📊 epidemiology

Multi-model forecasting of respiratory disease activity in Germany during the 2024-2025 season

This paper presents and evaluates a collaborative multi-model forecasting system for respiratory diseases in Germany during the 2024–2025 season, demonstrating that while ensemble nowcasts and forecasts were generally well-calibrated and outperformed benchmarks, accurately predicting peak timing and magnitude remained challenging.

Original authors: Bracher, J., Wolffram, D., Amaral Lind, R., Bardeck, N., Boehm, M., Contreras, S., Doenges, P., Guenther, F., Kaiser, R., van de Kassteele, J., Kuhlmann, A., Lange, B., Nemcova, B., Priesemann, V., Re

Published 2026-11-03 (v2)

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Bracher, J., Wolffram, D., Amaral Lind, R., Bardeck, N., Boehm, M., Contreras, S., Doenges, P., Guenther, F., Kaiser, R., van de Kassteele, J., Kuhlmann, A., Lange, B., Nemcova, B., Priesemann, V., Reinacher, U., Rodiah, I., Sandmann, F., the RESPINOW Study Group,, Schienle, M.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Every autumn and winter, a familiar cycle returns: respiratory illnesses sweep through communities, filling clinics and hospitals with people struggling to breathe. For public health officials, the challenge is not just treating the sick, but understanding the rhythm of these outbreaks while they are happening. They need to know how many people will be affected next week, when the peak of the season will arrive, and how severe it will be. To answer these questions, scientists use mathematical models—tools that take current data about infections and project them forward in time. These projections help hospitals prepare beds, guide vaccination campaigns, and inform the public. However, predicting the future of a virus is notoriously difficult, especially when the data itself is messy, arriving late or changing as more information is gathered.

In the 2024–2025 season, a team of researchers in Germany set out to test how well a group of different forecasting models could work together to track these respiratory waves. They focused on two main types of data: visits to general practitioners for acute respiratory infections, which represent the broad, everyday sickness, and hospitalizations for severe cases, which represent the most critical burden on the healthcare system. They also tracked confirmed cases of seasonal influenza and respiratory syncytial virus. The team operated a collaborative platform where five independent research groups submitted their predictions every week. The goal was to see if combining these different models into a single, collective forecast would provide a clearer picture than any single model could on its own, and to see how well they could handle the inevitable delays and corrections in the data.

The researchers found that the system worked well for catching up on the present. When the data for a specific week was incomplete because reports were still arriving, the models successfully estimated the true numbers, a process known as nowcasting. These estimates were generally reliable, though the models stumbled slightly during the Christmas and New Year holidays. During this break, the usual patterns of reporting changed, causing some models to temporarily overestimate the number of infections in the weeks immediately following the holidays. Once the teams adjusted for this seasonal disruption, the estimates returned to normal accuracy.

When it came to looking ahead, the models performed better than simple guesses based on the previous week's numbers, but they faced significant challenges in predicting the exact timing and height of the peak. Most models correctly identified that the season would rise and fall, but they consistently predicted that the peak would happen too early and would not be as high as it actually turned out to be. For instance, while the severe hospitalization cases peaked in mid-February, many models suggested the turning point would occur in late January. This tendency to see the curve flatten out sooner than reality meant that the forecasts were often too cautious about the severity of the upcoming weeks.

Despite these difficulties with the peak, the combined forecast from all the models proved to be one of the strongest approaches available. It did not consistently beat every single individual model, but it offered a robust and well-calibrated view of the situation. The study also highlighted the value of looking at different age groups separately. Because respiratory waves often hit younger and older populations at different times, breaking the data down by age allowed the models to spot patterns that were invisible when looking at the total numbers alone. This stratified approach revealed statistically significant improvements in accuracy that were missed when only looking at the country as a whole.

The project also served as a test of how scientific collaboration functions after the intense pressure of the pandemic. The team discovered that keeping a diverse group of modelers engaged and submitting data every week was harder than expected, with some groups unable to provide predictions in real time. To ensure a complete evaluation, the researchers had to fill in missing weeks later using the data available at the time, a method that introduced a small risk of bias but allowed for a full picture of the season. Ultimately, the study confirmed that while predicting the exact height and timing of a respiratory wave remains a difficult task, a network of different models working together provides a valuable, reliable tool for public health monitoring. The experience showed that with better handling of holiday reporting quirks and continued refinement of age-specific data, these forecasting hubs can become even more effective guides for navigating the winter season.

Technical Summary: Multi-model Forecasting of Respiratory Disease Activity in Germany during the 2024–2025 Season

Problem Statement
Respiratory diseases constitute a major source of morbidity and mortality, necessitating robust public health monitoring. While forecasting systems were extensively developed during the COVID-19 pandemic, the post-pandemic landscape presents distinct challenges: overlapping seasonal epidemics of multiple pathogens, altered surveillance practices, and the need to transition from emergency response to routine seasonal monitoring. The specific challenge addressed in this study is the real-time nowcasting (estimating current incomplete data) and short-term forecasting (projecting near-future activity) of respiratory disease activity in Germany for the 2024/25 season. The study focuses on four key indicators: general practitioner consultations for acute respiratory infections (ARI), hospitalizations for severe acute respiratory infections (SARI), and laboratory-confirmed cases of seasonal influenza and Respiratory Syncytial Virus (RSV). A critical technical hurdle is that all indicators are subject to retrospective data revisions, requiring models to integrate nowcasting steps to correct for reporting delays before generating forecasts.

Methodology
The study utilizes the RESPINOW Hub, a collaborative forecasting platform that aggregates predictions from multiple independent research groups. The operational period spanned from October 17, 2024, to March 24, 2025.

