Short-term forecating of infectious diseases can contribute to situational awareness and resource planning during infectious disease outbreaks. During the COVID-19 pandemic, such forecasts have received considerable attention, and it is increasingly acknowledged that multi-model systems can improve the reliability of results. So-called Forecast Hubs have therefore been launched in various countries in order to coordinate modelling efforts, enable a coherent comparative evaluation of different models and their combination into a forecast ensemble. In the opening talk of this session, we will provide an overview of the German and Polish COVID-19 Forecast Hub, which we operated from May 2020 through April 2021, when it was merged into a larger effort led by the European Centers for Disease Prevention and Control. We will discuss practical aspects of coordinating real-time forecasting with numerous independent research teams as well as statistical and epidemiological challenges we encountered. Particular attention will be given to a pre-registered evaluation study which we conducted between October 2020 and April 2021. The results indicate that while deaths can be predicted with some success, case forecasts are very challenging, and in particular abrupt changes in trends are difficult to predict. Ensemble forecasts overall showed good relative performance, for most forecast targets outperforming most or all individual models.

Short term forecasts of the case incidence are a key component in designing public health countermeasures to control an ongoing epidemic. To perform such forecasts one usually estimates quantities related to the growth of cases, e.g. the reproduction number or the growth factor, which are readily available on the national level. However there is considerable heterogeneity in the speed at which cases proliferate within one country, making estimates of these quantities on the subnational level desirable for accurate forecasts. As subnational estimates pose more difficulties due to small incidences and spatial correlation, we combine epidemic modelling with techniques from small area estimation and spatial statistics to facilitate estimation and thus forecasts. The usefulness of our approach is demonstrated by simulation studies and through application to German case data on COVID-19.

The MOCOS agent based model is an advanced continuous time epidemic model which is in use for policy recommendation for the polish ministry of health since the beginning of the COVID-19 pandemic. Although the main focus of the model was and is risk analysis and forecasting for the COVID -19 pandemic, it can easily be adapted to the class of respiratory diseases. A special focus of the MOCOS model is the detailed representation of the effect of non pharmaceutical interventions like contact tracing, smartphone based tracing , regional lockdowns and school testing. We present a short overview about the main algorithmic features of the MOCOS model and review the forecast performance of the model based on the submissions to the European forecast hub. We also discuss the Poland specific difficulties of model based policy recommendations. We close the presentation with some challenges of short and medium forecasting for agent based models .

With the outbreak of the Covid-19 epidemic in Germany, a spread model was developed at Fraunhofer ITWM, which is successfully used for short-term forecasts. The results have entered the German-Polish and European Forecast Hub and still form the basis for advising the state of Rhineland-Palatinate. We start the talk summarizing briefly the integral equation-based model which fits contact and detection rates to reproduce counted numbers of cases and deaths. In a first analysis we consider the situation in Germany in spring 2021. In this phase of the pandemic, data was still available at high quality and daily resolution for different regions and age groups. Moreover, protection could easily be described by three states: susceptible, vaccinated, or recovered. These facts made it possible to fit even the characteristics of the virus, like incubation time and infectious period, without consulting literature. In contrast to many other groups at that time, we fitted explicitly the detection rate and the offset between infection and detection. This enabled us to reconstruct the actual effect of post-test isolation. Comparing the different trajectories of reporting data around the staggered Easter holidays in selected German states, we bring strong evidence that, in spring 2021, post-test isolation made a stronger contribution to the containment of Covid-19 than vaccination or contact restrictions. For this purpose, we are fitting contact rates, which change with contact restrictions, and detection rates, which change with the testing regime. Freezing two of the three rates for restrictions, testing, and vaccination at their values before Easter 2021, and continuing the third one as fitted, we can estimate the effect of each individual measure. It turns out that, in particular, tests at schools have played an important role. In the remaing part of the talk we will discuss recent improvements to our code dealing with transitions between variants and incorporation of waste water measurements.

In PDYN 1.5—the epidemiological agent-based model developed at ICM UW—for each agent the probability of contracting infection is calculated daily. Once the agent gets infected, the course of infection consists of a chain of subsequent disease states that the agent goes through. These states represent the severity of symptoms such asymptomatic, mild symptomatic, requiring hospitalisation or requiring ICU treatment. The probabilities of transition from each state to other states determine the total probabilities of each particular course of the whole infection.

The probability of infection as well as transition probabilities between states may be reduced due to the agent’s immunity gained from vaccination or past infection within the frame of a multi-step immunity mechanism. The proposed immunity mechanism captures well-known characteristic of COVID-19 vaccines, i.e., much better efficacy against the severe outcomes of the infection than against the infection itself. The mechanism allows for differentiation of the probabilities of transition to particular severe disease states in the vaccinated and non-vaccinated groups of agents.

Using PDYN 1.5 calibrated to the Polish epidemiological data, we were able to show that our model along with the above mentioned immunity mechanisms is able to capture the reduction of the risk of hospitalisation relative to the risk of the infection itself resulting from vaccination or past infection. Moreover, our method proved its validity in the predictions of delta and omicron VoCs waves in 2021/2022 winter season that were published in Covid-19 European forecast hub. Additionally, to gain further insight into our model of immunity, we carried out an in-silico epidemiological study in which we evaluate the vaccine efficacy (VE) in our simulations and compare it to the real-life data.

In this talk, we would like to present these results as well as explain the mentioned immunity mechanisms implemented in PDYN 1.5 in more detail.