Spatial Heterogeneity in Cancer Incidence: Assessing Behavioral and Environmental Influences Using Machine Learning and Multiscale Geographically Weighted Regression

Cancer incidence exhibits substantial spatial disparities linked to environmental, behavioral, built-environment, healthcare-access, and socioeconomic conditions, yet the spatial scales at which these relationships operate remain insufficiently understood. This study develops an explainable spatial epidemiology workflow that integrates Random Forest, SHapley Additive exPlanations (SHAP), Ordinary Least Squares (OLS), Geographically Weighted Regression (GWR), and Multiscale Geographically Weighted Regression (MGWR) to examine county-level incidence for all-site, colon, breast, and skin cancers across Texas, USA. Random Forest and SHAP were used to identify outcome-specific nonlinear predictor relevance, and OLS, GWR, and MGWR were used to compare global, local, and multiscale spatial associations before and after RF-SHAP feature screening. MGWR generally achieved higher model fit than OLS and GWR. Before RF-SHAP screening, MGWR R² values were 0.701 for all-site cancer, 0.516 for colon cancer, 0.499 for breast cancer, and 0.694 for skin cancer, compared with OLS R² values of 0.343, 0.338, 0.257, and 0.368. RF-SHAP reduced predictors by about one-half and consistently improved AICc. The results show that environmental exposures, activity-related conditions, transportation access, screening, food insecurity, and chronic health indicators contribute to spatial differences in cancer incidence. The framework links nonlinear machine-learning evidence with spatially explicit interpretation for transferable epidemiological analysis.