University of Illinois Urbana-Champaign, Champaign, IL
RMA DelaysOur Admin Portal website is currently experiencing technical difficulties, and it could result in delays with RMAs being processed. We are currently working to resolve these issues. We apologize for the inconvenience.
Join us on August 11th for an ActiGraph webinar hosted by Xtalks:
Oncology Research and Care: Reimagining Digital InnovationRegister Now
Recognizing Physical Activity Patterns Individually Using Hidden Markov Models
- Presented on May 31, 2013
Accelerometers offer valid and objective measures to assess free-living physical activity (PA) (Kodama et al, 2002; Lyden et al, 2012). While these devices provide rich information with frequent readings over a period of time, the large amount of data is difficult to analyze and special techniques to extract features are needed. Summarizing accelerometer derived data is also problematic since people are active at different times and exhibit diverse activity patterns. Using threshold cut offs to categorize the data ignores individual differences. Hidden Markov Models (HMM) are a way to recognize patterns and classify the data on an individual basis. Additionally, HMM provide other useful information that includes the probability the subject transitions from one category to another.
Purpose To recognize patterns in daily physical activity and extract features from a large accelerometer data set using HMM.
Methods Accelerometer determined PA data with by minute readings for seven consecutive days using ActiGraph monitors was obtained from the National Health and Nutrition Examination Survey (NHANES) 2005-2006 data for 1312 participants, 47% male, aged 6-85 with an average age of 32.09 yrs. (SD = 22.07). For each subject, a four state Gaussian continuous HMM was fitted to the bivariate distribution of step and intensity counts to classify daily PA patterns
Results The data were classified into four categories via HMM: sleep or non-wear, low or sedentary, moderate and vigorous activity. Average step and intensity counts for each category were estimated for each subject. The probability of transitioning between each state was also estimated for each subject. For example, for the average woman performing a moderate activity, the probability that she will continue doing the activity is .5, while the probability that she will change to a sedentary activity is .15. For the average man, the probability of staying with the moderate activity is .62 and moving to a sedentary activity is .12. Thus, along with classification of PA on an individual basis, HMM offers additional information about the subjects’ behavior.
Conclusion Using HMM to identify sleep, sedentary, moderate and vigorous PA provides an individualized classification of the data and extraction of valuable features using both step and intensity counts.