Machine Learning Models For Predicting Laboratory-based Physical Activity Type From Consumer Wearables In Children

First predicting the type of activity (e.g., lying down, enrichment) then developing activity-specific equations could improve estimates of free-living physical activity energy expenditure in children. While studies have shown pattern recognition approaches trained on accelerometer data can predict children’s free-living activity type, there is potential to improve the classification of children’s activity using accelerometry (accel) and heart rate (HR) data. PURPOSE: This study evaluated underlying accel and HR features from consumer wearables to predict children’s activity type compared to accel and HR data from research-grade devices. METHODS: Children (n = 191, 5-12 years, 53% male, 57% White) completed a 60-minute protocol consisting of simulated free-living activities (e.g., walking, running, soccer, etc.). These activities were directly observed and categorized into four activity classes: Lying Down, Enrichment, Fundamental Movement Skills (FMS), and Sports/Games. Random Forest (RF) models were trained on the underlying accel and HR data from wrist-placed consumer wearables (i.e., Watch Series 7 and Vivoactive 4S) and the combined accel and HR data (i.e., GT9X accel+ Actiheart 5 HR) from a wrist-placed, research-grade accelerometer and a chest-placed HR monitor. Thirteen time and frequency domain features were extracted at each second, and 21 additional features were extracted by looking at a neighborhood of 60-seconds and using sliding windows. Leave-one-subject-out cross validation was used to evaluate model performance. RESULTS: Underlying data from Watch Series 7 exhibited the highest accuracy (90.8%, 95%CI: 90.7%, 90.9%) followed by GT9X accel+ Actiheart 5 HR (87.6%, 95%CI: 87.6%, 87.7%), and Vivoactive 4S (85.5%, 95%CI: 85.4%, 85.6%). Watch Series 7 also exhibited the highest sensitivity (87%, 91%, 92%, 78%) and specificity (98%, 95%, 94%, 97%) across the four activity classes, respectively. CONCLUSION: These results demonstrate the potential of RF models trained on underlying data from consumer wearables to predict children’s activity with similar (or better) precision than research-grade devices. Future studies should evaluate whether models trained on data from consumer wearables can predict children’s activity with high accuracy in free-living environments. Research reported in this abstract was supported by the National Institute of Diabetes and Digestive and Kidney Diseases under Award Numbers F31DK136205 & R01DK129215. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.