Comparing machine learning and statistical models for quantification of heat-attributable mortality across Europe

Extreme heat poses a major and growing risk to human health, yet accurately predicting its impacts on mortality remains challenging. In this study, we compared established nonlinear statistical models - including the epidemiological standard distributed lag non-linear model (DLNM) - with machine learning (ML) approaches both for predicting excess mortality and quantifying the heat-attributable mortality across Europe. We evaluated random forest regressions (RFs) and neural networks (NNs) trained on pooled European data, contrasting their performance with two-stage DLNMs and locally fitted generalized additive models. In each model, we included the lagged effect of temperature and additionally explored the inclusion of multiple environmental exposure variables (such as air pollution and humidity).

We assessed each model’s out-of-sample skill for predicting excess mortality, with our preliminary findings suggesting that the ML frameworks tend to improve skill across Europe. Notably, we find evidence that pooled ML models improve predictive performance for countries with fewer observations, suggesting that they are better able to learn from shared, diverse regional information. We also compared the spatial patterns and magnitudes of heat-attributable mortality estimated by the ML models with those from the DLNM, providing a benchmark. Together, our findings highlight the potential for ML-based frameworks to inform future heat-health impact assessments.