|
Abstract:
|
Sleep parameters such as sleep onset, duration and quality are important markers that have known to be closely related with health outcomes. The conventional tracking methods in observational studies using sleep diary are subject to self-report bias and missingness. The Polysomnography (PSG) as gold standard, is limited in application due to challenges in implementing in real living environment. The wearable devices have provided opportunities of tracking wrist movement in high frequency over a prolonged time period and are increasingly used as surrogates to quantify sleep in both commercially and in research studies. However, the existing sleep algorithms using accelerometry data mostly classify sleep window as periods when activity intensities or variations are below a self-determined threshold, separately for each night. We proposed a latent class modeling approach that builds upon the existing the existing sleep algorithms and classify sleep windows using multivariate measures generated from accelerometry devices, including the raw three-axis activity data, the light exposure, temperature, date as well as the empirical prior distribution of sleep at a given time stamp.
|
