Wearable VO2 Max Estimates: How to Read Validation Evidence

To judge wearable VO2 max accuracy, start with one distinction: a laboratory VO2 max is a measurement, and the number on your wrist is a model's estimate. Whether that estimate is "accurate" depends on which device, which activity, which reference test and which people were studied. This guide shows how to read that evidence, what three published studies do and don't cover, and how to keep a simple log so you can judge your own readings.

Lab measurement vs. wearable estimate

VO2 max is traditionally measured in a lab. You run on a treadmill or ride a stationary bike to exhaustion while a mask measures the air you inhale and exhale. A wrist or ring device skips the mask. It estimates the value from inputs such as pace, heart rate, resting heart rate, age, sex and weight, using a model the manufacturer designed.

A manufacturer's description of that model tells you what the number is built from. It doesn't tell you how close the number comes to a lab result. Only independent comparison against a reference test can do that.

Prerequisites: does your device and activity qualify?

Algorithms differ, so check your own device's documentation for these conditions before you interpret a value:

• Activity type: some devices produce an estimate only from certain activities, often outdoor runs or walks with a GPS signal. Others may use different inputs.
• Minimum effort or duration: a device may need a minimum length of activity, or several qualifying sessions, before showing a first value.
• Profile data: the model may rely on the age, sex and weight you entered, so wrong profile data affects the estimate.
• Your own history: if you mostly lift, cycle or run indoors, you may have few or no qualifying sessions. A missing or stale value then reflects the input rules, not necessarily a faulty device.
• Not a diagnosis: the score is a fitness estimate. Don't use it to judge a medical condition; ask a clinician about health concerns.

Manufacturers revise these rules, so check again before relying on them. Don't read today's rules back into older studies, which used their own procedures.

Reading the validation evidence

Three published papers are useful examples. Each is narrow, and none speaks for every device.

A wrist tracker against a treadmill test (2019, healthy young adults)

This study tested 30 healthy people (17 men, 13 women), aged 18 to 35 (mean about 21.7). Each wore the tested wrist tracker for seven days and followed instructions to obtain its cardio fitness estimate. The reference was an incremental treadmill VO2max test with a verification phase. A non-exercise prediction equation was also compared.

• Measured VO2max averaged 49.91 and the device estimate averaged 52.53, a statistically significant difference (P = 0.03).
• The authors reported an intraclass correlation coefficient (ICC) of 0.87, a reliability/agreement statistic, and described the Bland-Altman analysis as showing consistency.
• Mean absolute percent error was 10.2% for the device estimate, versus 7.8% for the non-exercise equation.
• The device's estimated heart rate was lower than a heart rate extrapolated from pace, and the gap between estimate and measured VO2max was inversely related to that heart rate difference (r = -0.45).

The same type of tracker in outdoor runners (2020)

A second study recruited adults aged 18 to 45 who were free of chronic disease or injuries that would prevent a graded exercise test, and who could run continuously for at least 15 minutes. They also needed a smartphone with GPS for the tracker. The reference was a treadmill graded exercise test to volitional fatigue with breath-by-breath gas analysis. The authors averaged the data into 15-second and 60-second epochs. The device estimate showed a positive bias of 1.59 mL·kg⁻¹·min⁻¹ against the 15-second laboratory value and 0.30 against the 60-second value, and mean absolute percentage error was under 10% for each comparison.

Notice how the reference choice changes the bias. Even the "gold standard" involves analysis decisions, such as how the lab averages its data.

Five wrist-worn trackers (2019, 24 healthy participants)

This study tested 24 healthy participants and five commercial wrist devices against indirect calorimetry. Not every device produced both a VO2max and an energy expenditure output. Error exceeded 10% for all evaluations except one device's VO2max estimate (7.3%), which still significantly underestimated VO2max. The authors concluded these wrist-worn trackers were most likely not accurate enough for sports or health care uses.

Taken together, the studies show that results vary by device and protocol, and that "reasonably close on average" can coexist with sizeable individual error.

Correlation is not agreement

A high correlation says two measures rise and fall together. It doesn't say they give the same number. A watch that reads 8 units too high for everyone could correlate nearly perfectly with the lab and still be wrong for each person. To judge agreement, look for:

1. Bias: the average difference (Bland-Altman).
2. Spread: how widely individual differences scatter around that bias.
3. Error size: mean absolute percentage error, which stops over- and under-estimates from cancelling out.

In the first study, the reported ICC was 0.87, but the mean absolute percent error was 10.2% and the mean difference was statistically significant. In the second, bias was small on average. Both can be true, because average bias and individual error are different things.

Applicable activities and validation populations

These findings are limited to the tested devices, their estimation methods, the activities used and the healthy participants enrolled. They do not establish how other devices, methods, populations or activity contexts would perform. Evidence for one model also can't validate a different model, even from the same maker.

A step-by-step workflow

1. Identify your device and app. Note the model and generation.
2. Find the current activity rule in its documentation and confirm that you do qualifying sessions.
3. Find a study naming your exact model. If none exists, label the evidence "adjacent."
4. Check participants and reference. Compare age, health status and fitness level with your own, and note whether the reference was a lab gas-analysis test.
5. Read agreement, not just correlation. Look for bias, spread and percentage error.
6. Log your own readings (below) and look at the trend, not any single number.
7. Don't rank devices from separate studies, since samples, protocols and references differ.

Illustrative audit log

The numbers below are made up to show the format. They are not test results.

The value changed only after outdoor runs, and the week 3 dip coincides with a different route. Without logging terrain, you might read it as a fitness change. The first study linked the estimate gap to a difference in heart-rate estimation, so add a note on heart-rate data quality for each run. That is a contextual note; the study did not test how the device was worn.

Acceptance checklist

• [ ] I know whether my number is a lab measurement or a device estimate.
• [ ] My device and activity match the manufacturer's current requirements.
• [ ] I found a study naming my model, or I marked the evidence as adjacent.
• [ ] The study's sample resembles me in age and health.
• [ ] I checked bias, error size and spread, not only correlation.
• [ ] I logged activity, route and conditions for each reading.
• [ ] I'm using the number as a personal trend, not a diagnosis or a ranking against other brands.

If you want to understand how training can influence aerobic capacity once you trust your baseline, Speediance covers it in its VO2 max workouts guide.