High-resolution modeling and projection of heat-related mortality in Germany under climate change

Climate change, driven by anthropogenic greenhouse gas emissions, is causing a global temperature increase. Europe is warming at almost twice the global rate, leading to more frequent, intense, and longer-lasting heatwaves. Heatwaves pose significant risks to human health, public health systems, food security, work productivity, and infrastructure. Higher temperatures are strongly linked to increased mortality rates, particularly for vulnerable populations like the elderly. The burden of heat-related mortality has risen globally, with Germany experiencing substantial heat-related deaths in recent years. Existing models for estimating heat-related mortality often suffer from limitations, such as treating daily effects as independent and mismatches in the temporal and spatial resolutions of temperature and mortality data. This study addresses these issues by developing a high-resolution multi-scale model to estimate heat-related mortality in Germany, enabling better localized risk assessment and future projections under various climate change scenarios.

The connection between temperature and mortality is often examined using generalized additive models (GAMs), distributed lag models (DLMs), and similar statistical models. However, these models have limitations. They often treat the effects of temperature on each day as independent, failing to capture the cumulative effects of consecutive hot days. Furthermore, the spatial and temporal resolutions of temperature and mortality data often don't align, leading to the use of aggregated temperature data which can reduce model precision. Most models also overlook the impact of regional temperature variations, such as the urban heat island effect. This study aims to overcome these limitations.

This study developed a multi-scale machine learning model to estimate daily heat-related mortality in Germany at the district level. The model integrates data with varying temporal (daily to weekly) and spatial (1x1 km grids to statewide) resolutions. Mortality data was obtained from the Federal Office of Statistics of Germany, including daily and weekly counts categorized by state, gender, and age group. Population data was obtained from the regional database of Germany. Temperature data was sourced from three different resources: Copernicus European Regional ReAnalysis (CERRA), Helmholtz Munich (1x1 km resolution), and Deutscher Wetterdienst (DWD) daily station data. An attention-like model was used to interpolate daily average temperature from DWD stations or climate projections to the district level. Climate projections were obtained from the EC-Earth3 model for three Shared Socioeconomic Pathways (SSPs): SSP126, SSP245, and SSP370. Two distinct neural network architectures were evaluated: a linear model and an exponential model. The models predict multiplying factors for a baseline mortality rate based on temperature input, accounting for lag effects using convolutional layers. The Poisson loss function was used for model training and evaluation. Multiple model instances were trained to reduce prediction variance, and the average output was used as the final result.

The model estimated approximately 48,000 heat-related deaths in Germany during 2014–2023, with most deaths concentrated during specific heatwaves. The analysis of the July 7–14, 2023 heatwave showed a total of 1100 estimated heat-related deaths (28% of the total for the year). The model revealed a lagged relationship between excess heat and daily mortality risk, persisting for up to two days after the temperature peak. Projections under different SSPs showed a significant increase in heat-related mortality by 2100: a 2.5-fold increase under SSP245 and a 9-fold increase under SSP370, assuming static sociodemographic conditions. Inclusion of non-static demographic projections showed a further increase in risk, up to 40-60% more heat-related deaths compared to static scenarios. The model accurately predicted daily mortality for warmer days (RMSE: 91.4 for training, 83.9 for validation; R²: 0.7887 for training, 0.8272 for validation) and showed good agreement with RKI reports when using similarly aggregated data. The model also demonstrated the importance of considering minimum daily temperatures in predicting heat-related mortality, suggesting that high nighttime temperatures are particularly dangerous.

The findings address the research question by providing high-resolution estimates of heat-related mortality in Germany and projecting future risks under climate change. The model's high spatial and temporal resolution allows for a more nuanced understanding of heat-related mortality, including the identification of local hotspots and the impact of heatwaves. The significant projected increases in heat-related mortality highlight the urgent need for adaptation strategies. The model's ability to integrate climate projections enables policymakers to assess potential future impacts and inform decision-making related to heat-health action plans. The model's good agreement with existing reports from RKI, while also highlighting differences due to data aggregation, validates its approach and demonstrates its value as a tool for improving risk assessment and mitigation efforts.

This study provides a novel multi-scale machine learning model for estimating and projecting heat-related mortality. The model's high resolution offers valuable insights into the spatial and temporal dynamics of heat-related deaths, highlighting the impact of heatwaves and the need for targeted interventions. The substantial projected increases in heat-related mortality under future climate change scenarios emphasize the urgent need for adaptation measures. Future research could focus on refining the model by incorporating socio-economic factors and further exploring the vulnerability of different demographic groups.

The model assumed uniform baseline mortality and temperature-mortality relationships across districts, which may not be entirely accurate. Obtaining precise temperature data for every district is challenging, and within-district variations in temperature are not fully accounted for. The climate downscaling model used in projections may introduce uncertainties. The projections assumed constant population and baseline mortality rates, neglecting potential improvements in healthcare or adaptation measures. The model primarily focused on temperature as a predictor, potentially overlooking other contributing factors to mortality.