Cluster Analysis to Find Temporal Physical Activity Patterns Among US Adults

Abstract

Physical activity (PA) is known to be a risk factor for obesity and chronic diseases such as diabetes and metabolic syndrome. Few attempts have been made to pattern the time of physical activity while incorporating intensity and duration in order to determine the relationship of this multi-faceted behavior with health. In this paper, we explore a distance-based approach for clustering daily physical activity time series to estimate temporal physical activity patterns among U.S. adults (ages 20-65) from the National Health and Nutrition Examination Survey 2003-2006 (NHANES). A number of distance measures and distance-based clustering methods were investigated and compared using various metrics. These metrics include the Silhouette and the Dunn Index (internal criteria), and the associations of the clusters with health status indicators (external criteria). Our experiments indicate that using a distance-based cluster analysis approach to estimate temporal physical activity patterns through the day, has the potential to describe the complexity of behavior rather than characterizing physical activity patterns solely by sums or labels of maximum activity levels.

Machine learning in physical activity, sedentary, and sleep behavior research

“…ML applied to compositional data still requires further development, and additional efforts are necessary to understand how to effectively utilize ML with such data. Thus far, there has been a tendency to rely on the K-means clustering algorithm and its variants for profiling physical activity, sedentary, and sleep behaviors [16,19,72,79], likely due to their simplicity and efficiency. However, future research directions in profiling analysis should explore and assess alternative clustering algorithms.…”