This study proposes a novel heatstroke risk prediction model that explicitly incorporates the air–sea temperature difference (ΔT) to account for coastal cooling effects induced by sea breezes. Conventional heatstroke indices, such as the Wet Bulb Globe Temperature (WBGT), rely primarily on terrestrial meteorological conditions and may overlook the influence of marine thermal environments. To address this limitation, 10-min interval meteorological data from the Japan Meteorological Agency’s AMeDAS station in Tsu were integrated with real-time sea surface temperature data obtained from a National Buoy in Ise Bay, resulting in a comprehensive dataset of 1,101 samples. A Random Forest classifier implemented from scratch using NumPy was employed to predict five distinct heatstroke risk levels.
The model demonstrated high generalization performance, with a 5-fold cross-validation accuracy of 98.0 ± 0.6% and a macro F1-score of 0.956. Permutation feature importance analysis under a WBGT-excluded condition indicated that ΔT contributed 6.9% to predictive performance, ranking as the third most important feature after air temperature and apparent temperature. Threshold analysis further identified ΔT ≥ 4 °C as a practical warning threshold, associated with an 84% probability of high-risk conditions. This threshold was empirically validated using real-world meteorological data from June 22, 2026. These findings suggest that incorporating marine environmental parameters can improve localized heatstroke risk assessment in coastal areas and provide a scientific basis for automated, site-specific early warning systems.