Fred Hutchinson Cancer Research Center, Seattle, USA
Our office will be closed Thursday and Friday, November 26th and 27th for the Thanksgiving holiday. We will reopen at regular business hours on Monday, November 30th.
Functional data analysis methods for physical activity measurements using accelerometers
- Presented on 2015
Purpose: To propose a novel functional data analysis framework to fully characterize activity intensity, duration and frequency based on accelerometer data.
Background: Traditional approaches reduce accelerometer data into simple summary measures, such as time and bouts in intensity categories (sedentary, light, moderate and vigorous activities). However, these approaches depend on specifying cut points for intensity or thresholds for bout length (e.g., 10 minutes), which are often subjective, and often omit refined information within categories.
Methods: 6507 women aged 63 to 99 years wore an Actigraph GT3X+ accelerometer in the Objective Physical Activity and Cardiovascular Health Study (OPACH). Accelerometer data were transformed and analyzed by a functional data analysis framework. Two functional indices were proposed to characterize the distribution of intensity and bout length continuously, free of cut points or thresholds. Functional principal component analysis and functional regression models were adapted to understand major modes of variations of activity profile and their association with health outcomes. These methods were applied to accelerometer data from the OPACH.
Results: The functional indices demonstrated the distribution of activity intensity and bout lengths of older women. Two principal components were found to explain most variation in activity intensity profile, one indicating the overall activity level across all intensity categories and the other describing relative shifts of light versus moderate and vigorous activities. These components were significantly associated with glucose and insulin levels among older women.
Conclusions: A flexible and interpretable modelling framework was proposed to utilize the rich information in high resolution accelerometer data.