How Accurate Is the Running Powertest? Validation Against Lab Spirometry in 25 Runners
You finish a hard run, your watch shows a VO2max, and somewhere in the back of your mind the question appears: how would a lab see me? Nobody wants to run with a mask just to find out — so we did it for you.
Together with MSH Medical School Hamburg, runners went through a treadmill ramp to exhaustion with lab spiroergometry (COSMED K5), the gold standard. Then we gave our Powertest nothing but their running speed and compared.
The short version: Across 25 runners, our VO2max reads on average 2.0 ml/min/kg below the lab, and a single test typically sits 4.0 ml/min/kg away (7.2%), with a correlation of 0.85. That's almost exactly where our cycling validation landed — and on foot we don't even have a power meter.
Why running is the harder case
On a bike, a power meter hands us watts directly. On foot, there is no crank to measure, and every runner moves differently: stride, ground contact, shoes, technique. So the question is blunt: if all the model gets is your speed, how close does it come to the lab?
Our Powertest turns your speed into running power and fits the same metabolic model we use on the bike (Mader). From that it derives VO2max, VLamax, your threshold and your training zones. One conversion is the core assumption: how much oxygen a runner needs per watt.
How we validated it
Methodology: How We Built This Comparison
Our platform holds over 15,000 Powertests from more than 1,000 athletes. But validating VO2max needs something rare: the same runner, measured by both our test and a metabolic cart.
Our data was collected as part of a joint research project with MSH Medical School Hamburg: the 2025 running cohort ran a ramp to exhaustion on a laboratory treadmill at 1% incline, speed rising by 1 km/h per minute, with gas exchange measured throughout by a COSMED K5 in mixing-chamber mode (Dynamic Micro-Mixing Chamber), recorded as 10-second values. 31 runners were tested, one test per athlete. One measurement day with physiologically impossible references was removed as a whole, six tests in total (details below), which leaves 25.
For every test, only the treadmill speed went into our production pipeline: no oxygen data, no heart rate, no extra sensor. The oxygen data was deliberately withheld, so the model could never peek at the answer, and VO2max was computed exactly the way it is for every runner on the platform. The lab reference is the highest 30-second average of measured VO2 (formally a VO2peak criterion).
The 25 runners range from 38 to 73 ml/min/kg, women and men (9 and 16), recreational to well-trained. The joint research program with MSH is ongoing; a scientific publication is planned as the cohorts grow. Data as of October 2026, powertest engine v1.11.16.
Our validation standard
Running follows the same rules as our cycling validation, and every validation we publish follows them:
- Independent reference. A university laboratory measures, not us: spiroergometry, with the highest 30-second VO2 average as the reference.
- Blind input. Our model never sees the oxygen data, only what you record yourself: time, power and heart rate on the bike, speed on foot.
- The product, not a lab version. We run the exact production pipeline every athlete gets and name its engine version.
- One test per athlete. Nobody counts twice.
- Exclusions by rule, with reasons documented. A test is removed only by a rule that never looks at our own result.
- Agreement, not just correlation. Bias with its 95% confidence interval, mean absolute error, Pearson r and the 95% limits of agreement (Bland–Altman).
- No tuning on the validation cohort. No model parameter is fitted to the athletes we validate on. The oxygen-per-watt constant dates from 2021, before the first lab cohort; the ramp factor is calibrated on more than 1,000 tests from our platform. The lab cohorts serve only as a check.
- Limits named. What is not validated yet stands in the article.
The result
| Metric | Running, treadmill ramp (n = 25) |
|---|---|
| Bias (Powertest − lab VO2peak) | −2.0 ml/min/kg [95% CI −3.8 to −0.1] |
| Mean absolute error (MAE) | 4.0 ml/min/kg (7.2%) |
| Correlation (Pearson r) | 0.85 |
| 95% limits of agreement | −10.8 to +6.8 ml/min/kg |
In plain language: a slightly conservative reading of about two points, and a typical single-test difference of four. For comparison, our cycling validation with 43 athletes shows a bias of −1.8 and a mean absolute difference of 3.3 ml/min/kg — running sits barely behind, without a single watt measured at the foot. The full cycling result is in how accurate the Powertest is on the bike.
Is that good? Context helps: the lab reference itself is not a perfect ruler. Repeat spirometry on the same runner varies from visit to visit as well. The lab is a reference, not an oracle.
