Agent-based tracking of disease spread with dynamic models of travel behaviour in a pandemic

Epidemic models are essential tools to coordinate all aspects of the response to pandemics. Models can inform policy makers on strategies for vaccinations and testing, but also to trigger mitigation measure such as the 'lockdowns' encountered during the COVID-19 outbreak. However, historically epidemic models have relied on crude approximations of infections processes such as the spread of droplet between individuals over space, and time. Here, we propose to use and further develop a well-established technology -agent-based models of daily travel behaviour- to simulate epidemics with considerably increased resolution. These models produce by design a network of encounters between individual agents (people, vehicles, etc.) at very high spatial (e.g. the bus used by the agents, or the church visited) and temporal resolution (each second of the day). The technology enables tracking chains of infections, as well as the health status of all agents over time. The strength of the approach lies in the diversity of behaviours that can be attributed to agents, which in turn enables to simulate a wide range of targeted policies: timing, duration, and spatial extent of containment policy, but also restrictions specific to age, professions, etc. Agents can be generated with any attributes, as long as suitable conditional distributions are known, e.g. morbidity by sex, frequency of home-office by income level, or mobility tool ownership (e.g. car, season ticket etc.). Since September 2019, the IVT is tracking the travel behaviour of a panel of 1,200 persons providing a unique insight in their behavioural response. We have detailed calibrated agent-based and activity-based simulations of daily life available for Switzerland and the region of Basel. The existing models that have originally been developed to simulate transport will be supplemented to incorporate detailed epidemic transitions in collaboration with the epidemiologists on the team ("episim"). Concretely, we propose to address the following questions. After a careful initial and then continuously updated calibration we will first simulate activity-focused containment policies against COVID-19 with 4 versions of the transport model. The measure of effectiveness for each policy will be defined jointly with the team, but will likely focus on intensive care occupancy, acute COVID-19 cases, days lost to illness, and days lost to quarantine measures. Confidence intervals will be calculated for each effectiveness metric. •Phase 1: Current agent- and activity-based average workday models including detailed vehicle-based public and private transport simulation.•Phase 2: Integrate a weekend implementation of the Switzerland model which also includes new arrivals and departures from/to abroad and inclusion of small children. •Phase 3: Integration of the behavioural adaptations in response to the constraints imposed and in response to the risk assessment of the agents.•Phase 4: integration of the interactions within the households and social networks to capture spatial mixing of the population.Second, we will compare results of the analysis of the four models against epidemic surveillance data, and against the output of traditional epidemic models based on differential equations. The comparison will focus on COVID-19 cases and intensive care hospitalizations for the following situation: "Backcasting" (reproduction of previous epidemic), as well as "Forecasting" with and without therapeutic medical interventions and contact tracing. In addition, we will develop dashboards to allow non-experts to interact with the simulations (hospital administrators, general public, etc). From a practical perspective, this proposal aims to optimize the current response to COVID-19, by exploring complex and spatially-heterogenous policies (Cantons and trinational regions with border effects). Our results will also be relevant for preparedness against other pathogens with pandemic potential. Finally, this proposal will also address a fundamental question in disease modelling: what are the respective merits of the 'traditional' equation-based models vs 'new' agent-based models in terms of projection accuracy, parametrization, and computational cost.