Evaluating Effectiveness of Heat Alert Systems Combining Epidemiological and Machine Learning Techniques

Heat Early Warning Systems, designed to alert the public to forecasted high temperatures, have been implemented across most European countries within the last 20 years. However, evaluation of their effectiveness in protecting public health poses significant methodological challenges. Unlike property damage from floods or thunderstorms, which can be directly attributed to extreme weather events, the impact of hot weather on human health involves complex causal pathways. Statistical methods are required to account for confounding factors influencing outcomes such as mortality and cardiovascular illness. Additionally, frequent revisions and updates of heat alert thresholds hinder the assessment of the impact of these changes over time.

This study introduces a methodological framework to asses the effectiveness of heat alert days in reducing all-cause mortality. By examining both pre- and post-implementation periods of Heat Early Warning Systems, the approach accounts for temporal shifts in mortality patterns. A key challenge is identifying heat alert-eligible days prior to HEWS implementation, as alerts are issued based on weather forecasts rather than actual recorded temperatures. Simply applying alert criteria to temperature data fails to retrospectively recognize heat alert days: the alerts are triggered by weather forecast, which do not necessarily correspond to actual recorded temperatures. To address this issue, a random forest classifier was implemented to retrospectively identify days that would have triggered a heat alert. Finally, a time-series regression using distributed lag models was combined with a difference-in-differences approach to evaluate whether heat alerts were associated with reductions in all-cause mortality. This method might contribute to further improvement of efficient heat alert systems.