Meteorological vs. Search-based Inputs for Heat-related Illness Forecasting

Objectives: Forecasting daily heat-related illness is often framed as a modeling problem, yet the practical behavior of different input structures under future conditions remains underexamined. Methods: This analysis compared two alternative configurations: one based on physical heat exposure measures and another that additionally incorporated a search-based index reflecting public attention. Instead of focusing solely on in-sample fit, the models were evaluated across sequential validation windows and a fully separated 2025 period. Results: During rolling validation, performance differences were small. In the independent future window, model responses diverged, with the linear specification remaining comparatively stable while the non-linear approach showed greater fluctuation. The search-derived signal did not replace established environmental indicators; however, when combined with a regularized linear model, it may function as a context-dependent auxiliary signal rather than a primary exposure proxy. Conclusions: Taken together, the results suggest that even a parsimonious input structure paired with a linear model can support practical short-term forecasting, offering a preliminary empirical basis for the design of low-cost early-warning systems in public health.