Modeling Infectious Disease Propagation in Airports Using Spatial Simulations on AWS

Last year, the joint blog post “Modeling the Spread of Infectious Disease Using Spatial Simulations on AWS” by Johns Hopkins University (JHU) and Amazon Web Services (AWS) demonstrated the steps involved in building and scaling spatial simulations on AWS for the Delineo Disease Modeling Project. The project simulated the complete cycle of infectious disease development in an urban area of around 240,000 people over a period of 100 days, providing valuable insights into the disease, such as forecasting lower and upper bounds of disease curves, and identifying clustering of regions and patterns.

Building on this work, the authors now aim to further pinpoint the transmission hotspots and specifically study the spreading of the disease in a large size indoor environment, such as an airport. To achieve this, they utilized Unreal Engine to create a realistic 3D simulation of entity movement and interaction within the airport space (Fig. 1).

Figure 1. Video of the crowd movement simulation inside airport (video opens in separate window)

Introduction and Objectives

Airport environments are inherently confined spaces where viral transmission rates can be particularly high. This work builds a virion transmission model tailored specifically for the airport environment, coupled with a realistic 3D simulation environment to study the aggregate people movement behavior.

The authors have carefully customized the transmission model to account for the unique spatial dynamics, population mix, airflow considerations, and mask-wearing practices prevalent in airport settings. The spatial simulation leverages Unreal Engine’s built-in realistic physics system using a 3D model of an airport to offer valuable insights into the complex interactions, collisions, and avoidance behaviors between entities within the environment.

The model’s robust, real-time depiction of crowd motion and infection probabilities provides a vital tool for airport modeling, with potential for airport authorities to utilize them and derive insights into real-world systems. This will enable them to implement data-driven safety measures and improve overall disease prevention strategies, ultimately enhancing the health and safety of both passengers and staff.

The work performed as part of this blog aims to achieve three key objectives:

1. Understand Real-Time Infection Dynamics: This work models the real-time transmission dynamics of infectious diseases within airport environments, taking into account factors such as proximity, duration of exposure, and the unique characteristics of different airport zones.
2. Identify Prevention Measures: The simulation highlights potential transmission hotspots and visualizes the effectiveness of various preventive measures, including enhanced cleaning protocols and changes to air filtration systems.
3. Inform Public Health Policies: By providing a clear, intuitive representation of how infections can spread in crowded airport spaces, this model serves as a powerful educational tool for both the public and airport authorities, informing decision-making processes and leading to the development of more effective safety protocols and public health policies.

Infection Model

The current model builds upon the foundation laid by the well-known Wells-Riley model, but addresses several key limitations to provide a more accurate and comprehensive assessment of infection risk. While the Wells-Riley model is useful for its simplicity and application to airborne transmission in well-mixed environments, it falls short in capturing the nuances of real-world disease transmission dynamics. Specifically, it does not account for variations in individual susceptibility, pathogen load, environmental conditions, or the critical role of close-contact transmission, which occurs through direct or close-range interaction between individuals.

In contrast, the new model incorporates these critical factors, offering a more nuanced and precise analysis of infection risk. By considering variables such as individual breathing rates, the pathogen’s infectious dose, variable environmental conditions, and the dynamics of close-contact interactions, our new model provides a more comprehensive and adaptable framework for understanding disease transmission. This enhanced accuracy is crucial for developing targeted interventions and improving public health outcomes. As such, our model represents a significant advancement over the Wells-Riley approach, addressing the inherent complexities of infectious disease transmission in a more effective and holistic manner.

The model begins by considering the probabilities of people getting infected through inhaled virions following Equation 1.

In Eq. 1, the probability of a person getting infected is given. N_c is the inhaled virions by an individual, N₀ is the upper limit of virions inhaled to get infected, with 900 copies being used, and this parameter is obtained from [1] and [2]. represents the probability of not getting infected, and subtracting it from 1 gives the probability of infection.

To model the close-contact infection and facility virion-concentration based spread, a general equation was derived as shown in Eq 2.

