8.93 (5.95%) | 0.15 | 0.67 (0.45) | 8.93 ± 1.63 | 10.99 ± 1.60 | 2.17 (25.78%) | 0.00 | 0.73 (0.53) |
| COROS (8 / 17) | 167.46 ± 12.50 | 175.38 ± 5.18 | 11.44 (7.15%) | 0.13 | 0.13 (0.02) | 9.75 ± 2.12 | 11.69 ± 2.01 | 1.93 (22.63%) | 0.08 | 0.79 (0.62) |
Success of a single test: 78% Huawei, 65.22% Garmin, 47.06% COROS; failures were more common at atypical (lower) thresholds.
Compression: watch LT HR spread was about 40-65% of the lab spread (SD 5.18-9.93 vs 12.50-15.12 bpm; our arithmetic); the watches overestimated low thresholds and underestimated high ones, so runners far from the average are misplaced most (the authors).
Equivalence: none of the three met the equivalence criterion for LT HR or LT pace.
Bland-Altman mean deviations (text): LT HR 2.86 (Huawei), 6.28 (Garmin), 5.44 (COROS) bpm; LT pace 0.98, 2.06, 1.77 km/h; limits of agreement are shown only in figures, described as "notably wide".
Numbers that do not add up: (1) the abstract gives MAE and MAPE of 10.66 bpm and 6.32% for Huawei, 11.44 and 7.15% for Garmin and 8.93 and 5.95% for COROS, while Table 4 gives 10.13 and 6.63% for Huawei and the Garmin and COROS values swapped.(2) Mean differences from Table 4 (watch minus lab: Huawei −1.60, Garmin +6.53, COROS +7.92 bpm; our arithmetic) do not match the TOST mean differences (−2.44, −6.52, −2.65 bpm) or the Bland-Altman means (2.86, 6.28, 5.44) for Huawei and COROS.(3) The text says the Garmin watch "exhibited underestimation in LT HR prediction" while Table 4 and the Bland-Altman text show overestimation. (4) The text says COROS failed tests differed from successful ones in LT HR and pace "(p < 0.01)", while Table 3 gives p = 0.74 and 0.36. (5) Huawei's successful-test lab LT HR is 165.12 in Table 3 and 169.20 in Table 4.(6) 71 + 17 + 11 = 99 participants, not 100. (7) The Discussion gives the LT pace MAPE range as 12.70-25.63% (Table 4: 25.78%). (8) The abstract calls the LT pace overestimation "significant ... for all devices" while giving p = 0.08 for COROS; COROS LT pace differs by +1.94 km/h in Table 4, −1.53 in the TOST text and 1.77 in the Bland-Altman text.The published correction changes only the corresponding author.
Mechanisms
Measured: the watch estimates sat close to the group mean of the lab values but explained little of the between-runner variation in LT HR (R² 0.02-0.45), and put LT pace above the lab value in every device (significant for Huawei and Garmin; COROS p = 0.08, 8 tests).
Authors' interpretation: the algorithms appear to predict near the population average, perhaps by design (a conservative LT HR keeps runners from training too hard); inputs that are themselves estimates (an age-formula HRmax, a watch VO2max, a best-ever 10-km time) and pace-guided stages that do not reach a heart-rate steady state limit accuracy; heart-rate-guided protocols (Huawei, Garmin) were more stable than pace-guided ones (COROS, the lowest success rate).They cite evidence that treadmill running elicits lower blood lactate at submaximal speeds, which may explain part of the pace overestimation.
Inference: a mean absolute error of 9-11 bpm is about the size of the age formula's individual error for HRmax (RMSE 9.1-10.5 bpm, REF-3494); since RMSE is never smaller than MAE, the watch's error is at least that large (our arithmetic; different quantities, threshold vs maximum).
Inference: the Huawei arm was given each runner's laboratory peak heart rate; in ordinary use, with 220 − age, its error would probably be larger.
Practical Application
Can I trust my watch's lactate-threshold heart rate? Only as an estimate: in 100 runners watch LTHR missed the lab value by 9-11 bpm on average and tracked it weakly (r = 0.13-0.67). Label it estimated.
Rank a watch LTHR with the estimates: its error is about that of an age formula, so it should not gate decisions; confirm it with a hard steady effort, an HRV ramp (REF-1231) or a lab test before building zones on it (B, recreational runners).
