|
Abstract:
|
Wearable computing sensors have shown increasing promises to capture subjects’ detailed biometric profiles in free-living conditions. Biomarkers extracted from sensor data are now included in several randomized trials as novel endpoints to assess intervention or treatment effects. Unlike the traditional scalar endpoints that are often observed only once, sensor data collected from the devices could be multivariate objects or functions and are observed repeatedly over days to capture the biological fluctuations. The conventional causal inference framework based on counterfactual outcomes might not be directly applicable. We propose a novel statistical method that defines and estimates the treatment effects by accounting for random assignments and meanwhile handling uncertainty through repeated observations. Our work was motivated by the daily actigraphy and smartphone data from the Individualized Comparative Effectiveness of Models Optimizing Patient Safety and Resident Education (iCOMPARE) study, which was a multisite randomized trial of work hour standards for internal medicine interns. We will evaluate the intervention effects on interns’ sleep, cognition, and motor behavior.
|
