Since the COVID-19 pandemic started, efforts for vaccine development and production began in many different countries. It is argued that an effective vaccine can be the best solution to end and minimize the impacts of the pandemic. While some countries have already started vaccine testing and production, it is estimated that approved vaccination solutions for large-scale implementation can be ready as soon as January 2021 or even earlier [1
]. Once a vaccine becomes available, the next challenge would be the vaccination of large numbers of people in a short period of time to minimize further human and economic impacts of the pandemic [2
]. Rapid mass vaccination requires many local vaccination clinics. Considering that a vaccine will become available for the general public use in the near future, it is important to start planning ahead of time to be able to implement mass vaccination effectively and efficiently [4
]. Therefore, access to mass vaccination modeling and simulation tools has become very important for public health units that are going to plan, manage and run different types of mass vaccination facilities.
Use of drive-through vaccination has been tested during past public health emergencies such as the H1N1 and during the COVID-19 pandemic for rapid testing and it has shown some promising outcomes [5
]. A significant advantage of the drive-through clinics is their lower virus transmission possibility compared to walk-in clinics because people are isolated in their cars and are not in direct contact with other people [7
]. They are only in contact with the immunization staff that are likely vaccinated first and are equipped with personal protection equipment. However, studies show that rapid vaccination using drive-through clinics requires proper site selections and design, human resources management, and careful attention to operational and logistical details [8
COVID-19 mass vaccination can be implemented using a combination of traditional approaches such as health clinics, pharmacies, nursing homes, schools, workplaces, places of worship, and innovative and temporary facilities such as drive-through and large stadiums, and parking lots. Use of large temporary facilities allows for rapid and safer mass vaccination [9
]. Drive-through facilities have lower disease transmission, low virus exposure, have large throughput, provide better security, and are more accessible and comfortable particularly for individuals with mobility issues or are geographically dispersed [2
]. The drive-through method has some limitations and shortcomings as its usage is influenced by weather conditions, requires suitable and available spaces, and significant logistical planning. Moreover, drive-through can cause traffic issues in the localities and may expose staff to carbon monoxide exposure [6
]. Despite these, drive-through facilities for mass vaccinations are recommended and are being seriously considered as part of the COVID-19 immunization process [13
Drive-through vaccination was largely used in the USA in 2009 during the H1N1 influenza pandemic. Subsequent studies found that with reasonable processing time and negligible carbon monoxide exposure, drive-through is a feasible and effective alternative to traditional walk-in clinics by providing faster vaccination under lower disease transmission risks [14
]. Drive-through clinics have been used for testing and provision of health services such as prenatal and pharmacy services during the COVID-19 pandemic as well [16
]. The use of drive-through during the COVID-19 pandemic demonstrated its effectiveness in the absence of enough clinical testing facilities [20
Drive-through mass vaccination facilities can be designed and implemented in various shapes and sizes [7
]. Drive-through facilities with more dispensing lanes can provide higher throughputs and prevent traffic overflow onto neighboring streets [10
]. Drive-through facilities need to be staffed with different skills including immunization, nursing, admin, IT, logistics, and security. Drive-through clinics can be set to work 24 h a day in two, three, or four shifts. Although drive-throughs have lower disease transmission risks, especially when staff and visitors use personal protective equipment and are vaccinated before vaccinating others, attention must be paid to the safety and the security issues related to traffic, drivers’ behaviors, and extreme weather conditions. Hence, large drive-through clinics may need to be supported by local emergency services such as police, fire, and paramedics [18
Several studies have modeled and examined mass vaccination and point of dispensing clinics using discrete event, agent-based, optimization modeling, and multi-criteria decision support systems for better layout, efficiency, resource allocation, and scheduling [2
]. Some of these models have been turned into software packages and tools. Examples of such tools included Clinic Planning Model Generator developed by Aaby et al. [22
], Maxi-Vac by the US Centers for Disease Control, and Real Opt by Lee et al. [10
]. However, new models and tools are needed to incorporate the specific challenges posed by the COVID-19 pandemic and emerging technological solutions in simulations and artificial intelligence. This paper aims to use the advancements in simulation tools and artificial intelligence to develop an application that can help mass vaccination planners to quickly assess potential outputs of different drive-through mass vaccination clinics.
We utilized an 80–20 train-test split to train a 5-layer feed forward neural network, which is shown in Figure 4
. We trained the network using the Adam optimizer with a learning rate of 0.001, minimizing the mean absolute error, in batches of 256. Training was stopped after 100 epochs, with the test and validation error displayed in Figure 5
Our final network is 10 Mb in size, which is small enough to be implemented even on the smallest of mobile devices and is able to infer results instantly. More precisely, by entering the parameters (within reasonable ranges) listed in Table 1
, our model is able to predict information such as cars and passengers passing through the vaccination center, average wait times throughout the day and overall completion metrics. An example of this can be seen in Figure 6
Our neural network model is capable of prediction orders of magnitudes faster than the full simulation model. More precisely, predictions were on average computed in 0.027 s, with minimum and maximum times of 0.025 and 0.039 s, respectively, in a simulation of 1000 random input variations. Distribution of the results can be seen in Figure 7
. Notice that the time scale we achieve here is on the order of milliseconds, whereas previously we required minutes to produce these same results. We achieve an improvement of over 3000× in terms of speed, largely due to the computation cost of simulating the entire event. Furthermore, we do all this locally as opposed to the need for cloud computation.
The drive-through model developed in this study is a meta-model (or models of models) that as described by Obsie et al. [33
], represents a deterministic proxy for stochastic simulation models. However, developing this type of meta-models requires running complex simulation for many combinations and then using the data generated by each run to create a simpler machine learning model that is a reasonable approximation of the initial simulation model. The meta-models can then provide quick predictions for different input values. While simulation-based or meta-modeling is not a new concept [34
], its application in different fields and simulation types are still limited. In this study, we applied meta-modeling approach in the context of mass vaccination.
The AI-based drive-through model developed here has been deployed as a web-application that can be used by potential users (www.adersim.org
). The approach serves our purpose of providing a general decision supporting tool for planning a large-scale drive-through vaccination. A general decision unit, such as a municipal government or a public health region, may simply input its unique parameters, and our model will generate accurate prediction of waiting time, capacity, etc., very quickly. Since our AI model is much faster than the simulation model, municipal governments or public health regions may compare multiple decisions, regarding the number of lanes, staffing levels, etc., and make the best choice. Further, our AI model is as small as 10 Mb, which can be implemented on any mobile device as well.
Our drive-through simulation has been developed and parameterized based on the past and current best practices available for drive-through design, settings, and operations. However, because the details of a potential SARS-Cov-2 vaccination protocols are not available yet, our drive-through mass vaccination simulation tool can be further tuned as new information becomes available. For example, the vaccination and recovery times may be very much influenced by the specific immunization protocols for SARS-Cov-2 vaccine, which may be different from previous vaccination. In the absence of real drive-through cases, such simulations can play a significant role in pre-planning.
Another potential application of our method in the mass immunization case would be an extension of this approach for other mass vaccination clinics such as large temporary walk-in clinics that is currently under development by the authors. In future, we would like to further refine our neural network to further reduce its complexity and enhance its accuracy. More importantly, it will be interesting to see whether the same network will serve more general cases as they arise in future.