Eq. 2 encapsulates the number of virions inhaled by each person (N_c) and evaluates if an individual will become infected based on the concentration of virions in the environment (c_v), the fraction of virions inhaled by the individual (f_inh), the duration of exposure (t_room), the virions dose inhaled given the distance from the infected person (N_close), the time spent within 2 meters of the infected person (t_close), the mask filtering rate (m_filter) (assuming 0.01 for surgical masks and 0.5 for fabric masks), and the air filtering (a_filter).

The concentration of virions in a facility (unit: particles/liter) can be derived as shown in Eq. 3. It is calculated by dividing the sum of exhaled particles by n infectious people that have been to the facility by the facility size; where d is the particle degradation rate, calculated by , t is the time since person with index i enters the facility, r_i is the emission rate of person (unit: particles/minute), m_i is the mask filtering rate if the infected person wears a mask; V is the volume of the facility (in Liters).

In order to derive empirical parameters for simulations, Covid-19 variant-based approximations from past studies are used. Estimates are made for delta variant patients from 3.563∗10³ to 2.354∗10⁵ copies per minute and 6.635∗10³ to 7.796∗10⁵ copies per minute for omicron variant patients [3]. For the latter simulations, the omicron variant was primarily used.Eq. 4 represents the fraction of virions inhaled. It is calculated by taking the sum of inhaled fraction of j droplet size classes; where f_vj is the fraction of viruses in droplet size class j, p_j is the deposition probability for droplet size class j.

The authors conducted an investigation into virus transmission dynamics, with a particular focus on the role of respiratory droplets. To inform this analysis, they leveraged published data on the size distribution of respiratory droplets [4]. First, the team calculated the fraction of expelled viruses within each droplet size class, based on the percentage of expelled droplets in those respective size categories. Next, they determined the fraction of viruses in each size class by combining the expulsion probability with the volume of expelled viruses. To ensure the accuracy of these calculations, the researchers carefully adjusted the fractions to account for potential discrepancies in the original data sources. Finally, the probability of virus transmission for each droplet size class was quantified, considering the deposition probability of the droplets. By summing these individual transmission probabilities across all size classes, the researchers were able to obtain the overall fraction of virions that are transmitted upon inhalation. The virion dose inhaled per unit time (in minute) given the distance from the individual to the infected person is presented in Table 1.

Table 1: Virions inhaled per minute given the contact distance from the infected person

Building Spatial Simulations in Unreal Engine 5

The authors leveraged the MassEntity framework, a data-oriented system available in Unreal Engine 5 that specializes in large-scale crowd and traffic simulation. Specifically, they utilized the MassGamePlay and MassAI plugins to generate the designated number of entities within the airport environment and apply AI-powered behaviors and movements on them (Fig. 2). This approach allowed for the intelligent simulation of the airport’s population at scale.

To provide directional guidance for the simulated crowd, the authors employed a lightweight navigation system called ZoneGraph. Rather than enforcing strict movement tracks on the entities, this population movement flow generation method takes advantage of Unreal Engine’s built-in realistic physics system (Fig. 2). This enables valuable insights into the complex collision, avoidance, and interaction dynamics between the individual entities as they navigate the airport space.

Figure 2. Visualization of Entities generated by MassAI and following ZoneGraph directions in an airport

Running Simulations on AWS

In this workflow, several modules including Unreal Engine (UE) and JHU’s Delineo codes work in tandem, and an AWS architecture encapsulating these components is shown in Fig. 3. First, logging tools are used to obtain the positions, unique identifiers of each person within the Unreal Engine simulation. This simulation runs on an optimized Amazon EC2 GPU instance (g5.24xlarge) on AWS. Once this is completed, the simulation outputs (i.e. dynamic entity log files) can either be downloaded locally onto the remote user’s machine or stored in an Amazon S3 bucket for later analysis. These files can be extracted to another EC2 C4 compute instance for post-processing and disease spread analyses using DMP Delineo, and the outputs in terms of disease statistics can be exported back to the remote client. During the simulations, either regular RDP or a high performance remote desktop protocol, such as DCV can be used to visualize the airport and the various people walking through specific areas.

Figure 3. AWS Architecture diagram showing the cloud services used to build the workflow

Results and Discussions

Every second, movement information, including the X, Y, and Z coordinates of each entity, is exported from the real-time Unreal Engine simulation. A percentage of the population is initially set to be infected. Then, the infection model described above is applied to the rest of the population to calculate each individual’s probability of getting infected, based on their cumulative movement track, facility virus load, and close contact moments with the infected individuals.

