Predicting and preventing COVID-19 outbreaks in indoor environments: an agent-based modeling study

How to mitigate the spread of infectious diseases like COVID-19 in indoor environments remains an important research question. In this study, we propose an agent-based modeling framework to evaluate facility usage policies that aim to lower the probability of outbreaks. The proposed framework is individual-based, spatially-resolved with time resolution of up to 1 s, and takes into detailed account specific floor layouts, occupant schedules and movement. It enables decision makers to compute realistic contact networks and generate risk profiles of their facilities without relying on wearable devices, smartphone tagging or surveillance cameras. Our demonstrative modeling results indicate that not all facility occupants present the same risk of starting an outbreak, where the driver of outbreaks varies with facility layouts as well as individual occupant schedules. Therefore, generic mitigation strategies applied across all facilities should be considered inferior to tailored policies that take into account individual characteristics of the facilities of interest. The proposed modeling framework, implemented in Python and now available to the public in an open-source platform, enables such strategy evaluation.

shows the parameter values used to create daily schedules for each agent in the facility. The first column is the type of activity that the agent does. The location of the activity is randomly chosen from a list of compatible locations in the facility. The duration of the activity is randomly chosen between the minimum and maximum duration. The number of occurrences during the day is also randomly chosen between the minimum and maximum number of occurrences. All random sampling for scheduling purposes assume a uniform distribution.

Name
Minimum Duration  Table 2 shows the list of parameters required for the Monte-Carlo Long-Term Transmission Algorithm (MC-LTTA). Some of those parameters are based on the epidemiology of the disease of interest and are, essentially, the probability for an infected individual to be asymptomatic for the duration of the infection, the first day after infection when an agent becomes contagious and the last day of their contagious period. In this study, we use the values of 0.4, 3 and 14 respectively for these parameters based on relevant reports about COVID-19 in the literature. A sensitivity analysis around these values is presented in the following section. The remaining parameters are user inputs such as the number of agents, the number of days to run the simulation for and the initial list of asymptomatic (or infected) individuals. As mentioned before, the pair contact probability computed from the ABM simulations is also a key input.

Sensitivity Analysis
We explore the influence of the following parameters: the latency of the disease (how long it takes before infected individuals become contagious), the percentage of the population that's infected at the beginning of the simulation, the probability that an infected individual is asymptomatic, and the duration of any pre-symptomatic period. In all cases, we repeat the simulations for the first facility described in the paper and at half occupancy. Figure 1 shows the influence of the latency value on the percentage of infections as a function of time. We use latency values of 1 to 4 days. As expected, the population of infected individuals starts increasing only after the initial latency period is completed. We also observe that the curves for lower latency value are smoother than the ones at higher latency values. This because at higher latency values, the transmission dynamic is similar to a cohort-based transmission whereby a group of individuals will get infected, go through the disease progression process and recover together.

Asymptomatic Probability
In Fig. ?? we vary the asymptomatic probability from 25% to 40%. The results show that the trajectory of the disease within the facility does not change significantly as a function of the asymptomatic probability in the 25% to 40% regime. We use a value of 40 % in the paper.

Considerations for Implementation
In this work, we use the CDC definition of close contact to model the transmission risk as a delta function using a proximity distance and cumulative duration cutoffs of 6 ft and 15 min respectively. As mentioned before, a delta function is not the most realistic way to model the transmission probability for at least two reasons: (1) transmission can occur when individuals are more than 6 ft apart and have been in contact for less than 15 min and (2) it doesn't take into account other important factors that may drive the transmission risk. Several publications have sought to derive expressions for the highly complex and multivariate transmission risk of a respiratory infection such as COVID-19 by taking into account respiratory droplet [1,2], sneezing and coughing dynamics [3,4], a comprehensive list of factors [5], etc. One or more of these expressions can be used within our proposed framework to obtain results that are more realistic. Similarly, as more data is gathered about new strains of SARS-COV-2, the transmission function can be updated accordingly.
Although the full importance of airborne transmission has not been estab-  lished yet, several studies have suggested that stagnant air with small droplet nuclei that includes the virus [6] as well as contaminated surfaces are potential infection routes. In other words, there is a non-zero probability that users of a facility get exposed to the virus without ever having direct contact with infected individuals. While this risk factor is left outside of the scope of this work, indoor air quality (e.g. quality of the ventilation, humidity, temperature, etc.) and surface cleaning policies can be used to assign a probability of agents getting infected without direct contact. This probability of infection can be incorporated into the MC-LTTA framework in a similar way that outside infections are incorporated.