One measurement day, removed as a whole
On one of the lab days, four reference measurements were physiologically impossible: runners held 17 km/h while the cart recorded an oxygen uptake of 31–39 ml/min/kg — only half to two thirds of what that speed demands. That pattern is typical of a leaking mask or a mis-calibrated volume sensor, and such faults only ever read low. Rather than hand-picking cases, we removed that entire day — all six tests, including the plausible ones. The rule behind it compares the lab value with the oxygen demand of the speed the runner actually held; it never looks at our Powertest. MSH confirmed the exclusion. If you keep every test that produced a result in (30 runners; one very short ramp ends before the model can fit it), the mean absolute difference is 5.8 ml/min/kg (13.3%); it's those few impossible references that drive it, not the model. Because such faults only ever read low, removing them can only help our numbers. That's exactly why we show you the unfiltered figure too.
The one error term: running economy
The single conversion assumption, oxygen per watt, varies a bit more between runners than between cyclists: measured against the same value of 11.4 ml/min/W that we use on the bike, 15 of the 25 runners (60%) sit within ±7%; on the bike it is 68%. Technique, shoes and stride all show up there. That's why the running result is a little wider — and why it's still remarkably close.
Your speed already carries the signal
We also tested whether running dynamics explain oxygen uptake beyond speed: ground contact time, vertical oscillation, cadence, stride length, leg stiffness and more. In this cohort (27 runners with plausible lab references) and across a dozen statistical methods, none beat a single fixed economy constant (best model 7.0% vs 7.3% error, p = 0.25), and vertical oscillation showed no measurable link at this cohort size. A 2024 meta-analysis (Van Hooren et al.) does find a modest link between such factors and economy; at the effect size reported there, proving it reliably would take around 60 runners. For you that means no extra hardware: a watch with GPS is enough.
Where we're honest about the limits
Treadmill yes, field test next. This validates the treadmill ramp against the lab. Most runners on our platform take their Powertest outdoors, on a track or road; that field version has not been put next to a metabolic cart yet. That is exactly the next step of our joint program with MSH: the same runners, a lab ramp and a track ramp, the same gas-exchange device. Until then, read your outdoor VO2max with the same corridor in mind, not as a lab-equivalent measurement.
Flat model, 1% treadmill. Our product models the treadmill as flat, while the lab ran at 1% incline. Accounting for the incline would move the bias to +0.9 ml/min/kg and the mean absolute difference to 3.6; the scatter around the bias stays the same. We still publish the flat number, because it is what the product computes today.
Why we don't simply subtract the bias. The obvious question: if we read 2.0 below the lab, why not add 2.0 and call it zero? Because tuning the model to the same 25 runners we validate on would turn an independent validation into curve fitting. A correction deserves to exist only once a new cohort confirms it. The same holds for the incline: the validation stays a test, not a training set.
The individual corridor is real. A bias near zero doesn't mean every test is near zero. Under the usual distributional assumptions, about one test in twenty is expected to fall outside the −11 to +7 ml/min/kg corridor. If your absolute VO2max matters medically or professionally, a spirometry remains the reference — and slots straight into your profile.
One lab, one cohort. 25 runners from one laboratory. The joint program with MSH continues, and a scientific publication is planned as the cohorts grow.
What this means for you
If you've been treating your running VO2max as a rough guess because "it's just a watch" — on the treadmill ramp, the data says it sits within a few points of the lab. And your training zones run on your own speed, the number you actually train and race with.
Ready to see your engine? Start your free trial and run your first Powertest this week.
FAQ
How accurate is the running Powertest compared to a lab test? On the treadmill ramp, across 25 runners: mean bias −2.0 ml/min/kg, typical single-test error 4.0 ml/min/kg (about 7%), r = 0.85, from running speed alone.
Do I need a footpod or running power meter? No. In our lab cohort, running-dynamics data from a footpod did not make the VO2max estimate measurably more accurate than speed alone. A GPS watch is enough.
Does this also hold for my outdoor test? Not yet validated. The outdoor field test is the next step of our joint study program with MSH Medical School Hamburg.
How does running compare to cycling? Almost the same: cycling 3.3 ml/min/kg mean absolute error in 43 athletes, running 4.0 in 25 runners. Details in our cycling validation.
Sources: Mader, A. & Heck, H. (1986). A theory of the metabolic origin of "anaerobic threshold". International Journal of Sports Medicine, 7(Suppl 1), S45–S65. · Internal validation cohort: laboratory study conducted as part of a joint research project with MSH Medical School Hamburg (running 2025, COSMED K5 in mixing-chamber mode, 25 treadmill ramps after removing one measurement day with physiologically impossible references; one test per athlete); analysis code and statistics documented internally, powertest engine v1.11.16.