Imagine a formula that could score each American’s unique risk of dying of COVID-19. People’s odds would determine their exact number in line for a vaccine.
The algorithm would take into account your age, your race, your full medical history and every one of your health insurance claims. It would marry that information with data about vaccine inventories and health care locations. You’d get an email, a text, or a phone call the week before your vaccine appointment telling you where and when to show up. If you turned down the shot, the next in line would take your spot.
The pandemic has brought such an approach far closer to reality than many might guess. Hospital groups in California, Boston, St. Louis and the Upper Midwest used medical records to score members’ risk of death when choosing who got first priority for shots, vaccine information, treatment or extra support.
It’s hard to know whether Americans would embrace a vaccine algorithm that tapped into some of their most personal information. Or whether people would accept a formula’s determination of who was at most risk of dying, blind to other values such as keeping teachers from missing work due to illness.
But if these objections could be overcome, some experts say it’s possible that the U.S. could save tens of thousands of lives during the next pandemic with a widespread system of vaccine microtargeting.
“We do have the data, we do have the computational capacity,” said Hossein Estiri, an assistant professor of medicine at Mass General and at Harvard who has worked on risk-based vaccine modeling. “It’s just that we haven’t figured out the politics to make this happen.”
Niall Brennan, former chief data officer for the Centers for Medicare and Medicaid Services, said more than a dozen states already possess the information they’d need to measure how much a vaccine would benefit each resident. He believes it can be done securely.
“It doesn't even have to be in one data warehouse or data lake,” said Brennan, who is now president and CEO of the Health Care Cost Institute. “It just has to be available in some systematic way. It definitely could have helped with vaccine targeting.”
More than 259,000 Americans have died of COVID-19 since the first Pfizer shot was administered here on Dec. 14. That accounts for nearly half of all Americans who have died in the pandemic.
A vaccine rollout based on better health data could have changed the outcome, according to an analysis by Nilanjan Chatterjee and Jin Jin, researchers at Johns Hopkins University’s Bloomberg School of Public Health.
“Our model showed that if you prioritized by actual risk you would have saved more lives,” Chatterjee said.
What’s certain is that the country isn’t there yet. Even Chatterjee acknowledges that the introduction of fine-grained priority groups based on vulnerability would have complicated the current vaccine rollout and likely slowed it down.
Craig Brammer, formerly with the Office of the National Coordinator for Health Information Technology, said microtargeting across the board wouldn’t have worked this time around.
“There has been some truly groundbreaking data science work, but it’s generally in local pockets where multiple data sources are aggregated into a sufficiently rich data set for precise modeling,” said Brammer, now the CEO of The Health Collaborative. “It wouldn’t have scaled. And even if it had, the pandemic would have been over by the time it rolled out.”
Back in October, when the federal government was deciding who would get the first doses of rapidly developed new vaccines, officials contemplated using risk scores for various segments of the population. The Advisory Committee for Immunization Practices of the Centers for Disease Control and Prevention went so far as to model how many deaths would be averted by prioritizing high-risk groups like seniors.
CDC epidemiologist Matthew Biggerstaff said in a presentation Oct. 30 that vaccinating all people age 65 and older right after health would save 2% to 11% more lives compared with prioritizing essential workers or adults with other illnesses.
Yet the committee ruled out that idea. It opted for a compromise accepting fewer lives saved in exchange for steps to keep society up and running. It established four phases of eligibility that mixed in highly vulnerable groups like seniors and less vulnerable people in professions like hospital staff and police officers.
The former is a data-driven choice; the latter a moral calculation, ethicists say.
Beyond a handful of state and local governments, almost no one attempted anything close to the kind of microtargeting that technology could enable. Rather than offering the vaccine by phone to a 67-year-old Latino man with diabetes, some states invited anyone fitting the description “obese” to show up in a CVS parking lot.
Guidelines for which underlying health conditions qualify vary from state to state, resulting in widespread confusion. Last month in Missouri, for instance, you could sign up if you had Type 2 diabetes but not if you had Type 1.
