1 Introduction

The COVID-19 pandemic has vividly demonstrated the impact emerging infectious diseases can have on almost all aspects of our everyday lives. The fight against the disease was omnipresent, ranging from mask mandates over vaccination strategies to country-wide shutdowns. Decision-makers within the public-health sector faced the challenge of choosing appropriate countermeasures for a disease that has only recently emerged. These decisions could, therefore, not be based on previous experiences but required a rapid (scientific) evaluation. During the pandemic, infectious disease modeling has gained importance and recognition as a valuable tool for understanding disease dynamics and predicting the future course. The models could assist decision-makers, e.g., by evaluating the effect specific countermeasures could have on the future disease course. Various methods exist for simulating infectious diseases. One of the most prominent and also the focus of this paper is agent-based modeling. Here, the population is represented at an individual level, with distinct characteristics and disease progressions for each individual.

The EpiPredict project aimed to develop a highly flexible simulation platform for the construction, execution, and analysis of infectious disease models. The platform should be easily adaptable to newly arising disease strains and should allow simulations with a fine geographical resolution, i.e., on a German county level. Therefore, high spatial accuracy was required to resolve the regionally varying population attributes. Due to these prerequisites, agent-based models were selected as suitable simulation tools, as they offer high modularity, easy extendability, and high regional resolution due to their inherent heterogeneity.

In the following, we will first introduce the EpiPredict platform, discuss the agents employed in our models in the context of AI, and then describe two platform applications. Finally, we will summarise the project and briefly present the projects that emerged due to the EpiPredict project and further research in agent-based infectious disease modeling.

2 The EpiPredict Platform

The EpiPredict project was a two-year project between 2020 and 2022 funded by the German National Research Platform for Zoonoses. During the preparation of the proposal, there was no mention of the COVID-19 virus. However, the overarching group of corona viruses was already one of the main aspects considered in the proposal. Due to the emergence of the pandemic, the focus was shifted even more toward the simulation of corona viruses with a specific emphasis on the COVID-19 virus. The EpiPredict platform, the main result of the EpiPredict project, includes many components designed to allow for an accurate simulation of COVID-19, e.g., including an individual-to-individual transmission due to the limited impact of smear transmission on COVID-19 spread. The modular simulation platform has a user-friendly interface, allowing non-experts to build and analyze epidemic models. At the platform’s core is an agent-based simulation framework that simulates each individual in a defined population, including their daily routines, e.g., going to work/school, shopping in a supermarket, or staying at home, as well as individual contacts and disease progression. As these agents are the key elements of the simulation platform, it is worth taking a closer look at the underlying concept of an agent employed in the EpiPredict platform.

2.1 Agents in Epidemic Disease Modeling

The application of agents in epidemic disease modeling brings specific requirements with it, ruling out a number of agent types. Agents come in different complexities. The most primitive form is the simple reflex agent [1] (p. 48ff.). It perceives its environment to obtain the as-is state and selects an action to perform based on a defined set of rules. Figure 1 illustrates the behavior of the simple reflex agent, which observes the current state of the environment through a sensor, selects an action based on the as-is state and a defined set of rules, and performs this action on the environment through its actuators.

Fig. 1
figure 1

Illustration of a simple reflex agent. Adapted from [1] (p. 49.)

Full size image

There are several extensions to this simple reflex agent, including goal-based agents, which do not base their action solely on a set of rules but instead on an overarching goal. In addition, utility-based agents are known, which evaluate their potential actions in terms of their impact on the environment based on a utility function. While in the field of AI, highly complex agents are used, e.g., learning agents in reinforcement-learning [2, 3], agent-based epidemiological models usually contain relatively simple variants, i.e., simple reflex agents [4] or simple model-based agents [5], as their low complexity makes them computationally more feasible for large-scale simulations [4,5,6,7,8].