The observation window for the simulation is set to 600 minutes (10 hours) to allow for 3-4 complete cycles of crowd entry and exit within the airport system. This timeframe accounts for the typical airport stay duration of 2-3 hours for most passengers. While extreme cases of individuals being stranded in the airport for more than 600 minutes may exist, these outliers are excluded from the scope of this study.

Fig. 4 illustrates the probability of getting infected for all entities under three different initial fractions of the infected population (i.e., 0.5%, 1% and 2%). The data points have been rearranged in descending order to provide a clear visualization of the distribution.

Figure 4. Distribution on likelihood of disease from high to low, based on different initial fraction of infected population

The result illustrates the relationship between the initial fraction of infected individuals entering the airport and the overall likelihood of disease transmission across the population. As the initial infection fraction increases, the probability of infection for the entire population rises accordingly. This can be attributed to the heightened probability of close contact between uninfected and infected individuals. The data reveals that the average likelihood of disease escalates more rapidly when the initial infected fraction grows from 0.5% to 1%, compared to the increase from 1% to 2%. This suggests a potential plateau in the correlation between the initial infection rate and the overall probability of transmission.

Figure 5. Distribution on likelihood of disease with close contact virions inhaled considered and not

Fig. 5 illustrates the distribution of disease likelihood under the same initial conditions (2%), but with the exclusion of virions inhaled from close contact with already infected individuals when applying the infection model. The substantial reduction in likelihood highlights the significant impact of close contact in driving disease transmission within the airport environment.

This observation can be partly attributed to the fact that most areas within an airport are open spaces with relatively high ventilation rates. However, the simulation reveals that during the 600-minute observation period, high concentrations of people occur within specific zones of relative confinement, such as the check-in counters, security areas, and cafes. It is likely that these are the key hotspots for potential disease transmission.

These findings underscore the necessity and importance of maintaining social distancing between individuals, particularly in densely populated areas to effectively control the spread of diseases. Additionally, the results suggest that even with an initial infection fraction as high as 2%, it remains challenging to accumulate high concentrations of virions within a single facility. This further emphasizes the critical role of close contact in driving disease transmission in airport settings.

Mitigation strategies designed to reduce the impact of disease spread will need to take these high-traffic, high-risk areas into account and implement additional precautionary measures to enhance overall safety and control the transmission of infectious diseases.

Conclusion & Future Work

This work demonstrates the valuable insights that can be gained by leveraging spatial simulations and a virion infection model to understand the complex epidemiological dynamics within confined airport environments. The current approach utilizes a cumulative infectious model to calculate the probability of infection based on the movement patterns through the simulation.

Moving forward, the authors plan to enhance the infection model by implementing dynamic, real-time infection probability calculations concurrent with the spatial simulations. This will allow them to understand the effectiveness of different modeling approaches applied to the same underlying aggregate movement patterns. The authors aim to establish a bidirectional coupling between the disease model and the spatial simulation, further integrating these critical components.

Additionally, the authors plan to vary all factors in our simulation, from different movement flows to initial infected setups, to achieve more comprehensive and holistic results. Regarding Unreal Engine visualization, future efforts will focus on increasing the number of entities, improving the physics of spatial movement, and introducing more complex movements and behaviors such as sitting, eating, and checking luggage.

References

[1] Basu S. Computational characterization of inhaled droplet transport to the nasopharynx. Sci Rep 11, 6652. (2021) https://doi.org/10.1038/s41598-021-85765-7

[2] Bale, R., Iida, A., Yamakawa, M. et al. Quantifying the COVID19 infection risk due to droplet/aerosol inhalation. Sci Rep 12, 11186. (2022). https://doi.org/10.1038/s41598022-14862-y

[3] Na, Liu et al. Comparison of the amount of SARS-CoV-2 exhaled by Delta and Omicron patients Journal of Infection, Volume 85, Issue 5, 573 – 607 (2022) https://doi.org/10.1016/j.jinf.2022.08.028

[4] Xiaojian, Xie et al. Exhaled droplets due to talking and coughing J. R. Soc. Interface.6S703–S714 (2009) http://doi.org/10.1098/rsif.2009.0388.focus