Leveraging Accelerometry and Heart Rate Data from Consumer Wearables to Predict Physical Activity in Children: A Device Agnostic Approach.

Introduction: This study examined the potential of a device agnostic approach for predicting physical activity energy expenditure (PAEE) from research-grade and consumer wearable accelerometry and heart rate (HR) raw data compared with indirect calorimetry in children.

Methods: Two hundred thirty-one 5- to 12-yr-olds (52.4% male) of diverse skin tone and body weights participated in a 60-min protocol with multiple activities at varying intensities. Children wore two of three consumer wearables (Apple Watch Series 7, Garmin Vivoactive 4S, Fitbit Sense) and a research-grade accelerometer (ActiGraph GT9X) on their nondominant wrist, and a chest-placed, research-grade HR monitor (Actiheart 5, ECG), concurrently. Children also wore a K5 criterion measure of PAEE (i.e., COSMED K5). Cross-sectional time series (CSTS), generalized additive mixed effects model (GAMM), and random forest (RF) were used to estimate minute-by-minute PAEE from features extracted from raw accelerometry and HR data. Variance explained ( R2 ), in addition to other metrics, evaluated agreement between estimated and criterion measurements.

Results: For the research-grade devices (i.e., ActiGraph accelerometry and Actiheart HR), R2 values were 0.74, 0.74, and 0.76 for CSTS, GAMM, and RF, respectively. For Apple, R2 values were 0.77, 0.76, and 0.78; Garmin's values were 0.73, 0.73, and 0.75; and Fitbit's values were 0.63, 0.65, and 0.67 for CSTS, GAMM, and RF, respectively. Across all other evaluation metrics, a similar pattern was observed with Fitbit performing the worst but with little variability between the modeling approaches or the other devices.

Conclusions: Except for Fitbit, accelerometry and HR data from consumer wearables predicted PAEE comparably to research-grade devices, and there was little variability across modeling approach. These outcomes support deploying a consumer wearable device-agnostic approach for PAEE estimation in children.