Past Activity Best Predicts Weekly Exercise, BMJ Study Finds
What you did last week serves as the strongest predictor of whether you will exercise this week, according to an observational study published in BMJ Public Health. Researchers examining 662 million user-week records from the app Strava found that a person’s recent activity history outweighs weather, pandemic policies, and social connections in forecasting adherence to physical activity routines.
-
Key Clinical Takeaways:
- Recent activity history across the preceding one to three weeks improved prediction accuracy for weekly exercise adherence by nearly 20 percentage points during the COVID-19 pandemic.
- High-income countries showed an immediate increase in recorded activity after March 2020, whereas lower- and upper-middle-income regions experienced greater disruptions.
- Greater access to outdoor sports amenities, particularly parks and pitches, consistently correlated with longer recorded weekly exercise durations across all pandemic phases.
Research Team Analyzes 5.17 Million Active App Users
The research team evaluated exercise records spanning from January 1, 2019, through June 30, 2021, reported. By filtering an initial pool of 30 million accounts down to individuals who joined before 2019, posted at least 30 activities, and lived in regions with more than 1,000 users, the authors analyzed 5.17 million active participants across 1,441 administrative regions in 65 countries. Weekly physical activity adherence was defined as recording at least 150 minutes of exercise, aligning with World Health Organization recommendations.
Before the pandemic, the mean weekly exercise duration across the cohort sat at approximately 90 minutes. Running accounted for 41% of all logged activities, cycling made up 36%, walking contributed 9%, and hiking added to the predominantly outdoor pattern, while structured gym workouts represented just 4% of the dataset. To identify what drove these patterns, the authors compared linear regression, neural network, and graph neural network models. They found that adding activity duration from the prior one to three weeks boosted model accuracy by nearly 20 percentage points during the pandemic when compared against profiles using weather, user demographics, and pandemic indicators alone.
Income Levels and Urban Features Affect Activity Resilience
The study revealed a sharp divergence in how different global regions responded to public health disruptions. Users residing in high-income countries displayed an immediate rise in recorded activity following March 2020, while middle-income settings faced prolonged declines. App representation leaned heavily toward developed economies, meaning the data underrepresented low-income countries.
Urban features played a measurable role in exercise resilience. More economically developed cities experienced larger activity drops during strict lockdowns, but those same urban centers also demonstrated greater long-term recovery through June 2021. Higher sports-amenity density correlated positively with longer recorded weekly exercise durations. Parks and outdoor pitches consistently linked to higher activity volumes, whereas indoor fitness centers and swimming pools showed negative or nonsignificant associations during periods of restriction.
Weather and Gender Influence Exercise Activity Levels
Environmental conditions influenced activity levels predictably. High wind speeds, heavy precipitation, and thick cloud cover showed negative associations with recorded exercise, while better visibility correlated with increased activity. Temperature displayed an inverted U-shaped relationship, with optimal ranges fluctuating depending on local humidity levels.
Demographically, female users exhibited stronger activity resilience than males within the model estimates. This gap widened significantly for walking and hiking, where the difference between female and male activity increased more than fourfold compared with pre-pandemic levels. However, the study authors noted that only 25% of users in the dataset self-declared as female, which limits the broader interpretation of gender comparisons. Social-network size maintained a positive, though modest, association with adherence, improving prediction accuracy by roughly 0.3 percentage points while demanding substantial computational overhead.