Epidemiologically validated spatial modelling reveals fine-scale heat health risks

Abstract

Heat exposure drives substantial global disease and mortality burdens, making spatially explicit heat health risk assessment essential for targeted adaptation. Existing studies apply diverse modeling approaches yet rarely evaluate them systematically because reliable epidemiological validation data remain scarce; conventional risk indices further depend on subjective indicator aggregation and expert weighting. Here we show an integrative geographically neural network weighted regression (GNNWR) framework validated against epidemiologically derived heat-attributable mortality fraction at the subdistrict level. The framework achieves stronger predictive performance and better generalization than ordinary least squares, multiplicative model, random forest, geographically weighted regression (GWR), multi-scale GWR, and other geographically weighted machine learning models, with consistent alignment between training and test sets and no evidence of overfitting. By learning spatial weights through a neural network, it simultaneously captures spatial non-stationarity and nonlinear relationships among meteorological, air pollution, demographic, socioeconomic, and built environment variables without requiring predefined index construction. It enables prediction in data-sparse areas and generates consistent high-resolution (100 m grid), subdistrict, and district-level risk maps, while identifying fine particulate matter (PM2.5) as the second strongest risk factor associated with heat-related mortality risk, after temperature. This validated, data-driven approach provides a transferable methodological foundation for identifying high-risk populations and informing precise public health interventions and urban climate adaptation strategies.