A multiscale geospatial analysis integrating explainable machine learning to characterize spatial disparities in melanoma mortality associated with environmental justice and healthcare access in the contiguous United States

This is the authors' abstract. We don't add key points for this paper.

Background Melanoma mortality exhibits profound geographical disparities associated with complex sociodemographic, environmental, and healthcare factors. Traditional spatial models may not fully capture the multi-scale heterogeneity underlying these geographic disparities. This study aimed to characterize the spatial heterogeneity and operational scales of factors associated with county-level melanoma mortality across the United States. Methods We conducted a county-level ecological study across the contiguous US (2018–2024). To address mortality data suppression by the Centers for Disease Control and Prevention (CDC) (deaths < 10), a representativeness analysis was performed. Predefined core covariates (median age, UV exposure, dermatologist density) were integrated with 45 Environmental Justice Index (EJI) indicators. Feature selection was performed using an XGBoost-SHAP pipeline followed by stepwise AICc optimization. Multiscale Geographically Weighted Regression (MGWR) was employed to evaluate spatial non-stationarity and operational scales, comparing its performance against ordinary least squares (OLS) and standard geographically weighted regression (GWR) models. Results The analytical sample comprised 1,156 counties, representing 88.6% of the total US population. MGWR demonstrated superior explanatory power ( R 2 = 0.666, AICc = 2247.84) and substantially reduced residual spatial autocorrelation. The XGBoost-SHAP framework identified eight core predictors. MGWR revealed substantial variation in spatial scales: baseline spatial variation operated locally, while median age demonstrated a regional-scale positive association, with coefficients varying across geographic regions. High-volume road proximity demonstrated regional inverse associations. Dermatologist density, ozone exposure, and tornado frequency operated at global scales; dermatologist density and ozone exposure showed widespread inverse associations (potentially reflecting broader metropolitan healthcare and socioeconomic patterns), whereas tornado frequency was positively associated with melanoma mortality. Notably, UV exposure, hurricane frequency, and mobile home proportions exhibited global-scale coefficient patterns but lacked statistically significant local effects ( P ≥ 0.05). Conclusions Melanoma mortality exhibits multiscale spatial patterns. The limited association observed for environmental UV exposure suggests that demographic structure and healthcare-related factors may play important roles in explaining county-level mortality disparities. These findings support spatially targeted public health strategies, including improved dermatological care access in aging and disaster-prone regions.