Global physical activity resilience during the COVID-19 pandemic: an observational study of 5.17 million exercise app users across 65 countries
Yichun Fan, Yuchen Chai, Fábio Duarte, Stefanie Jegelka, Siqi ZhengIntroduction
Physical activity (PA) is a critical health behaviour often disrupted by public health crises. Prior research has primarily relied on self-reported surveys, limiting timely and scalable insights. This study examines how large-scale digital exercise data can enhance real-time PA prediction and support interventions to strengthen PA resilience during global public health crises.
Methods
We analysed PA data from 5.17 million individuals across 65 countries, collected via a smartphone PA tracking app between January 2019 and June 2021. The dataset includes longitudinal exercise histories, sociodemographic characteristics, geographic locations and social network connections. We developed statistical and machine learning models—including linear regressions, neural networks (NNs) and graph neural networks (GNNs)—to predict weekly PA adherence (defined as at least 150 min of exercise per week in line with WHO recommendations) before and during the COVID-19 pandemic. Models predicted each week’s adherence using the individual’s exercise history from the preceding 1–3 weeks, together with sociodemographic, weather and pandemic-related variables. We also applied surrogate models and econometric analyses to assess the associations of individual and urban-level attributes with PA resilience globally.
Results
In a US county subsample where all model classes could be compared (Middlesex County, MA), the GNN achieved the highest predictive performance for weekly PA adherence, with 89.9% accuracy pre-pandemic and 82.1% during the pandemic. At the global level, the NN model demonstrated similar predictive performance (accuracy 87.6% and 82.8%). Personal exercise history was the strongest predictor, improving model accuracy by 20 percentage points, while increased model complexity or the addition of social network features yielded modest gains. Marked global disparities were observed: cities in lower- and upper-middle-income countries experienced greater reductions in PA during the pandemic compared with cities in high-income countries. Females, socially connected individuals and residents in cities with better access to outdoor sports infrastructure were more likely to maintain regular PA during the pandemic.
Conclusion
Digital PA data can enable accurate, real-time prediction of PA adherence and inform individual and environmental factors influencing global PA resilience during crises. This approach holds promise for informing targeted interventions and urban policy strategies to enhance health resilience and reduce global health disparities.