Most sleep apps ask the watch to do a lot overnight, and by morning the battery is gone. We wanted the opposite.
We built a sleep analysis system with sleep scientists where a deep CNN classifies sleep stages from the watch sensor data. It reaches 94.2% accuracy and uses under 18% of the battery overnight. It also predicts REM and disrupted sleep, and in testing it cut average sleep latency by 22 minutes.
The part we like most is how it got built: a 5-person team, 14 weeks, a clinical-grade dashboard for healthcare professionals, and human trials completed ahead of launch.
If you have a health or wearable idea and you are not sure whether the model, the battery or the data is the hard part, tell us what you are working on and we will say honestly where we think it breaks.