Housing, Neighborhood Context, and Soil Metals as Predictors of Model-Based Small-Area Diabetes Prevalence in a Medium-Sized City: An Explainable Spatial Machine-Learning Analysis

Diabetes prevalence varies across urban neighborhoods, yet the relative predictive contributions of environmental contamination, housing conditions, and geographic accessibility remain incompletely characterized. We applied an explainable spatial machine-learning framework to census-tract variation in model-based small-area diabetes prevalence across Greensboro, North Carolina. Environmental and neighborhood variables from 475 sampled residential locations were aggregated across 37 census tracts; tract prevalence estimates and 95% confidence limits came from the CDC PLACES 2024 release, so the findings describe relationships with modeled estimates rather than with clinically ascertained diabetes. Eighteen predictors, selected on conceptual grounds before model fitting, were evaluated using Random Forest under five-block spatial cross-validation. The primary model achieved an out-of-fold R2 of 0.275 (RMSE 4.213; MAE 3.314 percentage points), leaving approximately 72.5% of between-tract variation unexplained; the median R2 across 50 alternative spatial partitions was 0.224. Held-out SHAP analysis attributed 51.12% of summed absolute contribution to environmental metals, 21.78% to housing, 17.69% to geographic accessibility, and 9.40% to soil characteristics. Because the metal domain contained ten predictors against two housing and two accessibility predictors, attribution per predictor was highest for housing (40.04%) and geographic accessibility (32.52%). Individual predictor rankings were unstable across partitions and are not interpreted, and median soil cadmium and lead did not differ between High–High prevalence clusters and the remaining tracts. A 999-permutation test indicated performance beyond the permutation null (p = 0.001). These ecological associations are not evidence of individual-level exposure effects or causality.