Title: Modelling wearable data via quantile-based distributional data analysis
Authors: Vadim Zipunnikov - Johns Hopkins University, Bloomberg School of Public Health (United States) [presenting]
Abstract: With the advent of continuous health monitoring via wearable devices and digital sensors, users now generate their own unique stream of continuous data such as minute-by-minute heart rate or blood pressure. Aggregating these streams into scalar summaries ignores the distributional nature of data and often leads to the loss of critical information. We propose to capture the distributional nature of wearable data via user-specific quantile functions and develop flexible methods to analyze these functions as hybrid functional-distributional data. Traditional approaches of using a single distributional summary such as moments or extremes become special cases of the proposed method. Specifically, the use of L-moments to represent quantile-functions via interpretable decompositions allows us to define an interpretable distance between distributional observations. The quantile functions and L-moments are shown to be flexible to be employed within generalized scalar-on-function regression models and for analyzing joint and individual sources of variation of multimodal data. The proposed methods are illustrated in a study of the association between accelerometry-derived digital gait biomarkers with Alzheimer's disease (AD) and cognitive functioning. Our methods allow digital biomarkers of gait such as step velocity, cadence, stride regularity and mean stride time to be highly discriminatory for subjects with AD and impaired cognitive performance.