What you did last week may be the strongest clue to how much you'll exercise this week

When COVID-19 disrupted daily routines worldwide, millions of exercise records revealed striking differences in how people stayed active. Researchers examined the personal habits and urban conditions behind these changing patterns.

Want to read later? Download your clean, ad-free, print-friendly PDF copy by clicking here.

In a recent study published in the journal BMJ Public Health, researchers evaluated whether digital exercise records could predict adherence to physical activity and identified individual and urban factors associated with resilience in physical activity during the coronavirus disease 2019 (COVID-19) pandemic.

Background

About 31% of the global population does not meet the recommendation of at least 150 minutes of moderate-intensity exercise weekly. Physical inactivity is associated with cardiovascular disease, obesity, diabetes, cognitive impairment, and depression. Public health emergencies can disrupt routines through facility closures, movement restrictions, and infection concerns.

During the COVID-19 pandemic, studies reported differing changes in activity, but many relied on small, self-reported surveys vulnerable to recall and social-desirability bias. These limitations make global comparisons difficult. Further research using longitudinal digital records is needed to clarify how personal and urban conditions shape activity resilience across settings.

About the study

The researchers analyzed exercise records from Strava, a physical activity tracking app, covering January 1, 2019, through June 30, 2021. From 30 million initial users, they retained people who joined before 2019, posted at least 30 activities, and lived in regions with more than 1,000 users. They excluded implausible activity durations and retained the 10 most common activity types.

The final dataset included 5.17 million users across 1,441 administrative regions in 65 countries, providing 662 million user-week records. Weekly physical activity adherence was defined as at least 150 minutes of exercise recorded in Strava, using a duration threshold based on World Health Organization recommendations.

The team combined activity histories with user characteristics, weather, pandemic indicators, lockdown policies, social-network information, sports-amenity locations, population density, and nighttime-light measures used as a proxy for economic development. They compared linear regression, neural network, and graph neural network (GNN) models to predict weekly exercise duration and classify adherence using the 150-minute threshold.

GNN analysis was limited to a connected subsample of 4,271 users in Middlesex County, Massachusetts. Performance was assessed using accuracy and F1 scores, which account for both missed active users and false-positive classifications. Shapley Additive Explanations (SHAP) ranked predictors by their contribution to model outputs. A surrogate model and regression discontinuity design examined individual- and city-level associations and short-term changes around pandemic onset.

Study results

Recorded activity was concentrated in developed countries and several large developing countries. Running accounted for 41% of activities, cycling 36%, walking 9%, and hiking contributed to the predominantly outdoor pattern; gym workouts represented 4%.

Before the pandemic, mean weekly exercise duration was about 90 minutes. Activity shifted after the pandemic began in March 2020, but changes differed by country income group. Users in high-income countries showed an immediate increase, whereas regions in lower- and upper-middle-income countries experienced greater disruption. The app data underrepresented low-income countries.

Personal exercise history was the strongest predictor of weekly adherence. Adding activity duration from the preceding one to three weeks improved accuracy by nearly 20 percentage points during the pandemic in the Middlesex County analysis, compared with models using user profile, weather, and pandemic variables alone. This also improved the models' ability to distinguish active from inactive users. The best GNN achieved 89.9% accuracy before the pandemic and 82.1% during it.

Neural networks achieved similar performance globally, with accuracy of 87.6% before and 82.8% during the pandemic. Adding social-network structure improved pandemic-period prediction accuracy by only about 0.3 percentage points, while requiring substantial computation. SHAP analysis also confirmed that recent activity history was more influential than weather and social connectivity.

High wind speeds, heavy precipitation, and greater cloud cover were negatively associated with activity, while better visibility was positively associated. Temperature showed an inverted U-shaped relationship, with the apparent optimal range varying by regional humidity. Social-network size was positively associated with adherence before and during the pandemic.

Female users showed stronger activity resilience than males in model estimates for this period. Differences were particularly pronounced for walking and hiking, where the female–male gap increased more than fourfold compared with pre-pandemic levels; gaps for yoga, workouts, and weight training increased approximately 1.5 to 1.6 times. Only 25% of users self-declared as female, limiting wider interpretation of these comparisons.

The immediate pandemic effect varied among administrative regions, ranging from positive to negligible or negative. Short- and long-term effects showed a modest positive correlation. Developed-country regions generally increased or recovered over time, while Colombia and India had some of the largest longer-term decreases through June 2021.

More economically developed cities experienced larger activity reductions during lockdowns but also greater long-term recovery. Higher sports-amenity density was positively associated with longer recorded weekly exercise duration across pandemic phases.

A country-level analysis found this relationship was stronger where COVID-19 case burdens were higher. Parks and outdoor pitches were consistently associated with more recorded exercise, whereas indoor facilities, including fitness centers and swimming pools, had negative or nonsignificant associations.

Conclusion

The findings indicate that digital exercise records can help predict weekly physical activity adherence, particularly when models include a person’s recent activity history. The pandemic’s effects varied widely across cities, with more persistent declines observed in many middle-income settings.

These results could inform timely, targeted support and urban planning for public health emergencies. Still, Strava users may not represent the wider population, particularly because the sample comprised established app users recording predominantly outdoor activities.

Missing information on age, income, and education adds uncertainty, and the findings should not be generalized directly to the entire population.

Source:

Journal reference:

• Fan, Y., Chai, Y., Duarte, F., Jegelka, S., & Zheng, S. (2026). Global physical activity resilience during the COVID-19 pandemic: An observational study of 5.17 million exercise app users across 65 countries. BMJ Public Health. 4. DOI: 10.1136/bmjph-2025-003670, https://bmjpublichealth.bmj.com/content/4/4/e003670