The study will draw on 4 data sources: routinely collected ED data from the AKTIN ED Registry, additional information from a regional dataset of routine data drawn from the ED Klinikum Stuttgart, survey data from geriatric patients, and environmental exposure data provided by the German Meteorological Service (Deutscher Wetterdienst [DWD]) and the Federal Environment Agency (Umweltbundesamt [UBA]).

Retrospective data will be sourced from the AKTIN Registry [27]. The registry is implemented on the underlying AKTIN infrastructure, a nationwide federated data network operated within the Network University Medicine [28]. For the study, a total of 56 EDs throughout Germany, including clinics of different care levels in cities with different population sizes, provide electronic health records from all ED visits documented in the respective ED in the years 2019 to 2024.

Across participating EDs, approximately 6.35 million ED visits will be available for the study period, with an anticipated exclusion of up to 10% due to missing key variables. This retrospective ED dataset contains patient-specific information on age, sex (male, female, or other), and the zip code of the patient’s place of residence, as well as administrative data on mode of transport to the ED, time of admission, and mode of disposition (discharge to home, transfer to other hospital, inpatient admission, or death). Furthermore, it contains the following treatment information: ED diagnoses in the form of ICD-10-GM (International Classification of Diseases and Related Health Problems, Tenth Revision, German Modification) codes with a labeling of the main ED diagnosis, additional information on diagnosis certainty (secured, suspicion of, condition after or exclusion of diagnose), chief complaint in the form of Canadian Emergency Department Information System – Presenting Chief Complaints (CEDIS-PCL) codes, urgency level coded as Manchester Triage System (MTS) or Emergency Severity Index (ESI), vital parameters, Glasgow Coma Scale (GCS) and isolation status [29-31]. In case of an inpatient admission following the ED visit, inpatient treatment information according to the German social laws regulating the billing of inpatient services (§ 21 Krankenhausentgeltgesetz, KHentG) is available. This information contains the inpatient discharge diagnosis, also coded as ICD-10-GM. Multiple ED visits by the same person during the study period cannot be attributed due to data protection requirements.

All ED data will be sourced via established AKTIN registry processes in a standardized format and provided to the study team as a consolidated study dataset for analysis [28]. Before any analyses are conducted, we will perform systematic data quality checks, including verification of variable formats (dates, times, and ICD-10-GM codes), plausibility checks for key variables (eg, age range, sex, and time stamps), and identification of potential duplicate records based on site, admission date and time, and local encounter identifiers. Obvious duplicate entries and technically invalid records will be removed in accordance with predefined rules.

For the analysis of basin-valley differences in Stuttgart, additional regional information beyond the regular dataset is available for the period from 2011 to 2023. This dataset comprises approximately 409,000 cases and enables longer-term analyses. Compared to the AKTIN ED Registry, the dataset provides more detailed information on emergency patients and their treatment. In addition to diagnoses and symptoms, treatment times, duration, diagnostics, place of residence, and physician requirements are recorded. This allows for an analysis of not only the emergency presentation but also the diagnostic and therapeutic effort involved.

Additional prospective data on the individual functional level of old patients will be collected at 3 participating study centers (Department of Geriatric Medicine Aachen, Hospital Fürth, and University Clinic for Geriatric Medicine Oldenburg). All geriatric patients admitted to the geriatric ward via the participating EDs at the 3 study centers will be screened and invited to participate in the prospective substudy. Written informed consent will be obtained during inpatient treatment by trained study personnel at the participating centers. In addition to routinely performed comprehensive geriatric assessment, the substudy will collect standardized patient-reported and functional outcomes, including the Parker Mobility Score (mobility) [32], the SF-12 Health Survey (health-related quality of life) [33], and parts of the questionnaire on social situation [34]. A project-developed questionnaire will assess self-reported heat exposure and heat-related problems in the days preceding hospital admission. A follow-up survey approximately 6 weeks after discharge will assess longer-term outcomes, such as reduced mobility and increased level of dependency (eg, discharge to a nursing home). In addition, inpatient data will be used to describe specific characteristics of the geriatric cohort using the Hospital Frailty Risk Score [35] and investigate outcomes as length of stay and in-hospital mortality. To maximize recruitment and follow-up adherence in this vulnerable cohort, only geriatric centers with established research experience are included. Study personnel will be dedicated to the substudy, experienced in geriatric care, and located directly on the wards to enable repeated opportunities for study information, support with questionnaires, and communication with relatives or legal representatives when required. Contact information for follow-up will be updated by the day of discharge, particularly in case of changes in the discharge destination (eg, nursing home). Follow-up will be conducted within a 2-week contact window around the 6-week time point, with repeated contact attempts if needed. Based on treatment capacities and prior patient volumes at the participating centers, we estimate that approximately 281 patients will meet the inclusion criteria during the planned recruitment period. Based on experience from previous geriatric studies, we expect that approximately 50% will consent to participate (approximately 139 participants). For the 6-week follow-up, a dropout rate of approximately 50% is anticipated.

