High-Dimensional Heterogeneity in Heat-Related Health Impacts Across Germany Districts.

Abstract

BackgroundMultiple factors can drive heat susceptibility, but are typically analyzed individually. This study investigates high-dimensional heterogeneity across multiple domains in the effect of hot days on emergency hospitalizations across German districts. MethodsWe obtained emergency hospitalization rates per 1,000 in 3-day windows for the summers of 2005-2023 from the German Diagnosis-Related Groups-Statistic, linked with spatio-temporal temperature information. A hot day was defined as at least one day above the district-level 98th percentile heat index. Multiple characteristics of the demographic, health/social, and environmental domains were selected as potential effect modifiers. We implemented a generalized random forest, where outcome and propensity scores were modeled with fixed effects estimators to control for seasonality and district-level confounding. We divided the conditional average treatment effects (CATE) into quartiles to map their spatial distribution and identify key effect modifiers via the variable importance and observed indicator profiles across quartiles. ResultsWe found substantial variation across districts regarding heat-related emergency hospitalizations. The top three indicators for predicting this effect heterogeneity were the percentage of people seeking protection, population below age 6, and air pollution levels. The observed indicator profile for the most susceptible districts was characterized by a larger percentage of vulnerable populations, higher car density, and fewer hospital beds. ConclusionThere is substantial variation in heat susceptibility across and within districts in Germany. We propose an approach that can help understanding such complex interplay of indicators from different domains driving heterogeneity. Research in ContextO_ST_ABSEvidence before this studyC_ST_ABSWe screened PubMed and Web of Science for studies published from 2000 to March 2026 using search terms for "extreme temperature", "hospitalizations", and "susceptibility". Previous literature investigated the role of demographic, health and social, and environmental indicators as potential drivers of effect heterogeneity in heat-related hospitalizations. However most research focused on few effect modifiers analyzed separately, mainly driven by constraints of existing methods. To date, no study used data-driven approaches to account for the joint, intersectional structure of most indicators. Added value of this studyThis study drew on a comprehensive list of candidate indicators based on previous literature and a comprehensive enumeration of hospitalizations linked with high-dimensional temperature data and regional characteristics. We modeled these indicators of the demographic, health and social, and environmental domains jointly, rather than separately, using a replicable generalized random forest approach. This allowed us to identify the most important drivers of effect heterogeneity across a country while accounting for the joint, intersectional structure of the candidate indicators. We identified substantial within- and across-district heterogeneity, mainly driven by indicators of the demographic domain, followed by indicators of the environmental, and health and social domains. Implications of all the available evidenceWarning systems and adaptation measures for heat are frequently designed around generic categories of vulnerable groups or around a single dominant risk factor, which risks both overlooking places where several moderate disadvantages combine to produce high susceptibility and misdirecting resources toward areas identified by one indicator alone. Therefore, understanding heat susceptibility as the joint, intersectional product of multiple district characteristics instead of single indicators can directly inform heat action plans and heat-health warning systems.