M. Hegarty-Craver, D.S. Temple, D.E. Dausch, H.J. Walls, E.A. Preble, M.D. Boyce, R.P. Eckhoff
RTI International,
United States
Keywords: wearables, smartwatch, mobile applications, cloud computing, artificial intelligence and machine learning, modular design
Summary:
Background: Consumer wearables provide an opportunity to continuously and objectively measure health remotely, expanding the reach of biomedical research. A recent survey reported that over 50% of adults in the United States under the age of 65 years own a wearable. While technically feasible, using consumer wearables for longitudinal research and Artificial Intelligence (AI) applications remains a challenge because these devices do not need to conform to medical standards. Frequent updates can change the way that the reported health metrics are calculated, which can obscure treatment effects or mean that AI model needs to be re-trained. Vendors also use different algorithms to calculate the same or similar metrics, like resting heart rate, which means that research findings or AI models are device specific. To effectively leverage the data from consumer wearables, researchers need access to the minimally processed sensor data that wearables vendors use to calculate the metrics that they report. RTI International developed the Architecture for Localized Precision Health data Acquisition from Wearables (AlphaWear) to acquire and analyze high-resolution interbeat interval (IBI) and 3-axis accelerometer data from compatible consumer wearables. IBI Data: Our team deployed AlphaWear using Garmin Fenix series smartwatches to understand exposure risk to respiratory viral illness in free-living conditions. We developed an analytics pipeline to transform the IBI data into continuous measures of heart rate (HR) and heart rate variability (HRV) that we standardized by physical activity (PA) and time-of-day to address the impacts of known confounders. This allowed for continuous calculation of risk in contrast to most methods of illness detection that rely on resting measures. We also found that the addition of HRV measures was sensitive to the early immune system response and allowed for earlier detection compared to HR measures alone. Accelerometer Data: AlphaWear has also enabled our team to characterize eating behavior in controlled settings using a Garmin Vivoactive 5 smartwatch versus a research-grade ActiGraph LEAP smartwatch. Detecting when and how meals are consumed is important for advancing the field of precision nutrition because these behaviors affect metabolic and physiological responses. We developed a model using features that were extracted from the 3-axis accelerometer including the average vector magnitude, the variance of the x-axis acceleration, and the dominant frequency over 3s (ActiGraph) and 5s (Garmin) windows as well as the number of z-axis crossings and the time that the vector magnitude remained above a pre-selected threshold for 30s. The models developed using research- versus consumer-grade wearables performed similarly. Discussion: AlphaWear provides control over how data are processed so that researchers can adjust analytics to the population that they are studying (e.g., elderly individuals versus younger, more active adults) or research question that they are trying to answer (e.g., HRV around mealtimes for individuals with eating disorders). Transparent versus black-box processing provide metrics that can be used sustainably in AI applications to provide generalizable models of health and disease. Future directions for AlphaWear include expanding the number of supported wearables and developing analytics pipelines to run more efficiently in hybrid environments.