Wilson Cusack
@wilsoncusack
Every morning Eight sleep: congrats on getting 2hrs of deep sleep! Apple Watch: yeah… it was 30m
7 replies
12 recasts
71 reactions
shazow
@shazow.eth
There's a good chance neither of these are correct. They generally correlate motion/hr with average person behaviour, but most people aren't average. Need a head strap that can measure brain activity to get good sleep stage metrics. I used to have a MyZeo back in the day and watch-based sleep metrics were completely uncorrelated. (Sadly not sure if there's a good available option today.)
1 reply
0 recast
4 reactions
Wilson Cusack
@wilsoncusack
It's pretty darn good! https://www.apple.com/healthcare/docs/site/Estimating_Sleep_Stages_from_Apple_Watch_Sept_2023.pdf
1 reply
0 recast
0 reaction
shazow
@shazow.eth
Sure, pretty darn good, but the unintuitive thing is that the error rate is not random. If the model does not fit you personally, it will likely be often wrong (rather than wrong 38% of the time or whatever). These numbers are against an average population. I don't see numbers in this paper but there's probably a non-trivial sub-population whose results are more often wrong than correct, and it's hard to tell which population you're in unless you cross-validate yourself. That is, intuitively people assume that say a 50% error rate means it'll be wrong 50% of the nights for everyone, but in this case it's closer that it'll be wrong for some people 100% of the time and 0% of the time for another half of people. (Hypothetical extreme numbers to make a point)
1 reply
0 recast
0 reaction
