Environmental predictors and machine learning models for climate-health risk prediction from extreme weather events: a global systematic review

Climate change is increasing the frequency, severity, and duration of extreme weather events, including heatwaves, floods, and heavy rainfall, posing substantial risks to population health worldwide. Traditional epidemiological approaches capture retrospective associations but are limited in modelling the non-linear, multivariable relationships between climate exposures and health outcomes. This review synthesised evidence on the application of machine learning models to predict health outcomes associated with extreme weather events and identify the environmental predictors most consistently reported as important. A comprehensive search across 11 electronic databases was conducted following the PRISMA-2020 guidelines, identifying peer-reviewed studies published between July 2010 and July 2025. Eighteen studies met the inclusion criteria and were assessed for risk of bias using the Prediction Model Risk of Bias Assessment Tool. Heat-related exposures dominated the included studies, with 17 of the 18 studies focusing on heatwaves, while only one examined heavy rainfall in relation to dengue outcomes. Temperature-based variables were consistently used as predictors across all studies. Socioeconomic and demographic variables were included in 10 studies, but were ranked as the most influential predictors in only three. Random Forest was the most frequently evaluated algorithm and was identified as the best-performing model in seven studies spanning high-income and upper-middle-income settings. However, the Generalised Additive Model outperformed machine learning models in two studies, and predictive performance varied substantially across contexts. Although the evidence base is concentrated in high-income countries, external validation was absent across all included studies, limiting the confidence in model generalisability. Despite the dominance of heat-related studies, the absence of models for flood-related health outcomes represents a critical gap, given the global burden of flood-related morbidity. Key limitations include limited integration of socioeconomic determinants, uneven geographical distribution of evidence, and sparse data availability with coarse spatial resolution in low- and middle-income countries, constraining model transferability. Significant variability in predictive performance, combined with the absence of external validation, indicates that the operational deployment of these models remains premature. Addressing these gaps requires expanding the modelling approach to include rainfall and flood-related health outcomes, improving data availability and spatial resolution, and developing transferable model architectures to support reliable early warning systems.