Today, more than a dozen states have opened eligibility to all adults. But that doesn’t guarantee an appointment. Vaccine sign-up systems have crashed in Massachusetts, Florida, Tennessee, and Washington, D.C., fueling frustrations.
“We could have relied more on risk stratification,” said Chatterjee, the Johns Hopkins researcher, describing a system where specific groups of people were scored by their vulnerability to severe illness or death. “All of this information on age, preexisting conditions, demographics and socioeconomic patterns could have been incorporated in our vaccine distribution planning at an earlier stage.”
Chatterjee recently published a study that assigned COVID-19 risk scores for the general U.S. population and the elderly Medicare population incorporating information on age, race, ethnicity, gender, area of residence and preexisting conditions. The study factored in findings by other researchers that, for example, obesity or uncontrolled diabetes raises your risk of coronavirus death twofold and elderly 25-fold compared with younger age groups. It examined how these combined factors stack up in the population.
At the request of USA TODAY, Chatterjee extended his analysis to show what would have happened if these scores had been used to guide vaccine distribution.
He took the weekly vaccination rates state by state since January and assumed that they had been distributed to more vulnerable groups.
In the end, Chatterjee concluded that at least 27,000 fewer people would have died since January. The analysis looked only at lives saved among people 65 and older; other vulnerable groups also would benefit.
“If you’re both old and have a preexisting condition, the risk gets multiplied,” Chatterjee said. “Those people should be prioritized over those who are just old and don’t have a preexisting condition. A model like ours that says that could have been helpful early on.”
Algorithms in the wild
Chatterjee’s solution is more than theoretical.
In California, health care data science company Cogitativo teamed up with health insurer Blue Shield of California to build an algorithm using data on its 4 million members. The company assigned each member a COVID-19 risk score used to pinpoint who should receive vaccine information first.
Jamie Chan, vice president of clinical quality at Blue Shield of California, told USA TODAY that Cogitativo’s model was “eye-opening.” The company used risk scores to target messaging and services to specific members, like homebound seniors with specific health conditions, and sent them food early in the pandemic. More recently, those high-risk members were sent vaccination appointment details.
Chan says it’s a pity this kind of algorithm and rich health insurance data aren’t being used at the state level.
“Unfortunately the state’s used crude levels of data because it’s easier,” she said. “I don't think that our government, nationally as well as at the state level, understands what information we have and what we can do to help guide appropriate decision making and care.”
Brennan, the former Medicare data officer, said “all-payer claims databases” kept by some states have the information needed to assess risk levels for individual residents.
“Theoretically, you could absolutely build a model that inputs all of the vaccine eligibility criteria and data from APCDs,” he said.
Large medical centers, like Mass General Brigham in Boston, are using the data they have on patients to build COVID-19 risk scoring algorithms as well. There, Estiri built an algorithm that uses clinical data on more than 16,000 patients to assign each their risk of dying of COVID-19.
“This technique could be extremely valuable for determining who is most likely to benefit when resources are limited, such as informing vaccination distribution,” he wrote in a study published in February.
Still, implementation even in his own hospital has been a tough sell, Estiri admits. The algorithm was approved only for research purposes, so to implement it at the emergency department would require a new review process and new data.
Other health systems have had more luck implementing COVID-19 risk algorithms. In December, BJC HealthCare in St. Louis started scoring patients on their risk of death from any cause, including COVID-19, to help guide care at the hospital.
Within 24 hours of a patient’s admission, doctors can access a patient’s mortality score through a widely used electronic health record system known as Epic, says Randi Foraker, director of the Center for Population Health Informatics and professor at Washington University School of Medicine in St. Louis, who designed the algorithm. Upon reviewing a patient’s medical chart in Epic, the doctor would see the mortality score and treat it accordingly.
“We used electronic health records and claims data including demographics, diagnoses, medications that have been prescribed and laboratory values,” said Foraker, describing the algorithm. The largest hurdle wasn’t implementing the technology: “We had the governance structure worked out, and we had the trust,” she said. The hard part was constantly updating the model with new data on COVID-19 risk factors like underlying conditions.
