This blog post is part of the series on AI and vaccination. Also see: How AI Can Help with Vaccination AI for Managing Vaccine Supply Chain Covid Vaccine needs Blockchain Transparency; U.K. Sees the Light
What population to vaccinate first? In what order? These are simple questions with wickedly hard answers when it comes to scale and to the enormous number of decision constraints combined with the fragmented, incomplete data. AI is helping, because it can reduce the uncertainty and perform human tasks at scale. But it’s the people who make the decisions. (People also create algorithms.)
Vaccination Approaches Vary Globally and Locally. AI Helps Everywhere.
Indonesia with the population of 270 million people spread across 17,000 islands vaccinates the most mobile population, ages from 18 to 59, to help the country’s hard-hit economy. AI supports decisions.
Russia vaccinates people within the age bracket of those who participated in clinical trials of its vaccine. Israel vaccinates the entire country.
In the US, CDC recommends that healthcare personnel and residents of long-term care facilities should be offered the first doses of COVID-19 vaccines. The US recommendations aim to decrease death and serious disease, preserve functioning of society and reduce the extra burden of COVID-19 on people already facing disparities. Each US state and local government has its own vaccination plan.
This is all complicated by the moving target of reaching herd immunity quickly, without clarity in precise numbers. In the US, the government aims at 70-80% of the population. The spread of researchers’ opinions is between 50% and 90%. WHO says, “Until we better understand COVID-19 immunity, it will not be possible to know how much of a population is immune and how long that immunity last for, let alone make future predictions.” In other words, with more data, we will see adjustments. This is totally normal in AI, but might not look normal in the public eyes.
It’s Hard to Predict The Speed of Vaccination, but It Will Be Easier Over Time
AI predictions are complicated by the availability and quality of supplies. Vials, syringes, alcohol are all data points. Imagine, all the necessary vaccines and supplies overcame the weather, cyber threats, and even natural disasters – and there is no one to give an injection. AI staffing and capacity simulation models for first responders and healthcare providers help reduce (but not eliminate) uncertainty. Organizations are guessing when they might get back to business as usual.
Vaccination Constraints and Unknowns Create Enormous Complexity
AI optimizes decision support by considering the enormous complexity and interplay of the goals, supplies, demographics, economics and so much more. The AI’s immediate help is in solving two major problems:
1) Efficiency and fairness in vaccinating those who are interested and
2) Identifying and contacting those who won’t engage because they are uniformed, hesitant or have some reasons to forgo the vaccination.
Propensity models for vulnerability are not new, they were quickly repurposed to prioritize vaccine distribution. There is still very little data. The data on entry and exit points will accelerate clarity and confidence in predicting the effect of the vaccine, and consequently, more accurately determine the herd immunity threshold. Sharing the data in healthcare has always been a problem because of the strict privacy regulations. Fortunately, the regulators let healthcare organizations use federated learning, so they can share models, but keep the data private.
Another imprecise question is who are the essential workers? How to identify the vulnerable ones among them? For example, the people who are paid hourly and cannot afford to miss work while getting the vaccine? AI can help, but it cannot and should not design the equitable vaccine distribution.
What are the ways to reach the right population? Even an old data quality problem of the contacts and addresses (very well known to the marketers) is an issue. Data labeling has been a issue. AI capabilities for inferring the data and active learning are in play here – not fully fixing, but improving the data. Combining the data at a state or federal level, let alone globally, is still unsurmountable.
Celebrate Successes, They Are Hard-Earned
Media thrives on sensations. It amplifies disparate precedents, while glazes over hard earned successes. For instance, multiple publications kept repeating the same news about the algorithm mistake that left out frontline doctors in one hospital. I perfectly understand the anger and frustration of these people. But I also want to defend data scientists across the globe who rushed to help, worked without breaks and often pro bono since March of 2020. Some of them worked side by side with the medical personnel, witnessed death and tough dilemmas. To deal with unimaginable complexity, they were stitching data piece by piece, inventing creative approaches and repurposing the existing ones.
I know many doctors, nurses and essential workers for whom vaccination is already done. It was smooth and showed the light in the end of the tunnel. Celebrate successes! Be grateful to those who made them happen.
Follow Svetlana on Twitter @Sve_Sic