Dynamic Algorithms to Estimate Physical Activity Energy Expenditure and Mode from a Hip-Mounted Accelerometer

• Activity Number: 597
• Type: Invited
• Date/Time: Thursday, August 4, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): John Staudenmayer*+
• Companies: University of Massachusetts at Amherst
• Keywords: accelerometer ; actigraphy ; dynamic models ; statistical learning

Abstract

We consider the problem of using hip mounted 1-second epoch accelerometer data to estimate physical activity in a free-living population. We develop two related algorithms to estimate energy expenditure (METs, metabolic equivalents) and physical activity mode (what the person who wore the accelerometer was doing). The algorithms combine two ideas. (1) We develop a statistical learning method that uses features in the accelerometer signal to predict the outcomes (METs and mode), and (2) we use dynamic models to take advantage of the observation that both METs and mode tend to vary slowly over time for each person. We estimate the algorithm's parameters and evaluate its performance using both data collected in the lab (accelerometer monitoring combined with indirect calorimetry and directed activities) and direct observation of free-living subjects by trained observers.