Can Your Wearable’s VO2max Estimate Track Real Fitness?

Your wearable's VO2max estimate probably contains a real signal. Treating its displayed number as a laboratory measurement is the mistake. The best free-living research pipeline estimated VO2max with an error of 2.902 mL/kg/min, but it used roughly seven days of continuous multisensor data and a model built for research—not the proprietary software inside a consumer watch.

That difference matters because cardiorespiratory fitness (CRF), commonly measured as VO2max, tracks health outcomes that deserve attention. In patients with established cardiovascular disease, a meta-analysis of 21 prospective cohorts including 159,352 people found that the highest versus lowest CRF category was associated with lower all-cause mortality (hazard ratio 0.42, 95% CI 0.28–0.61) and cardiovascular mortality (0.27, 95% CI 0.16–0.48). Those are associations, not proof that a watch score changes anyone's prognosis, but they explain why VO2max attracts so much interest.

The best result came from a very different kind of wearable study

The strongest validation evidence comes from Spathis et al.'s 2022 study in the Fenland cohort. Researchers collected about seven days of free-living wrist accelerometry, heart rate, heart-rate variability, moderate-to-vigorous physical activity, resting heart rate, and anthropometric data from 11,059 UK adults. Participants also completed supervised cycling ergometry in a clinic to measure VO2max.

The team trained a dense neural-network model in 8,384 participants and evaluated it in a held-out test set of 2,675. The full multisensor model produced an R² of 0.671, a correlation of 0.821, and a root mean square error (RMSE) of 2.902 mL/kg/min (95% CI 2.806–3.002). RMSE describes the typical size of prediction errors: smaller is better.

Exercise physiology often treats about 3.5 mL/kg/min—roughly one metabolic equivalent, or MET—as a practically meaningful change in fitness. That is a convention, not a diagnostic cutoff. Still, an error around 2.9 mL/kg/min puts this research system near a range where repeated measurements could become useful.

The result also identifies what made the model work. Anthropometrics alone had an RMSE of 4.051 mL/kg/min, and resting heart rate alone had an RMSE of 4.007. Combining anthropometrics and resting heart rate improved the error to 3.159 mL/kg/min. The full sensor set did better because it captured more of how a person actually moves and recovers across daily life.

Fancy modeling helped less than better inputs. A linear model using the same features had an RMSE of 2.956 mL/kg/min, only 0.054 higher than the neural network. Traditional non-exercise equations performed far worse, with R² of −3.2 and a correlation of 0.389. In this setting, sensor fusion mattered more than algorithmic fireworks.

Why that does not validate the number on your wrist

A 2026 systematic review and meta-analysis by Dosis et al. reviewed 18 studies of unsupervised, free-living CRF estimation. Its conclusion cuts against consumer-watch marketing: commercially available wearables produce a large degree of error for individuals. The review attributes much of the gap to calibration. Commercial algorithms are primarily built from short periods of structured exercise data, whereas the Fenland model used a week of continuous free-living monitoring and multiple signals.

Those approaches answer related but different questions. A watch may infer fitness from a GPS-tracked run, pace, and heart-rate response; another estimate may lean heavily on resting heart rate when exercise data are sparse. Neither setup automatically reproduces the performance of a research model that combines movement, heart-rate variability, activity intensity, resting heart rate, and body measurements over days.

The review also found that many studies used internal validation only. Models often look better when tested on people drawn from the same data source used to build them. Spathis et al.'s dedicated held-out test set is therefore more persuasive than a model assessed only within its development sample, though it still needs replication in more varied populations.

Consumer studies show signal, not a settled accuracy claim

Consumer hardware is not physiologically blind. A 2024 community cohort study by Zhang et al. found associations between smartwatch-derived heart-rate and physical-activity measures and cardiorespiratory-fitness measures. That supports the basic premise that watches capture information related to fitness. It does not establish that a given person's displayed VO2max is accurate enough to substitute for a lab test.

A 2025 Apple Watch validation study by Lambe et al. compared Apple Watch VO2max estimates with a laboratory standard. The available abstract information does not provide extractable RMSE, standard error of estimate, or Bland-Altman limits of agreement. Without those individual-level accuracy measures, the study cannot settle how close an Apple Watch value is likely to be for any one wearer.

This distinction is easy to miss. A device can correlate with fitness across a group and still make errors large enough to mislead an individual. Population-level association is a starting point, not a permission slip to make clinical decisions from a dashboard.

Who should be especially cautious?

The most accurate free-living result came from a relatively healthy UK adult cohort. The evidence does not establish comparable accuracy in people with VO2max above 55 mL/kg/min, severe obesity, advanced age, or diverse clinical conditions. Prediction models commonly struggle at the edges of the population used to build them.

Mechanical efficiency adds another problem. Non-exercise estimation methods often assume that people convert movement into work with similar efficiency. The 2026 review reports that this assumption can fail in older, frail, or musculoskeletal-impaired people and may systematically underestimate VO2max. A 2024 seismocardiography study illustrates why researchers continue to test alternative sensors, including chest vibration signals collected at rest. It does not yet provide a consumer-wearable answer.

Wear time is another unresolved variable. The Fenland result depended on roughly seven days of continuous collection. No evidence here shows that one, two, or three days converge on the same estimate. A weekly trend built from inconsistent wear behavior can look more scientific than it is.

Fitness matters; the estimate remains a measurement problem

CRF's relevance extends beyond endurance performance. A 2025 systematic review and meta-analysis found that fitness was associated with mortality risk across BMI categories. A separate meta-analysis in cardiovascular-disease patients found lower mortality among those with higher CPET-measured fitness. These findings concern measured CRF and long-term outcomes; they do not validate consumer-device estimates.

Other observational evidence points in the same direction. A 2026 UK Biobank cohort study of 54,418 adults aged 39 to 70 without diabetes at baseline associated higher estimated VO2max with lower incident type 2 diabetes risk and examined biological aging as a possible mediator. A 2024 study of cognitively normal adults found positive correlations between VO2max and myelin-water fraction in several white-matter regions. Neither association proves that increasing a wearable score will prevent diabetes or preserve brain myelin, but both reinforce why accurately tracking fitness would be useful.

How to use your wearable without fooling yourself

• Follow the trend over months. If you wear the same device consistently and your estimate rises or falls alongside training changes, that direction may be more useful than the absolute number.
• Do not overread one workout. Resting heart rate alone had about 4.0 mL/kg/min RMSE in the Fenland analysis, and short structured sessions do not match a week of multisensor observation.
• Keep context with the metric. Changes in weight, training type, illness, heat, medication, and imperfect sensor contact can all affect the inputs that algorithms use.
• Use laboratory testing when the stakes are clinical. Cardiopulmonary exercise testing remains the appropriate method for cardiac risk assessment, diagnostic questions, or decisions that depend on precise VO2max.

A consumer wearable can be a useful longitudinal fitness instrument. It is not yet a validated replacement for a lab measurement at the individual level.

The defensible conclusion is moderately confident and deliberately narrow: your watch's VO2max estimate likely reflects real physiology, especially when it moves consistently over time. The 2.902 mL/kg/min research benchmark shows what passive, multisensor monitoring can achieve under strong conditions. It does not tell us that your specific Apple Watch, Garmin, or Fitbit reaches that benchmark. Until independent consumer-device studies consistently report small individual errors in varied populations, use the number as a trend line—not a diagnosis.