NJIT Machine Learning Expert Pan Xu Combats COVID Vaccine Inequity

An NJIT computer scientist studied COVID vaccine data from Minnesota to design equitable methods of distributing vital resources during any widespread emergency.

The resulting algorithm showed that giving everyone equal access to vital resources isn't necessarily the best approach, depending on the methods and desires of emergency authorities, explained Pan Xu, assistant professor in Ying Wu College of Computing.

For example, if an emergency disproportionately impacted people of certain ages, ethnicities, incomes, locations or races, then those communities should get greater percentages of whichever resources were needed to help them, Xu noted. That could mean withholding the resources from other groups who are better empowered to withstand delays.

How can we say a distribution strategy is fair or equitable?

"A key question that remains is, how do we quantify the equity? How can we say a distribution strategy is fair or equitable?," Xu observed.

Human instinct is to give everyone the same access, but Xu noted a February 2021 Washington Post report stating that, "White Americans are being vaccinated at rates of up to three times higher than Black Americans, though Black Americans have suffered a much higher death rate from COVID-19 than White Americans." He also cited a New York Times report from the same timeframe stating that, "In New Jersey, about 48 percent of vaccine recipients were White, and only 3 percent were Black, even though about 15 percent of the state's population is Black."

But the research could apply to many kinds of emergencies, such as natural disasters or wars, where perhaps the resource might be protective equipment, water or anything else. In the near term, Xu is presenting his work at the 2022 Association for the Advancement of Artificial Intelligence conference, which is virtual this year because of the pandemic. He is also planning to brief New Jersey Department of Health officials about his findings in case the pandemic heats up again.

"Essentially, we try to utilize tools from stochastic optimization to exploit results from machine learning [such as] the demographic distribution among arriving agents and how disproportional each group is represented in the arriving population compared with that in the general population," Xu said. "Techniques of machine learning give us the potential biases in the representations in the arriving population, while our algorithms aim to proactively adjust the allocation policy to combat potential biases in the arriving population such that the final vaccination rate is as close to preset targets as possible."

"There are two issues regarding how to define an equitable policy," he added. "The first is setting a target ratio we aim to achieve (i.e., target vaccination rate). Some will say we should put the target ratio of each group [as its proportion] in the general population, while others argue we should set the target ratio of each group as the infection rate of that group. The second issue is quantifying the gap between what we achieve and the target."

Xu's research partner was Yifan Xu of Southeast University, Nanjing, China.