Abstract
Obesity affects a rising percentage of children and adolescents, decreasing quality of life and increasing risks for comorbidities. While the causes are known, obesogenic behaviors result from complex interactions with the living environment. The BigO project utilizes large-scale behavioral and environmental data to support policy making. This work proposes a novel approach for modeling expected population behavior as a function of the local environment. Using data from 156 children, we trained models to predict physical activity levels in small geographic regions with 81% accuracy. These models enable the automatic generation of heatmaps, providing public health authorities with tools to visualize the spatial distribution of behavior.
Introduction: Addressing the Obesogenic Environment
Childhood obesity is a global epidemic driven by "obesogenic" environments—surroundings that promote low energy expenditure and high calorie intake. Public health authorities aim to design "active environments" but often lack the granular data needed to understand how neighborhood characteristics translate into population behavior. Traditional monitoring methods, such as surveys, are often too small-scale, while continuous GPS tracking raises significant privacy concerns. This research addresses a critical question: Can we build a reliable model that predicts the physical activity of a population using only publicly available environmental data?
Methodology
The researchers developed a pipeline to bridge individual sensor data with aggregate environmental characteristics.
Data Collection and Privacy
The study utilized data from the BigO project (Big Data Against Childhood Obesity), involving 156 children and adolescents (ages 9-18) in Thessaloniki, Greece. Participants provided approximately 90,890 hours of accelerometry and GPS data via smartphones and smartwatches. To protect privacy, individual residences were mapped to geohashes—rectangular grid cells measuring roughly 610m x 930m.
Feature Engineering and Modeling
Two sets of attributes were extracted:
- Behavioral Indicators: The average "activity counts per minute" for residents within each geohash.
- Environmental Attributes: Counts of four Point of Interest (POI) types (athletics/sports, fast food, public parks, and cafes/bars) extracted from OpenStreetMap and Google Maps.
A Random Forest (RF) classifier was trained to perform binary classification, predicting whether a geohash would exhibit "High" or "Low" physical activity relative to the mean. Given the dataset size of 42 unique geohashes, a Leave-One-Out (LOO) cross-validation procedure was employed.
Results and Insights
The Random Forest model achieved an overall accuracy of 81%. The model demonstrated strong precision (0.81 for both classes) and recall (0.88 for Low, 0.72 for High), making it particularly effective at identifying sedentary neighborhoods that may require intervention.
Qualitative Validation via Heatmaps
The model's predictions were used to generate heatmaps of Thessaloniki. Qualitative analysis showed that areas predicted as "High" activity (e.g., geohashes E, G, and H) correlated with high densities of parks, pedestrian streets, and sports facilities. Conversely, densely populated residential areas lacking these amenities (e.g., geohash F) were correctly flagged as "Low" activity zones.
The Power of Behavioral Imputation
The core insight is that the urban environment serves as a strong proxy for behavior. This allows for "behavioral imputation," where health authorities can generate comprehensive city-wide heatmaps even when sensor data is available for only a small fraction of the population. This "privacy-by-design" approach focuses on population-level averages rather than individual trajectories.
Limitations and Future Directions
While promising, the study was limited to 156 participants across 42 geohashes in a single city. Future research should expand to diverse urban fabrics (e.g., sprawling North American cities vs. dense European centers) to improve generalization. Additionally, incorporating more granular features—such as sidewalk quality, street lighting, or crime rates—could enhance predictive power. While the paper assumes a link between environment and behavior, these relationships remain complex; future iterations could explore temporal dynamics, such as seasonal changes or the impact of specific urban redevelopment projects.
Summary updated