A0719
Title: Integration of longitudinal physical activity data from multiple sources
Authors: Jingru Zhang - Fudan University (China) [presenting]
Haochang Shou - University of Pennsylvania (United States)
Hongzhe Li - University of Pennsylvania (United States)
Abstract: As various devices have been developed to collect physical activity data, a critical problem is how to integrate datasets across different conditions to better understand the characterization of physical activity. The key to this problem is to remove site effects while maintaining common features. However, since wearable sensor devices are deployed to record physical activity minute-by-minute continuously over multiple days, the longitudinal time-dependent structure makes the integration challenging. A new method is proposed to integrate longitudinal physical activity datasets, which model the shared information by common eigenvalues and eigenfunctions while allowing for site-specific scale and rotation. The proposed method is applied to NHANES datasets with different types of wearable sensors. The results demonstrate the method's superiority in removing site effects while preserving biological signals compared to existing approaches. A framework is developed for integrating longitudinal time-dependency datasets and provides insights into the analysis of physical activity data.