The retrospective weather and environmental data are collected from freely accessible datasets provided by the German Meteorological Service (DWD) and the Federal Environment Agency (UBA). The DWD provides data sets comprising hourly and daily aggregated data on air temperature, humidity, wind speed, and precipitation. The Federal Environment Agency provides datasets on air quality at an hourly level on ozone (O₃), nitrogen dioxide (NO₂), particulate matter (PM10 [particulate matter with a diameter ≤10 µm] and PM2.5 [particulate matter with a diameter ≤2.5 µm]), and sulfur dioxide (SO₂). For the regular use case Stuttgart, data on global radiation, radiation balance, as well as UV-A and UV-B radiation will also be available. The weather and environmental data of Stuttgart will be sourced from an environmental monitoring station located in the valley basin, approximately 1.3 km from the clinic. In addition, pollen count data will be acquired from the German Pollen Information Service foundation (Polleninformationsdienst [PID]). The PID operates various pollen measuring stations in Germany, where the pollen concentration of various allergenic plants and trees is measured and counted on a daily basis. The pollen types with the highest allergic potential for humans will be selected together with experts from the PID.

Area-level socioeconomic deprivation will be operationalized using the German Index of Socioeconomic Deprivation (GISD), a composite measure developed by the Robert Koch Institute that summarizes relative regional deprivation across education, employment, and income dimensions and is available at multiple spatial levels (eg, municipality and district) for different years. Patients’ residential postal codes will be mapped to GISD values and used to assess socioeconomic gradients and potential effect modification of heat–outcome associations.

We will use the proportion of ED encounters resulting in inpatient admission per hospital and day (hospital-day level; derived from encounter disposition and admission fields using AKTIN encounter identifier aggregated by clinic×date).

We will quantify daily counts and proportions of prespecified diagnostic and syndromic categories, including heat-sensitive conditions and shifts in major diagnosis group distributions (hospital-day level; derived from ED ICD-10-GM, inpatient discharge ICD-10-GM, and CEDIS-PCL).

We will analyze the daily distribution of referral sources and mode of transport to the ED (hospital-day level; derived from encounter-level pathway variables such as transport and referral).

We will use daily distribution of acuity and clinical severity indicators at presentation (hospital-day level; derived from triage categories, high-acuity proportions, and additional severity proxies such as vital signs or Glasgow Coma Scale).

We will characterize patient-reported heat-related symptoms and functional impact using structured questionnaires at prespecified assessment time points (participant-level, derived from questionnaires on quality of life, mobility, social background, and heat-related behavior).

This study was approved by the leading ethics committee at the Medical Faculty of the University of Magdeburg (29/24 dated March 08, 2024), as well as by the institutional ethics committee at RWTH Aachen (24-333 dated October 10, 2024).

The use of the ED data has been approved by the AKTIN Data Use and Access Committee and is subject to AKTIN’s data protection regulations. Written informed consent will be obtained from all patients participating in the ancillary prospective geriatric substudy. Both the retrospective registry-based study and the ancillary prospective geriatric substudy are registered in the DRKS and are governed by their respective approvals (AKTIN data access for retrospective analyses; site-specific ethics and informed consent for the prospective substudy).

Data are processed in compliance with applicable data protection regulations. The dataset used for analysis is pseudonymized at source and direct identifiers are not available to the study team. Data transfer and storage are secured (encrypted transfer, access restricted to authorized personnel, analysis on secure servers), and results are reported only in aggregated form.

Geriatric participants will receive no compensation.

The descriptive analysis of the overall population will characterize the overall cohort and support the planning of relevant subgroup analyses. It will include key aspects of the cohort based on multiple variables, including sociodemographic factors (age and gender), presenting complaints (CEDIS-PCL), initial triage assessment (MTS or ESI), patient referral pathways, hospital stay duration, discharge outcomes, diagnostic categories (ICD-10-GM), and socioeconomic status (GISD). Continuous variables will be summarized using mean (SDs) or median (IQRs), and categorical variables as counts and percentages. Distributions will be presented graphically (eg, histograms and density plots).

