“Harmonization of Longitudinal Physical Activity Intensity Measures from Multiple Sources”

Presented by Haochang Shou
Department of Biostatistics, Epidemiology and Informatics
University of Pennsylvania

Wearable devices and digital phenotyping have been increasingly included in observational and interventional studies. However, it remains challenging to combine and compare data collected across different studies or cohorts due to large variations in the choice of device types, acquisition and preprocessing protocols. The key to this problem is to remove unwanted study or site effects while maintaining common biological variations. However, since wearable sensor devices are designed to continuously record data at high temporal resolution over several days, the longitudinal time-dependent nature of the collected data complicates the integration process. We propose a new method to integrate longitudinal physical activity intensity time series data collected from accelerometers. The method models the shared information by common eigenvalues and eigenfunctions while allowing for domain-specific scale and rotation. We apply the proposed method to two cohorts within the NHANES study where the physical activity measures were collected with different generations of accelerometers and processed into different units. Results demonstrate the superiority of our method in removing site effects while preserving biological signals as compared to existing approaches.