Abstract:
|
There is a critical need to understand the temporal dynamics of depression using real-time objective measures. The Intern Health Study (IHS) seeks to identify predictors of depressive symptoms by following a large cohort of medical interns. For the entire medical internship year, the IHS uses a phone app to collect the interns' daily self-reported mood scores (1-10), and uses a smartphone and wristband to collect objective measurements such as minute level activity data, nightly sleep time and duration, heart rate data, and geolocations via a smartphone or wristband.
We introduce a flexible multivariate time series model to analyze multiple sensor data streams collected at different time scales with occasional missingness (due to failure to wear wristbands or carry smartphones). Our model predicts interns' mood and estimates the lagged effects of each data stream by sharing information both across time, to account for smooth time-varying associations, and across similar subjects. We illustrate our methods using data from the 2015-16 IHS cohort recruited at University of Michigan. Lastly, we discuss computational issues and the practical implications of our results.
|