It doesn't seem like a huge problem for ML to solve. Have someone wear a consumer EEG headband and an expensive medical device which measures their actual sleep

It doesn't seem like a huge problem for ML to solve. Have someone wear a consumer EEG headband and an expensive medical device which measures their actual sleep state. Penalize the consumer device when its findings are off of the medical device. If the consumer device is still inaccurate after many iterations, add another input like a wristband with multiple sensors. Use both the EEG band and wristband as inputs to the ML, and train the ML until it's close in accuracy to the expensive medical device. In my mind, there just has to be some combination of easy measurable inputs that, together, correlate very well with sleep state.