INTACT: A method for integration of longitudinal physical activity data from multiple sources

Wearable devices and digital phenotyping are increasingly used in observational and interventional studies to assess physical activity. However, integrating and comparing data across studies and cohorts remains challenging due to variability in device types, acquisition protocols, and preprocessing methods. A key challenge is removing unwanted study- or device-specific effects while preserving meaningful biological signals. These difficulties are exacerbated by the longitudinal and within-day correlations inherent in high-resolution time series data collected from wearable sensors. To address this, we propose INTACT (INtegration of Time series data from weArable sensors for physiCal acTivity), a novel method for harmonizing physical activity intensity time series from accelerometers. INTACT models shared information through common eigenvalues and eigenfunctions while allowing for domain-specific scale and rotation adjustments. We demonstrate the effectiveness of INTACT on two cohorts from the National Health and Nutrition Examination Survey (NHANES), where physical activity measures were collected using different generations of accelerometers and processed into different units. Our results show that INTACT outperforms existing methods in mitigating domain effects while preserving biological signals, enabling more reliable cross-study comparisons of physical activity patterns.