By Risto Miikkulainen, AVP of Evolutionary AI, Cognizant Technology Solutions & Professor of Computer Science, The University of Texas at Austin
Public intervention policy has been the main response in coping with the COVID-19 pandemic until now. To this end, the Pandemic Response Challenge, a global AI competition, was established to help develop intervention policies that best balance safety and economic impact. Even though the COVID-19 pandemic is now more punishing worldwide than it has ever been, in the last few weeks, hope has also emerged: Several vaccine trials have proven successful, limited vaccination programs have started, and the world is preparing for large-scale vaccination programs that could eventually end the pandemic.
This new situation makes the Pandemic Response Challenge more compelling than ever, for two reasons. First, it may take months, perhaps even years, to vaccinate the world’s population and eradicate COVID-19. Throughout that time, the world is vulnerable to the disease, and intervention plans will continue to be crucial in containing and mitigating its spread. Interestingly, vaccinations will make such plans more complex than before by creating a continuously changing situation: Plans will now need to be continuously adapted to the level and distribution of immunity in different regions. Such complexity is exactly what the XPRIZE competition was designed to address. Therefore, the techniques developed in the competition can be instrumental in learning effective intervention plans simultaneously with ongoing vaccination efforts as well.
Second, a different decision problem will soon emerge on managing the vaccination programs themselves. Vaccinating billions of people around the world, many for the first time, is a formidable logistical and public information challenge. It is, however, a decision problem that can be defined by variables that describe the context, actions that can be chosen, and outcomes that need to be optimized. The techniques developed in the competition can likely be utilized to cope with this challenge as well.
Regarding the first challenge, i.e. optimizing interventions with vaccinations, the XPRIZE competition is already set up to take into account vaccination policies to a limited extent. The dataset used for this competition, the Oxford COVID-19 Government Response Tracker, contains an intervention category H7, coding for ongoing vaccination efforts (i.e. the extent of population being vaccinated, such as key workers, clinically vulnerable groups, elderly groups, others). Remarkably, this data is available immediately as soon as any such actions are taken in any country---there is no history to be encoded into the dataset first.
How can the vaccine information be taken into account in the competition entries? Technically, the H7 information is background data, similar to obesity rates, demographics, weather, and other supplemental data. Note that although H7 is listed similarly to intervention actions in the Oxford data, it is actually not an intervention that can be prescribed at will (e.g., full vaccination of everyone may be desirable but not possible). Instead, H7 indicates a level of existing vaccination efforts, similar to other background variables. The predictor and prescriptor models can use it as additional input, and change their output accordingly. Note, however, that similarly to other background information, H7 cannot be updated once the predictor or prescriptor evaluation period has started in the competition---it only forms an informed starting point for the roll-outs.
While it is possible to be prepared for H7 in this way, the vaccination information is unlikely to affect the course of the competition significantly. The amount of data will be limited for several months -- only a few countries will be able to start their vaccination programs, vaccinate only a small subset of the population, and it will take several weeks for the vaccinations to become effective. The effect will probably first be seen in deaths and hospitalizations, and months later in the number of cases. By the time the competition concludes in February, vaccinations are just beginning to have a discernible effect.
However, there may be an opportunity to get an early understanding of such effects in some cases, and use the competition platform to demonstrate them. Since the competition allows specialization for particular regions, it may be possible to utilize vaccination information in a few selected regions to do better. For instance, UK started their vaccinations in early December; if their program progresses rapidly enough, data may exist on the first few stages by the time the prescriptor evaluation stage begins. At that point, it may be possible to retrain predictors for such regions with the latest H7 information before the end of Phase 2, and thus characterize and utilize their effect. Since H7 is likely to change within a few months, the evaluation horizon for such entries may be shortened. If the opportunity arises quickly enough, such optional evaluations may be possible, allowing us to learn more from the competition, and thus preparing for possible deployments after the competition.
Note that the Pandemic Response Challenge is intended to encourage technologies that are helpful in not just intervention plan management of COVID-19, but future decision problems as well. The motivation for the competition includes being prepared for future pandemics, coping with other natural and manmade disasters, and even global warming. However, the most immediate such new decision problem is the second challenge mentioned above---how should the vaccination programs be designed to be maximally effective, with minimal cost? Different countries have different resources to carry out vaccination programs, regarding manufacturing, distribution, storage, and administration of the vaccines. They have different demographics, mobility, and culture, requiring different targeting, delivery mechanisms, and public information campaigns. Available vaccines will differ in cost, availability, and ease of distribution, and different countries will be implementing vaccinations at different times. The prescriptions may include public information campaigns, vaccination requirements and checks associated with different activities. Provided that the competition is successful, the winning technologies could be adapted for this second challenge. The context, action, and outcome variables may be different, but the same idea of learning to predict the outcomes, and using the predictions to learn prescriptions, still applies.
The world was unprepared for COVID-19, scrambling to contain and mitigate it, and that effort is still ongoing. However, we can already learn from it, and due to efforts like the Pandemic Response Challenge, should soon have the tools to do well in the vaccination phase of this pandemic.