back
2 comments
Apple's VO2Max measures are not based upon that deep neural network development, and empirical seems to be conflating a few things. And FWIW, just finding the actual paper is almost impossible as that same site has SEO-bombed Google so thoroughly you end up in the circular-reference empirical world where all of their pages reference each other as authorities.

Apple and Columbia did recently collaborate on a heart rate response model -- one which can be downloaded and trialed -- but that was not related to the development of their VO2Max calculations.

Apple is very shrouded about how they calculate VO2Max, but it likely is a pretty simple calculation (e.g. how much is your heart responding based upon the level of activity assumed based upon your motion, method of exercise and movements). The most detail they provide is in https://www.apple.com/healthcare/docs/site/Using_Apple_Watch..., which mostly is a validation that it's providing decent enough accuracy.

What’s your source on Apple not using the neural network for VO2Max estimation? They’ve been using on-device neural networks for various biomarkers for several years now (even for seemingly simple metrics like heart rate).

FWIW, the article above links directly to both the paper and a GitHub repo with PyTorch code.

>FWIW, the article above links directly to both the paper and a GitHub repo with PyTorch code.

Neat, though the paper and the Github repo have nothing to do with Apple's VO2Max estimations. It's related to health, and touches on VO2Max and health sensors, but the only source claiming any association at all is that Empirical site. And given that this research came out literally years after Apple added VO2Max estimates to their health metrics, it seems pretty conclusive that it is not the source of Apple's calculations. Neat research related to predicting heart rate response to activity (which might come into play for filling in measurement gaps which happen during activity when a device isn't tight enough, etc).

>What’s your source on Apple not using the neural network for VO2Max estimation?

You're asking me to prove a negative. Apple never claims that they do any complex math or deep neural networks to derive VO2Max, and from my own observations of its estimates of mine, it seems remarkably trivial.

Trivial can still be accurate. But it hardly seems complex. Like, guess people's A1c based upon age, body fat percentage, demographic and you'll likely be high-90s accurate with trivial algebra.

>even for seemingly simple metrics like heart rate

Deriving heart rate from a green light imperfectly reflecting off skin, watching for tiny variations in colour change, is actually super complex! Doing it accurately is actually pretty difficult, which is why wearable accuracy is all over the place, though Apple is one of the leaders and has been for years. Guessing a number based upon HR and activity level isn't quite as complex.

Interesting, there is no mention of neural networks in that 2021 pdf from Apple (https://www.apple.com/healthcare/docs/site/Using_Apple_Watch...). Also, it seems like the only reference of VO2 Max computation with neural net is on empirical.health blog. https://imgur.com/a/2hX7XNc

I wonder now if all of the derived metrics on Garmin (Training readiness, Training load, Training status)are purely statistical algorithms

Apple was reporting VO2max for a very long time (much before 2023). I wonder what the accuracy was back then? Maybe they should the option for users to re-compute those past numbers based on the latest and greatest algorithm.