Agent-Based Modeling of COVID-19 Epidemic
"The same as biological life is an emergent property of chemistry, the phenomenon of the COVID-19 pandemic is the emergent property of human interaction"
- Morphogenesis, Nahuel Gonzalez
When improving a national or regional pandemic model, PANDEXIT software provides the ability to discover dependencies and quantify the effects of policy change, resulting in the optimal targeting of efforts and prioritization of changes.
Epidemiological research benefits from the ability to model contagion dynamics and social interactions. With epidemic simulation forecasting and scenarios can be tested and evaluated, informing mitigation strategies and increasing preparedness.
PANDEXIT Product Summary
PANDEXIT is a tool to help decision makers evaluate the outcome of different COVID-19 lockdown policies through a reasonable estimation of the evolution of the pandemic in several configurable scenarios.
Specifically, PANDEXIT is an agent-based simulation - a method that has a long-established history in academic research, both for pandemics and economic modelling. It follows every virtual agent (adult or child) through the day, matching its travelling, educational, working, and buying behaviour to the one corresponding to its demographic group, and simulating its interactions to better model the evolution of the pandemic under different lockdown scenarios in a detailed manner – be at a local, regional, and/or national level.
It’s a far superior approach then the commonly used statistical techniques that rely on limited aggregated data, often times after-the-fact analysis, and most importantly ignore individual human interaction points in which viruses thrive.
The level of granularity that is required to provide much more accurate predictions is one important aspect that distinguishes PANDEXIT from other available agent-based models. The other factor is that the algorithm has been fine-tuned over the last 30 years.
Using the PANDEXIT tool will provide a significant pillar of confidence in determining the best strategy in a way that considers the particular demographic, geographic, and economic conditions to avoid the costs of an erroneous approach.
Description of Current Situation / Background
The COVID-19 pandemic has forced a worldwide lockdown and quarantine of suspected cases, which includes restricting the businesses and organizations that are considered essential to remain open, and severely limiting the movement of people inside the country. However, eventually the lockdown and travel restrictions will need to be lifted gradually. That will require unparalleled consideration of public health concerns and minimising catastrophic economic consequences. Which policies to implement, in which order, how to gradually phase them in, and how to consider the peculiar conditions of each territory is not only a sovereign prerogative, but also a problem subject to the particular demographics, geography, political divisions, and economical structure of any country. Encompassing the complex scenarios presented by this novel crisis, a possible second wave of contagion, and future challenges of a related category will require original tools.
What is the Problem in The Current Situation?
The necessary information for critical decision making during unprecedented crises is generally lacking. Statistical models for previous pandemics might no longer apply due to the natural growth in the country or modified conditions. For instance, the SIR (susceptible, infected, recovered) model is an example of a widely used statistical model, where the three groups of persons and their size, together with their time evolution, is considered in order to describe the evolution of the pandemic. This does not at all consider the fact that the phenomenon of the COVID-19 pandemic is the emergent property of human interaction.
The bottom line is that classic statistical modelling cannot simulate how big data of demographics, behaviour, transportation routes, and work conditions all work together in minute detail at a nationwide level.
Subsequently, we have the current situation where determining the best strategy that considers the particular demographic, geographic, and economic conditions of a country to avoid the economical and public health costs of an erroneous approach appears to be remote.
How Does PANDEXIT Contribute to Solve the Problem?
Instead of using statistical techniques to capture the behaviour of relevant groups in an aggregated fashion, we use agent-based modelling that focuses on simulating every individual part of the system under consideration.
Agent-based modelling is a class of computational models for simulating actions and interactions of autonomous agents to assess their emergent effects on a system, i.e. the pandemic, as a whole. Agent-based models are capable of evaluating the outcomes for the specific conditions of a country or a region.
With access to big data describing demographics, behaviour, transportation routes, and work conditions, together with the extensive availability of computational resources, a country-wide simulation can be carried out in minute detail. All the calculated data is then processed by a proprietary Artificial Intelligence engine that delivers human readable information.
The result being that PANDEXIT will support decision makers in evaluating the outcome of different COVID-19 lockdown policies through a reasonable estimation of the evolution of the pandemic in several configurable scenarios.
Agent-based simulations of nationwide scale are usually carried out through massively parallel high-performance computing (HPC) and thus require intensive computational power and data access.
For PANDEXIT’s proof-of-concepts, ADGS used a cloud-based deployment using virtual machines that allows scalability while being cost effective in all stages of the project. For example, ADGS runs simulation with a set of 5 million agents (virtual persons, adults and children) and 5000 work places, together with 1000 selected locations (like hospitals, schools, and other essential businesses) can currently run without need for HPC with a time granularity of few minutes per simulation step. Increasing the size, the number of agents, and the resolution of the simulation can be carried out by scaling up the hosting server or by adjoining additional servers to the running cluster. As agent-based simulations are highly parallelizable by design, cloud deployment allows optimum allocation of computational resources while adjusting the cost to the current needs.
From the point of view of the availability of information, public access to big data describing demographics, behaviour, transportation routes, and work conditions, allows a countrywide simulation to be carried out with enough detail. However, access to fine grained data will be needed to capture the minute details of the country. The data can be stored and feed to the simulation in a privacy preserving way, completely anonymized, and open to auditing.
The data is then analysed and processed by our Artificial Intelligence engine that builds a dynamic human readable interface, locating foci on a map, and providing curves showing the infections, deaths and recoveries.
Current Stage of Development
ADGS has developed a proof-of-concept (PoC) simulation for the COVID-19 spread in Qatar as soon as the lockdown measures were put into effect in the country, in order to demonstrate the feasibility of the project, its own technical capability, and to show an approximation to the results obtainable with a full-fledged project.
The prototype simulations included a restricted set of agents, 50,000 in total, following fictional people living in Doha, the capital city, and its surroundings throughout their daily lives, with an assigned home and work to which they commute daily in several transports, like car, taxi, bus, metro, walking, etc. The last version models the entire population of Qatar, 2.8 million agents.
In another case, the Ministry of Defence of Argentina requested ADGS to build a simulation for their Military Academies and Military Bases that are located across the country, The agency provided a large dataset of information for 40,000 agents, specifically transportation, location, civilian addresses for the families as well as family data, ages, etc.
The aforementioned PoC simulations allow setting the infection parameters like the rate of infection, age ramp (how much are the elderly affected in comparison with younger age groups), hourly death probability for patients (unaided and in hospitals or clinics), degree of adherence, efficiency of masks (depending if weared by non-infected, asymptomatic or infected), as well as enabling the possibility of reinfections after recovery. Parameters and scenarios can be altered while the simulation runs. Meaning, policies, such as reopening or keeping schools closed, changing the rules, etc., can be enabled or disabled in the PoC simulation in order to track the evolution of the pandemic in the country.
Infections can be tracked individually (when the number of infected individuals is low) or by region. Feasibility studies on the possibility of extending the scope of the simulation to include a one-to-one matching of inhabitants and workplaces have been conducted successfully.
Both of the aforementioned government agencies are currently in the process of deciding how they will deploy PANDEXIT not only for the current phase of the pandemic but also for the inevitable subsequent waves.
We propose the creation of an agent-based simulation following every virtual agent (adult or child) through the day, matching its travelling, educational, working, and buying behaviour to the one corresponding to its demographic group, and simulating its interactions to better model the evolution of the pandemic in a specific region under different lockdown scenarios. The average calculation is 5 second per one million agents per day (one step).