0256 Modeling Individual Sleep Needs from Wearable-Device Data

Abstract

Abstract Introduction There are currently thought to be two processes that interact to regulate one’s sleep/wake cycle: a sleep homeostat, which depends on the amount of time you have been awake; and the circadian rhythm, an internal 24-hour molecular clock present in nearly every cell of the body. Both processes differ between individuals, whether that’s from a decision to pull an all-nighter or from a genetic inclination towards being an “early bird” or a “night owl”. This project focuses on building a model that can predict how sleep disruptions impact an individual’s sleep cycle, and the ability to recover from these disruptions. Methods A “two-process” model with the capability to simulate a theoretical sleep-wake cycle with disruptions was implemented in Python, and is being tuned using machine learning (ML) on actigraphy data. Actigraphy measures motion using a noninvasive accelerometer embedded into a wearable-device (e.g. FitBit), allowing for the study of sleep activity patterns in a real-world setting. Data was collected from approximately 20,000 individuals, who reside in 121 different countries. ML algorithm was trained and tested on synthetic data generated from the “two-process” model, and then applied to the actigraphy dataset to evaluate prediction accuracy. Results Our model has shown that recovery from sleep disruptions can be modeled mathematically, and can be individualized using actigraphy data. The mathematical model was able to successfully simulate the “two-process” interaction, and, by sweeping parameter values, those that yielded a single sleep bout of 8-9 hours/night were identified. ML algorithm tuning with actigraphy data is ongoing. Conclusion Our results suggest that it is possible to predict the sleep patterns of individuals despite disruptions in their schedules from sleep disorders, travel, work, or social and family commitments. Future research should focus on optimization of sleep in individuals using the informational output of our model, and extending it to incorporate the behavioral and/or clinical changes that have proven to improve an individual’s sleep. By incorporating this additional set of data, the resulting model can not only predict the impact of disruptions, but also provide recommendations for prevention and recovery based on an individual’s input about their lifestyle and their sleep. Support (if any)