Data Sources: The analysis relies on four primary data streams:
1. ARI: Syndromic surveillance from the ARE-Praxis-Sentinel (approx. 500 GPs), reported weekly to the Robert Koch Institute (RKI).
2. SARI: Sentinel hospitalization data from the ICOSARI system (approx. 70 hospitals).
3. Influenza & RSV: Laboratory-confirmed cases from the SurvStat system (notifiable diseases).
- Data were stratified by age groups (0–4, 5–14, 15–34, 35–59, 60+ for ARI; additional splits for 60–79 and 80+ for SARI).
- A key feature of the methodology is the handling of data revisions: SARI data undergoes almost exclusively upward revisions, while ARI revisions are symmetric and driven by weighted averages of sentinel sites.
Models: A total of nine models were contributed to the Hub, but the evaluation focused on specific configurations per indicator:
1. Nowcasting Models: KIT-simple nowcast, KIT-epinowcast, RKI-Pilot 01, and RIVM-GAM. These correct for reporting delays and are applied to horizons $h = -3$ to $0$.
2. Mechanistic Models: HZI-ODEmodel, a compartmental (SIR-type) model applied to pathogen-specific (Influenza/RSV) targets.
3. Data-Driven Statistical Models: KIT-hhh4 (endemic-epidemic time series), KIT-LightGBM (gradient boosting), KIT-TSMixer (neural network for time series), and MPIDS-PS embedding (phase-space matching). These were applied to syndromic (ARI/SARI) targets.
- Note: For the main ARI and SARI indicators, a total of eight models were run to generate nowcasts and forecasts. For Influenza and RSV, only one forecasting model (HZI-ODEmodel) was available for each. Most forecasting models utilized the KIT-simple nowcast as an input to handle reporting delays.
Ensembles and Benchmarks:
- Ensembles: Two ensembles were constructed: KIT-ensemble complete (average of all available models) and KIT-ensemble realtime (limited to models available in real-time). A European RespiCast-Ensemble was also included for ARI comparison.
- Baselines: A Persistence model (flat-line with uncertainty derived from past errors) and a Historical model (seasonal averages from previous years).
Evaluation Metrics:
- Primary Metric: Weighted Interval Score (WIS), a quantile-based probabilistic score.
- Secondary Metrics: Absolute error for point forecasts, coverage proportions for 50% and 95% prediction intervals, and WIS applied to log-transformed incidence to reduce scale dependence.
- Statistical Testing: Permutation tests were used to assess the significance of performance differences against the Persistence baseline.

Key Results

Nowcasting Performance: Nowcasting models generally showed reliable performance. However, several models (specifically KIT-simple nowcast and KIT-epinowcast) exhibited an upward bias in early January due to an inability to adjust for the extended reporting delays during the Christmas/New Year break. This was corrected in subsequent weeks for the simple nowcast model.
Forecasting Performance:
- Calibration: Forecasts were generally well-calibrated, though prediction intervals were often wide.
- Peak Prediction: A significant challenge was the anticipation of peak timing and magnitude. Most models predicted the peak too early (e.g., late January instead of mid-February for SARI) and underestimated the peak magnitude, resulting in "too flat" curves.
- Benchmark Comparison: Most models outperformed simple benchmark models (Persistence and Historical) overall. However, for age-stratified targets, performance differences were statistically significant, whereas for unstratified population-level targets, statistical power was often insufficient to declare significance.
- Ensemble vs. Individual Models: Unlike previous pandemic-era projects, the combined ensemble did not consistently outperform the best individual models. For ARI, LightGBM and TSMixer performed comparably to the ensemble. For SARI, the ensemble performed well but did not strictly dominate.
- Horizon Specificity: At a 1-week horizon, most models performed similarly to the Persistence baseline. At a 4-week horizon, models struggled to outperform the Historical baseline. Skill was primarily observed at intermediate horizons (2–3 weeks).
- Christmas Effect: The failure to explicitly model the Christmas break dip in ARI data led to overprediction in early January. A retrospective modification to the hhh4 model (adding a Christmas indicator) significantly improved its performance, suggesting this feature is critical for future iterations.

Significance and Contributions
The paper presents the first operational season of the RESPINOW Hub, serving as a bridge between pandemic-era emergency forecasting and routine seasonal surveillance. Its primary contributions include:

Operational Validation: It demonstrates the feasibility of a multi-model, multi-target forecasting system for respiratory diseases in a post-pandemic context, integrating both syndromic and pathogen-specific indicators.
Handling Data Revisions: The study highlights the necessity of coupling nowcasting and forecasting steps when dealing with retrospective data revisions, a specific challenge for German surveillance systems.
Performance Insights: It provides empirical evidence that while ensemble forecasts are robust, they do not automatically guarantee superiority over individual models in seasonal contexts, particularly when individual models (like LightGBM) capture specific seasonal features effectively.
Community and Organization: The paper documents the organizational challenges of maintaining collaborative forecasting efforts in "times of peace," noting that model participation was moderate and required retrospective completion to ensure a comprehensive evaluation. It emphasizes the value of the Hub in strengthening links between the RKI and academic institutions.
Lessons Learned: The study identifies specific technical improvements, such as the need to explicitly model holiday effects (Christmas break) and the potential benefits of using pathogen-specific stratified indicators to improve peak prediction accuracy.

The authors conclude that while the system is functional and generally outperforms simple baselines, predicting the exact timing and magnitude of seasonal peaks remains a difficult task. The work underscores the importance of continuous model refinement and the integration of specific seasonal features to enhance real-time situational awareness.

More like this