Predicting census-tract obesity prevalence using integrated environmental, built-environment, and neighborhood data: a scale-aware geospatial machine-learning proof-of-concept

Source abstract

Obesity prevalence varies markedly across neighborhoods, reflecting environmental, built-environment, and social influences, yet few studies integrate primary environmental measurements with contextual spatial data under a scale-aware predictive framework. As a proof-of-concept, we evaluated whether household-resolution measurements, laboratory-measured soil metals (Pb, Cd, Zn, Se, Cu) together with housing, soil-context, built-environment, and sociodemographic variables, can be harmonized into a spatial-block machine-learning framework to generate small-area obesity predictions for Greensboro, North Carolina, USA. Predictors were assembled for 475 geocoded households spanning 34 census tracts and aligned across heterogeneous spatial supports. Models were fitted using household-resolution predictors, with predictions aggregated to the census-tract level for evaluation. Five-fold cross-validation used coordinate-based spatial blocks, which reduced geographic proximity between training and testing observations but did not guarantee complete separation of census tracts. In the primary exploratory analysis, Gradient Boosting produced the lowest mean prediction error (root mean squared error, RMSE = 4.57 ± 1.53; R2 = 0.453 ± 0.235), although performance was modest and variable across folds; these values are primary exploratory estimates rather than validated generalization performance. A fully tract-aggregated analysis provided a complementary evaluation using 34 independent observations and indicated that tree-based models remained comparatively strong, although model rankings were not identical. Contextual demographic, housing, and spatial-gradient variables contributed more strongly than measured soil metals; because the PLACES outcome is itself modeled from demographic and socioeconomic structure, these leading predictors are partly circular with the outcome and are best read as markers of structural context. These findings demonstrate the feasibility of the data-integration framework but should not be interpreted as definitive estimates of predictive accuracy or causal environmental effects.