Runners far from average are misplaced most: the watches pulled high thresholds down and low ones up, so a fit runner's watch LTHR may sit below the true value and a less fit runner's above it (the authors' reading of the plots).
Do not set threshold pace from the watch test: every device overestimated LT pace by about 1-2 km/h against a treadmill test (part of the gap may be treadmill vs track).
A failed watch test is common: 22-53% of single tests returned nothing; repeating it is normal, not a sign of a problem with the runner.
Limitations
Small device subsamples: Garmin and COROS results rest on 15 and 8 successful tests; runners chose which watches to test, so the groups are self-selected.
Different conditions: treadmill reference (1% incline) vs outdoor track tests on another day; part of every difference may be day-to-day variation or treadmill vs track, not the watch.
One test, one firmware: no repeat test, so the watches' reliability is unknown; algorithms change with firmware. Watch and lab values from the same runners were compared with Welch's unpaired t tests, so the "no significant difference" results are not paired comparisons.
Favourable inputs: the Huawei test used each runner's laboratory peak heart rate.
Heart-rate source unclear for Garmin (optical in the Methods, chest strap in the Limitations).
Reporting: abstract and table MAE values disagree (Garmin and COROS swapped, Huawei different), mean differences disagree across Table 4, TOST and Bland-Altman, text and tables disagree on direction (Garmin) and significance (COROS failures), and the participant counts sum to 99.
Industry link: Huawei supplied equipment and gave feedback (Acknowledgments) while the conflict statement declares none; the Huawei arm was the largest and had the highest success rate.
Bearing on [SYNTH-38]: recreational runners and three 2024-2025 watch models; transfer to cyclists' auto-detected LTHR is untested.
How this entry was checked and changed
Why this level
test validation, n = 100 for one device but only 23 and 17 for the others (15 and 8 successful tests), read from the full text (open access, CC BY 4.0); the rubric's 100-or-more cross-sectional band does not apply to the device comparisons, and internal inconsistencies in the reported numbers lower confidence further
Full text (checked 2026-10-27)
One hundred healthy recreational runners: Huawei (n = 100), Garmin (n = 23), and Coros (n = 17) (full text)
The success rates for a single test were 78% for Huawei, 65.22% for Garmin, and 47.06% for Coros (full text)
Table 4, LT HR: Huawei DmaxMod 169.20 ± 15.12 vs watch 167.60 ± 9.93, MAE 10.13, MAPE 6.63%, p 0.24; Garmin 162.14 ± 14.81 vs 168.67 ± 8.25, MAE 8.93, 5.95%; Coros 167.46 ± 12.50 vs 175.38 ± 5.18, MAE 11.44, 7.15% (full text)
Abstract: Huawei MAE = 10.66 bpm, MAPE = 6.32%; Garmin MAE = 11.44 bpm, MAPE = 7.15%; Coros MAE = 8.93 bpm, MAPE = 5.95% (full text)
LT HR regression: Huawei R 2 = 0.13 and r = 0.36; Garmin R 2 = 0.45 and r = 0.67; Coros R 2 = 0.02 and r = 0.13 (full text)
Table 4, LT pace: Huawei 9.76 ± 2.28 vs 10.71 ± 2.19, MAE 1.22, 12.70%; Garmin 8.93 ± 1.63 vs 10.99 ± 1.60, MAE 2.17, 25.78%; Coros 9.75 ± 2.12 vs 11.69 ± 2.01, MAE 1.93, 22.63% (full text)
TOST mean differences −2.44 bpm for Huawei, −2.65 bpm for Coros, and −6.52 bpm for Garmin; Bland-Altman mean deviation 2.86 bpm (Huawei), 6.28 bpm (Garmin), 5.44 bpm (Coros) (full text)
71 completed outdoor testing with one smartwatch, 17 with two smartwatches, and 11 with all three (full text)
For the Coros watch the text reports differences in both LT HR and LT Pace (p < 0.01); Table 3 gives 167.45 ± 12.50 vs 159.10 ± 24.38, p 0.74 (full text)
Equivalence bounds were defined using Cohen's d = ±0.5 (full text)
Acknowledgments thank Hang Liu, Litian Zhou, and Yufeng Fang of Huawei Device Co., Ltd. for supplying the equipment; funding from the National Natural Science Foundation of China Grant (72071018) (full text)