B0727
Title: Nonhomogeneous hidden Markov models to leverage routine in physical activity monitoring with informative wear time
Authors: Corwin Zigler - University of Texas at Austin (United States)
Beatrice Cantoni - University of Texas at Austin (United States) [presenting]
Abstract: Missing data due to device nonwear is ubiquitous when commercial-grade wearable devices are deployed in free-living conditions for extended time periods. To accommodate the threat that device wear patterns may well associate with underlying activity, a nonhomogeneous hidden Markov model (NHMM) is offered that permits transitions between latent physical activity states to depend on time-varying exogenous input variables. It evaluates how data on both routine activity patterns and abrupt shocks can be used to justify further assumptions that wear time is missing at random, offering the potential to improve missing data imputation over existing methods that regard wear time as missing completely at random. The modelling relies on a Polya-Gamma data augmentation approach, leading to the development of an efficient Markov chain Monte Carlo (MCMC) sampling scheme with straightforward extension to missing data imputation. It is shown how the proposed methods can improve inference on evolving physical activity among a cohort of adolescent young adult cancer patients who exhibit some degree of regular physical activity and experience treatment changes during the observation period.
