Demographically-Informed Heat-Mortality Risk Curves via Risk Graph Neural Networks

This paper proposes Risk Graph Neural Networks (RGNNs), a hierarchical graph-neural architecture that incorporates granular census features to optimize coefficient vectors in Distributed Lag Non-linear Models (DLNMs). The design aims to retain interpretable temperature-mortality risk curves while adding demographic and geographic context. According to the supplied abstract, experiments across 10 regions in England and Wales and two unprecedented heat years found lower point errors and near-nominal uncertainty coverage for RGNN variants. The reported advantage was especially visible during the 2022 heatwave, when baseline methods reportedly failed to maintain calibration.