| Go do an actual sleep study some time. You'll be hooked up to a ton of sensors, and in the morning a group of doctors pour over the results, and they'll basically vote to decide on what the data means. "Oh this looks like REM". So, to calibrate a sleep tracking device, you have a person wear the device, while also doing the sleep study. You do this a bunch of times. You train some ML models to try and make the outputs from the sensor data, after processing, the same as the study data. After some degree of accuracy you declare success. Now, does it work? In broad strokes, yes. You can (easily!!) see the effect of alcohol on sleep quality. If you have a crap night vs a good night, sure, a wrist based consumer device can figure that out. Actual details? Eh. I wouldn't trust the devices for anything but directional data. The more sensors devices get, the better than ML model can be trained. Now it has been awhile since I last worked on this stuff (I actually just sat next to the people doing the work), so maybe there is some revolutionary new technique out there, but if not, it is still ML models trying to correlate things and match them up to what a bunch of fancier sensors said during studies. |