Project overview

Short-term forecasts provide an important complementary perspective to scenario modeling. They contribute to situational awareness, i.e., the timely and reliable assessment of the current intensity and trend of an epidemic. In this way, they support decision-making about interventions and the use of resources. Because many epidemiological indicators are available only with a time lag due to reporting delays, a thorough assessment of the current situation also requires a correction of the most recent data, a task known as nowcasting, which is conceptually closely related to short-term forecasting. This subproject will expand existing infrastructures to provide an operational nowcasting and short-term forecasting system for respiratory syncytial virus (RSV), pneumococcal, seasonal influenza, and COVID-19.

Both case numbers and the resulting burden on the health care system are considered. Particular emphasis is placed on age-stratified analyses and appropriate quantification of forecast uncertainty. Building on our previous work, we will take a collaborative multi-model approach combining nowcasts and forecasts from different models. Such combined ensemble forecasts have proven to be more robust than single-model forecasts in many situations and facilitate communication with decision makers. All results are made available to the public in real time following an established open science concept. This approach also provides the framework for model evaluation and validation. Predictive performance will be assessed using modern evaluation methods in a prospective and pre-registered evaluation study.

Particular attention will be paid to which modeling strategies are most promising at which time horizons, the advantages of stratified forecasting or nowcasting, and how best to combine different forecasts into an ensemble. While a basic set of models will be contributed by the leadership team at the Karlsruhe Institute of Technology as well as by four partner groups within the work package (see section Analysis Methods), the platform will also be open for contributions from external groups as in previous projects.

The goal of this subproject is to extend existing structures for collaborative short-term forecasting of COVID-19 (the German-Polish COVID-19 Forecast Hub, https://kitmetricslab.github.io/forecasthub) to various respiratory pathogens (respiratory syncytial virus, pneumococcus, seasonal influenza). The current platform, shown in Figure 4.1, includes predictions of COVID-19 cases and deaths based on a number of different models.

The proposed project will add additional indicators on the above diseases, including the resulting burden on the health care system, as well as age-stratified outcomes. In this expanded framework, it will also be necessary to include a nowcast component. By nowcasts, we mean the statistical correction of the most recent data points in a time series of epidemiologic indicators that are incomplete due to reporting delays and usually underestimate the true burden. Figure 4.2 illustrates the relationship between nowcasts and near-term forecasts, as addressed here, and longer-term scenarios, as addressed in SP3.

The platform will include nowcasts, short-term forecasts, and associated uncertainties, from both mechanistic models based on the findings of subprojects 1 and 2 and purely data-driven statistical approaches. The main output will be a combined ensemble prediction, which is expected to have better stability compared to single model predictions and thus provide better guidance, especially to decision makers. The generated probabilistic predictions are subjected to continuous evaluation using modern statistical methods. In this way, our subproject will provide a useful feedback mechanism for model development. The main research questions are as follows:

1. How reliable are the generated nowcasts, and up to what time horizons are reliable predictions possible for the various pathogens and targets? Are there advantages of integrating epidemiologic information as generated in TPs 1 and 2 compared with purely data-driven approaches?
2. How can the quantile predictions of the individual models be improved by prediction combination/ensemble approaches?
3. Are reliable predictions by age group possible, and can stratified models improve overall predictions? How can stratified predictions be effectively visualized / communicated?

In all these issues, special attention is paid to the appropriate quantification of forecast uncertainty in terms of probabilistic forecasts. We will also involve relevant expertise from the public health field to ensure that the predictions made are actually of practical use. Dr. Johannes Dreesmann, Head of Department at the Lower Saxony State Health Office, has agreed to be available as a contact person.

Dr. Johannes Bracher
Prof. Dr. Melanie Schienle, KIT Karlsruhe