The number of all-cause ED visits in relation to temperature and environmental factors will be investigated using a 2-stage analysis approach. In the first stage, given the available data, we plan to fit the site-specific number of daily ED visits using a quasi-Poisson regression model including a distributed lag nonlinear model for daily temperature to capture delayed temperature effects, using appropriate splines and lags [34].

These models aim to adjust for seasonal trends in temperature and air quality using smooth functions and will additionally include day of week, public holidays, and further covariates (O₃, NO₂, particulate matter [PM10 and PM2.5], SO₂, UV-A, UV-B radiation, and predefined pollen). Heat days, one of the main focuses of this study, will be defined a priori based on clinic-specific temperature distributions (days with mean air temperature above the 95th percentile), and alternative operationalizations (eg, heat index, tropical nights, and heatwaves) will be examined in prespecified sensitivity analyses. The analysis is planned to be stratified by age group (0-5 years, 6-11 years, 12-17 years, and ≥65 years). In the second stage, ED-specific distributed lag nonlinear model coefficient vectors are planned to be pooled using a random-effects meta-analysis approach to obtain an overall heat exposure–response association and to quantify between-site heterogeneity [34]. As sensitivity analysis, we will assess the robustness of our modeling choices, including alternative lag windows, splines, and alternative number of knots for temperature and lag. We will also investigate the sensitivity when using other alternative stratification strategies and missing data handling. A detailed statistical analysis plan will be prepared in advance of the data analysis.

Analyses of the prospective geriatric substudy will be conducted separately from the retrospective registry-based analyses and follow prespecified endpoints and analytical procedures as defined in the substudy protocol and DRKS registration. We will summarize baseline characteristics and patient-reported and functional outcomes using appropriate descriptive statistics. Associations between heat exposure measures and substudy outcomes will be evaluated using regression models suited to the outcome scale (eg, linear, logistic, or ordinal models, as applicable), with effect estimates reported alongside 95% CIs. Models will adjust for key prespecified covariates available in the substudy dataset (eg, age, sex, and frailty-related measures), and sensitivity analyses will address missing data and loss to follow-up where appropriate.

To create the syndrome definition, a literature search on ED visits during heat events will be conducted to identify relevant ICD-10-GM and CEDIS-PCL codes. The feasibility and validity of the created syndrome definitions will then be assessed using the above-described retrospective ED data linked with environmental data. We will perform descriptive analysis of the identified cases of the health indicator of interest (eg, heat-related ED visits) using time series and exposure-outcome plots to visualize a potential association between the frequency of codes and temperature. Logistic regression analyses will be used to quantify these associations and account for uncertainty. The results will be discussed with clinical experts, and the syndrome definition will be adjusted in an iterative process until one indicator is finalized. To validate the indicator, time series of the syndrome definition will be compared with temperature time series, and mixed model logistic regression analyses will be conducted again to quantify the associations and uncertainty between trends of identified cases and changes in temperature. Furthermore, we will be looking into other external data sources, such as heat mortality and heat warnings from the DWD, to assess if we can perform additional validation analysis by comparing them with the heat indicator created in this study.

The objective is to develop mathematical models using machine learning methods to enable simulation-based analyses, such as weather-dependent predictions. These models will vary in granularity and capture correlations between environmental data and disease-specific demographic factors, such as age and sex. Feature selection will be tailored to different disease types, incorporating demographic characteristics and clinical patient information. Based on this analysis, customized data-driven machine learning models will be developed. Various model types, including support vector machines, feedforward neural networks, and recurrent neural networks, will be explored. The proposed architectures were selected based on their strengths for the prediction task at hand: support vector machines provide robust performance in high-dimensional feature spaces with strong regularization properties; Feedforward neural networks offer flexibility in modeling complex nonlinear relationships between environmental and demographic variables; and recurrent neural networks are specifically suited for capturing temporal dependencies and lagged effects inherent in time series data such as seasonal ED usage patterns. These models will be trained and cross-validated using supervised learning on ED data. A literature review will be conducted to identify models that integrate expert knowledge, with a particular focus on time series models incorporating climate data for forecasting ED usage. An initial approach may involve time series regression with climate variables, allowing for the prediction of seasonal trends. Additionally, daily aggregated environmental data (see section Data Linkage) will be incorporated into the modeling process. This dual approach mitigates project risks by combining machine learning as a black-box model with an interpretable time series approach. The integration of both methodologies will enhance insights, particularly in feature selection. To ensure reproducibility, all source code will be made publicly available via the project’s GitHub repository [35].