From agent-based simulations of disease dynamics to expert support on public health interventions

Abstract Agent-based epidemiological simulators have long been used to predict the dynamics of infectious diseases. They allow to combine compartment approaches with representing contact networks derived from real or synthetic populations. Often, scenario techniques are applied to model the impact of public health interventions. During the COVID-19 pandemics the agent-based simulator COVASIM has been established for modelling the spreading of SARS-CoV-2. It includes four different contact networks related to homes, schools, workplaces and environment. During the pandemic it became obvious that the current simulation techniques were not sufficient to cover important needs in public health administrations and society. While most simulations were based on scenarios predicting expected numbers for the infected, hospitalized or critically ill, these studies could not provide easily strategic options on a variety of equivalent public health interventions, which are all suitable to keep disease dynamics within acceptable bounds, set e.g. by hospital capacity. Knowing about a variety of equivalent options, however, is of crucial importance to enable a fair burden sharing throughout all groups of society. Therefore, we applied machine learning techniques to create an inverse model for epidemiological simulations based on COVASIM. For given transmission rates it outputs a range of quantitative interventions which are equivalently suitable to keep epidemiological dynamics within predefined bounds. Currently, it covers interventions like school and workplace closings and other mobility restrictions. Future work will be directed to include other interventions, like vaccination campaigns, and we will work to make the approach more generalizable for other infectious diseases. The tool may turn out helpful fur public health administrations and stimulate the public debate on the best way a society may take in a pandemic event or other situations where public health action has to be imposed. Key messages • Machine Learning allows to invert agent based simulators. • Instead of disease dynamics simulations a set of equivalent public health interventions consistent with pre-defined constraints can be calculated.