Multiscale and Multimodal Characterization of Mobile Sensor Data
Abstract
Physiological and behavioral processes unfold on multiple time scales. Traditional time-series analysis tools are designed to capture stationary, single-scale processes, which may miss important information. Several methods have been proposed in recent decades to capture multiscale properties of time series, such as detrended fluctuation analysis. This report examines the way that multiscale measurement of physiology and behavior fits passive sensing data from the Fitbit Charge 4 mobile sensor. Physical activity and heart rate (HR) data from a large, long-term study of office workers were analyzed using traditional time-series analyses and a newly developed multiscale method: multiscale regression analysis. These analyses were conducted at the day and month level. Results indicate that multiscale analyses lead to substantial improvements in model R2 over single-scale analyses for autocorrelation analyses of HR and steps (13 percent to 108 percent increase) and for the cross-correlation or coherence between HR and steps (21 percent to 88 percent increase). Multiscale analyses that led to better fit statistics were most advantageous when considering physical activity as compared with HR. Overall results suggest that physiology and behavior in daily life are better captured by estimating multiscale rather than single-scale processes.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 2023
- Accession Number
- AD1192597
Entities
People
- Alexander F. Danvers
- Esther Sternberg
- Evan C. Carter
- Lidia S. Obegon
- Matthias R. Mehl
Organizations
- United States Army Research Laboratory
- University of Arizona