Speaker
Description
Standard recommendations of 7–9 hours of sleep do not reliably guarantee daytime alertness. Using wearable-derived sleep-wake data, we infer two latent physiological states—homeostatic sleep pressure and circadian phase—via a mathematical model, and generate personalized sleep-wake schedules aligned with each individual's circadian rhythm. In two prospective clinical trials, adherence to model-based schedules significantly improved alertness, with circadian alignment proving a far stronger predictor than total sleep time alone. These findings led to deployment of our algorithm across all Samsung Galaxy Watch devices—the first mathematical biology model adopted by a major tech company. Extending this framework, we show that mathematically derived circadian features from sleep-wake data accurately predict mood episodes in bipolar disorder, outperforming models that rely on richer but more invasive data. Finally, we introduce HADES-NN, a neural network method for optimizing circadian models under real-world discontinuous light signals, enabling future model personalization at scale.
