|
Abstract:
|
Accelerometry data provides a promising avenue as individual-specific data that can benefit precision health frameworks. Other studies analyze accelerometry data by using summary statistics (eg single-axis count data) and applying regression-based cutoffs to classify activity levels (Vigorous, Moderate, Light, and Sedentary). However, these cutoffs are often not generalizable across populations, devices, or studies. Thus, a more generalizable data-driven approach to analyze activity data is necessary. We holistically consider a subject’s activity profile using Occupation-Time curves, which describe the percentage of time spent at or above a continuum of count levels. We develop multi-step adaptive learning algorithms to analyze the Occupation-Time curves under both an L1 and L0 framework. The L1 learning algorithm incorporates a hybrid approach of fused lasso and Hidden Markov Model, as well as refinement learning steps, to identify activity windows of interest. Similarly, we develop a functional-data analysis approach under the L0 framework to identify and estimate important activity windows. We demonstrate these methods using simulations as well as real world data analysis.
|
