Sleep versus awake from a sensor on the wrist is always going to be subject to wide error bars.
Source: worked on an experiment trying to make people fall asleep while driving (in a simulated environment so it was safe- I built the simulated environment). It took a lot of EEG wires and gel in subject's hair etc. to accurately figure that out.
Source: Worked on the ResMed S+ and got a theory dump from the PhDs running the project.