Q&A: COVID-19 Mass Vaccination — Simulation, AI Application and Real-World Implementation


The work presented by Dr. Ali Asgary in this webinar with AnyLogic has also featured on Bloomberg and CTV News Ottawa. Both sources provide excellent context for how the simulation and AI-based developments are being used in practice.

As COVID-19 vaccines have become available, many challenges have needed resolving. Not least, ensuring sufficient supply and effective distribution.

At our webinar, March 2021, guest presenter Dr. Ali Asgary of York University, Canada, shared insight into drive-through COVID-19 mass vaccination development using simulation, machine learning, and its online application. The webinar shows how the model is being used by public authorities.

Read on for more details, the webinar recording, and Q&A answers with Dr. Ali Asgary.

COVID-19 Drive-through Vaccination Model and AI

Different methods are needed to vaccinate millions of people quickly and safely during the COVID-19 pandemic. Drive-through is one of the mass vaccination options that are widely applicable. To plan, design, and implement drive-through vaccination facilities, public health agencies are using simulation technology.

During the webinar, Dr. Ali Asgary presents the drive-through vaccination simulation model from his research “Artificial Intelligence Model of Drive-Through Vaccination Simulation”. He also explains the core structure of the model: screening, registration, vaccination, and observation. To accurately reflect those steps and the logic behind the drive-through vaccination process in the model as well as speed-up the model-building process, Dr. Ali Asgary’s research team used the Road Traffic, Pedestrian, and Process Modeling libraries that come with AnyLogic.

Mass vaccination simulation

The simulation shows waiting times and the number of vaccinations possible for different simulation input settings, including the number of vaccination lanes, the number of staff, and the time needed for procedures.

Using the AnyLogic Cloud API to expedite run times, the research team ran the simulation more than a hundred thousand times and created a large dataset from a wide range of input parameters. The machine learning model developed from this dataset can accurately predict output values and has been deployed as an online AI application (Try it!).

If you are interested in Dr. Asgary’s research, he and his research team are happy to partner for application cases. You can reach them at: asgary@yorku.ca.

Answers to the questions that were asked of Dr. Ali Asgary during the webinar

Does the model simulate any of the rare but potentially disruptive events, such as a patient having anaphylaxis, a car emergency, etc.?

The model has the potential for inclusion of such events, but they are not considered directly. However, because such emergency provisions are considered in the drive-through design (extra attention booths or emergency spacing) these events should be easily taken care of by taking the problematic patient of a car out of the process.

How were the ranges for the parameters determined?

The parameter values for screening, registration, vaccination, and recovery rates were used in some of the previous cases of mass vaccination and drive-through facilities, and from existing protocols. Due to the variance in possibilities though, these metrics were specifically set as parameters so that the model can easily be adapted as more fine-tuned data is known.

How do car agents select which lane to go to? If it’s random or based on the shortest queue, could heuristics be used to improve the results?

Cars go to lanes with shortest queue. This is important in cases where cars with more than one passenger are randomly entering the system.

However, the user can change the model’s parameters to change this. The “Dedicate Lanes to HOVs” and “Adjust HOV lanes dynamically” checkboxes will enable the simulation to dynamically allocate enough HOV lanes based on the past hour averages of the HOV queues and non-HOV queues. Additionally, the “Parallelize Service to LOVs” will allow simultaneous service delivery to vehicles with one or two passengers when there are enough staff available at the service booth.

What feature of AnyLogic has been used in this simulation model which cannot be addressed by other simulation software?

AnyLogic was used as it provides opportunities for effective visualization in both 2D and 3D, it has built-in ways to perform scenario analysis and optimization, and the ability to have agent-specific attributes and behaviors in combination with a discrete event model. AnyLogic Cloud and its API also provided a streamlined way to execute thousands of simulations quickly and to easily retrieve the data from the experiments.

Has this simulation been applied or is yet to be applied?

Our cloud version of the simulation and the AI version, as well as our publications, have been accessed many times since they have become public. We can safely assume that some of these are done by actual mass vaccination planners, some of whom approached us. This simulation model was used at a vaccination site at Denver, Colorado and one in Arnprior in Ontario, Canada. A third site in Chicago, Illinois is currently in the planning phase.

What's the role of AI in this model?

AI is not used to drive any of the internal components or logic within this model. Instead, it was used as an approximator of simulation results. Due to all the combinations of possible inputs, it can be time and computationally expensive to run hundreds of thousands of simulation runs.

A large dataset was able to be generated by running the simulation in the AnyLogic Cloud and fed into an AI agent to learn how to predict outputs based on the inputs. This enables users to be able to configure inputs and retrieve the results extremely quickly.

What is the input/output in your neural network model? What is the thing that you are trying to predict?

The AI application allows users to configure the same inputs that were used in the simulation model – including service times, shift information, number of open lanes and staff per lane, and additional settings. The outputs include number of cars passed, passengers passed and average wait time and a histogram depicting the distribution of wait times. Read more >>

What kind of AI algorithms were used here?

A supervised learning algorithm was used, training a neural network consisting of 5 fully connected, feed-forward layers using the ReLU activation function. Read more >>

Additional and related resources:

Related posts