Cardiovascular
Association Between Urban Greenness and Machine Learned Physical Activity Measures: A Cross-Sectional Study Using UK Biobank Hiroshi Mamiya* Yacine Lapointe Lapointe Lapointe McGill University
Background: Physical activity is a major preventable risk factor for chronic disease and mental health disorders. Residential proximity to urban greenness has been shown to positively influence physical activity. However, most studies have relied on self-reported measures of physical activity, which are prone to bias. Wearable devices can provide more objective measures of activity duration and intensity using accelerometry signals. This study compared associations between urban greenness and physical activity using self-reported measures and three accelerometer-derived methods, including a novel Machine learning–based approach.
Methods: We analyzed UK Biobank accelerometer data from 81,582 urban participants aged 40–69 years, collected in 2013. Exposure to urban greenness was measured using the Normalized Difference Vegetation Index within a 300 m residential buffer. Participants’ mean weekly duration of light physical activity was derived from self-reports and 7-day wrist-worn accelerometers. Three accelerometer-based metrics were generated: Vector magnitude, Activity counts, and a Machine learning–based measure. Linear mixed models accounting for within-city clustering were adjusted for diet, transportation, socioeconomic characteristics, and neighbourhood environmental factors, and stratified by household income.
Results: The Machine learning-based measure demonstrated the strongest agreement with gold-standard measures and the strongest positive association with greenness (Fig. 1), particularly among low-income households. Accelerometer-derived metrics showed curvilinear relationships, whereas self-reported physical activity showed weaker and inconsistent associations.
Conclusions: These findings support evidence that urban greenness promotes higher physical activity. Machine learning-based accelerometer processing yielded stronger and more consistent associations than traditional methods, highlighting its value for physical activity research.