2.2 Agents in the EpiPredict Platform

The previously discussed definition of agents shows the advantage of using agents in our simulation framework. Since we want to simulate a heterogeneous population, in which each individual potentially has different characteristics and acts depending on these characteristics, agents are a suitable basis. They allow the formulation of internal rules that can vary for each agent as they incorporate the individual’s characteristics, resulting in distinct behavior patterns. In addition, the simulated populations are in the range of a couple of thousands up to millions of individuals, and the agents, therefore, must be relatively simple to avoid high computational costs. For this reason, simple reflex agents are employed in the EpiPredict software. More complex agents could lead to dynamics closer to the real world, e.g., by including agents that deliberate on minimizing their infection risk and act accordingly. However, their complexity also makes large population sizes computationally more expensive, leading to increased simulation times, which would correspond to a reduced applicability of the simulation tool. To illustrate the usage of agents in our simulation framework, consider the following examples:

Example 1: Agent movement

In our simulation, agent movement is determined by so-called schedules, which correspond to timetables, including the locations agents should visit at defined times. Therefore, schedules constitute the rulebook for the individuals on which their actions (the movement) are based. During each timestep, i.e., the discrete unit of time used in the models, the agents observe the current simulation time, look up in their schedule which place they are supposed to visit at the defined time, and move to that place if they are not already there. Figure 2 illustrates the movement process of a simple reflex agent in our simulation. It is also possible to include additional conditions in the agent schedule. An example is including the agents’ current disease state, e.g., prohibiting movement to a specific location if the agent exhibits symptoms of a particular disease. Additionally, global simulation properties, e.g., the current infection numbers, can be set to influence the individuals’ movements.

Fig. 2
figure 2

Schematic illustration of a simple reflex agent movement in the EpiPredict platform

Full size image

Example 2: Agent Contact

Besides individuals, places are also implementations of simple reflex agents. Places in our framework are locations that individuals can visit according to their schedules. During each timestep, a place observes if it includes at least one infected individual and if the total number of individuals is greater than two. If these conditions, corresponding to the agents’ action rules (see Fig. 1), are met, the place randomly establishes contacts between the included individuals dependent on pre-defined contact rates, thereby acting upon its environment.

3 Platform Applications

A direct application of the EpiPredict platform during the EpiPredict project was the reconstruction of the first major COVID-19 outbreak in Germany. In February 2020, a carnival event in Gangelt, a small municipality near the border between Germany and the Netherlands, turned into a super spreading event, leading to the first failure of infection tracing and an uncontrollable disease spread. The simulation platform was successfully employed to reproduce the infection numbers after the initial super-spreading event. For this simulation, we generated a synthetic population of approximately 12,000 individuals, based on German census data [9]. The individuals were heterogeneous in age, sex, and eligibility to work. We employed schedules depending on these characteristics, i.e., children were assigned schedules that included visiting schools, while individuals eligible to work were assigned schedules that included visiting workplaces to allow heterogeneous behavior of the simulated agents and thus increase the accuracy of the simulation. Adjustments were made to the agents’ schedules, i.e., disabling movement to work and school within a specific date range, to reflect the complete lockdown applied within the affected county after the super spreading event. In addition to reproducing the case numbers, it could be shown that the employed interventions drastically reduced the number of infections by simulating the disease dynamics without the employed interventions. Computation times were in the order of minutes, even for this relatively small population of simple-reflex agents. Increasing agent complexity would drastically increase the computation time making, e.g., model calibration computationally unfeasible.

In addition to the simulation of human diseases, a modified version of the simulation platform was employed to simulate the spread of a Hantavirus within a mouse population. The Hantavirus is a family of viruses that spreads asymptomatically within rodent populations but may cause severe illness if transmitted to humans [10]. The goal was mainly to investigate the so-called spill-over effects causing the spread of the Hantavirus between different mouse populations. These effects could successfully be reproduced during the employed simulations.

