> It frequently counts me as sleeping when I sit down to watch TV in the evening, or read a book in bed. This is why I pushed for sleep on v1 of Microsoft Band

> It frequently counts me as sleeping when I sit down to watch TV in the evening, or read a book in bed. This is why I pushed for sleep on v1 of Microsoft Band to be manually triggered. Eventually cloud detected sleep was lit up, but we benefited from lots of gathered data at that point, but even so, it is a hard problem to solve. The best solution would be per user training, allowing users to correct auto detection and use that to improve future tracking, but AFAIK no one in industry is doing that. > I'm not sure how it makes an assesment of my sleep level. I don't know if it's just based on movement or if it takes heart rate data into account This is an incredibly hard problem. For sleep studies, multiple doctors look over all the vitals gathered (and for a full sleep clinic study, subjects are very wired up!) And to determine when the subject is in a given phase of sleep, they vote, and it is not always a consensus decision![0] Trying to then make a machine algorithm out of that is difficult, especially working with consumer level sensors. That said, across most users, data tend to be mostly correct, which is what these consumer devices target. Being able to say someone got 7.5hrs of sleep is good enough, the 10 minute bathroom break doesn't impact the overall numbers too much. Being able to give a good summary of weeks and months is huge, and being able to spot large regressions and improvements is key to user satisfaction. The ultimate goal is to gather a lot of data, noise or not, run it through machine learning, and start being able to give people actionable advice that they will see real results from. [0] this was relayed to me by the team who doing sleep classification, other sleep studies may happen differently.