4 Summary and Outlook

In summary, the EpiPredict project has resulted in the construction of a valuable agent-based simulation framework that can be easily employed by people of various backgrounds, especially non-experts, to simulate the spread of infectious diseases within highly heterogeneous populations. The framework uses simple reflex agents to represent places and individuals within the population. These were chosen to allow for a feasible computational complexity while simultaneously allowing highly autonomous actions of the individuals, thereby appropriately representing the real world. The framework was validated by successfully applying it to analyze the spread of various diseases within human and mouse populations. Looking ahead, several follow-up projects build upon the achievements of this project. One such project is SpaceImpact. While EpiPredict’s goal was to create a simulation platform that allows the flexible simulation of various diseases, SpaceImpact aims to forecast COVID-19 case numbers on a county level throughout Germany by using and extending the EpiPredict platform. The extensions include integrating parallelized simulations of multiple models, implementing a calibration procedure, and using regional mobility data. To allow these regional forecasts, a significant focus must be placed on the underlying regional data being used to simulate the counties, i.e., the population, place, and accurate, current infection data. This becomes especially challenging in later stages of the pandemic, where data availability decreases due to lower testing rates and less funding for surveillance. While SpaceImpact focuses on forecasting COVID-19 case numbers, the flexibility of the EpiPredict platform allows an easy adaption in the case of a newly arising disease to obtain valuable forecasts for the disease dynamics.

Another project that is a result of the EpiPredict project is OptimAgent (Optimal control of the epidemic under heterogeneity conditions—decision making perspective on agent based modeling) more specifically the subproject GEMS (Development of the German Epidemic Micro-Simulation System). This project is based on the expertise that has been gained during the course of the EpiPredict project. During EpiPredict, it became clear that developing a highly flexible, user-friendly simulation platform would come at the cost of computational efficiency, which was necessary due to the project’s aim. For large-scale simulations, i.e., the entire German population (\(\approx 80\) million individuals), the importance of computational feasibility increases significantly such that the applied approach should prioritize performance over usability. Within OptimAgent, the aim is to apply this lesson learned and create a large-scale simulation platform focused on performance, including, e.g., parallelization capabilities to allow the simultaneous simulation of a large number of individuals. It is an interdisciplinary effort to develop a standardized framework that can be used as decision support for public health during potentially emerging diseases, including, e.g., the simulation of various intervention strategies and the evaluation of their efficiency and effectiveness. This project adopts an individual-based modeling approach, where individuals represented by simple reflex agents are connected through networks that allow the spread of the disease without explicit simulation of the agents’ movements. Through the simple implementations of agents and the non-explicit simulation of movement, the software aims to be capable of rapid simulations for large population sizes.

Further developments can be expected in the broader landscape of agent-based disease modeling. Advancements in computational power should allow increased agent complexity, enabling the inclusion of complex internal behavioral models like BDI (Beliefs, Desires, and Intentions) [11] or PECS (Physical conditions, Emotional state, Cognitive capabilities, and Social status) [12], which would correspond to the implementation of model-based or utility-based agents [1]. Even more complex internal behavior models could potentially become interesting for future agent-based infectious disease models such as the SOAR architecture [13], which is a system that includes various types of learning mechanisms, thereby being an implementation of learning agents [1]. Adopting these techniques can enhance the accuracy of individual behavior representation, thereby leading to an overall improvement in model accuracy. It is important to note that more complex models generally require more data on which parameter assumptions can be based.

Furthermore, the integration of AI techniques in other regions of agent-based infectious disease modeling can be expected. AI can, for example, assist in the calibration endeavor for agent-based models, which is computationally intensive due to the models’ complexities and the large number of simulation runs required because of their stochasticity. So-called machine learning surrogate models have recently gained popularity and can be employed as calibration tools for agent-based models [14,15,16]. Surrogate models map the input parameters of an agent-based simulation to certain output variables, i.e., the case numbers for infectious disease modeling, without the need to perform simulations of the original model. This allows a computationally efficient calibration as parameter sets can be evaluated using the inexpensive surrogate model with appropriate parameter sets then being applied to the actual model.