Foraker is in the process of measuring whether the algorithm produced better outcomes for patients than would have occurred without it.
At Sanford Health, a health care system with 46 hospitals in the Dakotas, Minnesota and Iowa, physicians and researchers built an algorithm that prioritizes incoming patients for either treatment or vaccination, based on their risk of getting severe COVID-19. The algorithm, built using health records from 100,000 patients, is baked into Epic and triggered when a patient tests positive for the virus.
“For every positive COVID test, we run them through a digital filter that looks for things like obesity, kidney disease, age and ethnic background to say, ‘This person is at a much higher risk of being admitted to the hospital,’” said Jeremy Cauwels, Sanford Health's chief physician, who helped develop the algorithm. “Then, within 48 hours, we call them up to say we want to bring you in” for an infusion of antibodies, which can help prevent severe COVID-19 illness by helping the body neutralize the virus.
The algorithm also prioritizes for vaccination of vulnerable populations like Native American diabetics, leading the health system to contact patients via phone and mail with appointment scheduling information.
Sanford estimated that its delivery of antibodies to more than 2,700 COVID-19 patients, a process that was guided by the algorithm, prevented 15 deaths and averted 80 hospital admissions.
Challenges to the formula
For all the benefits that microtargeting might offer, critics are skeptical that a prioritization algorithm would have worked nationally.
One barrier is how siloed and unequal America’s health care system is.
“The U.S. does not have a national medical system or a system of universal health care,” Courtney Campbell, a professor at Oregon State University who teaches medical ethics, said in an email. “At least 25 million Americans do not have health care insurance coverage. Access to health care, including access to vaccines, thereby often is a commodity stratified by ability to pay rather than a public or communal good allocation by equality of need and risk profiles.”
Had the country even attempted to roll out the vaccine with an algorithm relying “on clinical evidence and social determinants of health,” wrote Campbell, it likely would have contained “its own biases that discriminate against some and prioritize others.”
Surprisingly, at least one study has found vaccinating people at the highest risk of death first may not always be the best way to minimize the number of people who die. According to Yonatan Grad, a professor of immunology and infectious diseases at Harvard's T.H. Chan School of Public Health, immunizing healthy young people and front-line workers who routinely interact with others could, under certain circumstances, more rapidly tamp down the spread of COVID-19 and therefore take away the threat to higher-risk groups.
For Arthur Caplan, director of New York University Langone's Division of Medical Ethics, a purely data-driven allocation of vaccines goes against this country’s history and culture.
“You can come up with algorithms all day long,” said Caplan, but America is more concerned with making public health decisions based on moral judgments than what the data says.
America doesn’t consider the odds when deciding to keep a dying family member on a breathing machine, says Caplan, and it doesn’t consider the risk of death when deciding who should be first vaccinated against a fatal virus: “These decisions are not data-driven.”
To Kirk Nahra, a partner at WilmerHale and an expert in health privacy law, building a system that uniquely prioritized Americans by their individual risk scores would have relied on data that didn’t exist, taken too long and been seen as too controversial given the nation’s concerns about privacy and health care data. A system that smacks of Big Brother also might scare some people from getting vaccinated.
“I think the government wanted to do it in a way that was as noncontroversial as they could,” Nahra said. “There already are people unwilling to get the vaccine. If you add a privacy concern, there’s no chance that number goes down.”
And then there is the ease of administration.
"There is a lesson to be learned about simplicity over complication, especially when you have 50 different state captains calling the shots,” said Josh Michaud, associate director for global health policy at the Kaiser Family Foundation.
A complex risk scoring system might have slowed the COVID-19 vaccine rollout, Michaud said.
Amesh Adalja, a senior scholar at the Bloomberg School and an infectious disease doctor practicing in Pittsburgh, said narrowly prescribing the vaccine for individuals would have been impractical in 2021.
“It might work in an academic paper,” he said. But when telling doctors and nurses who’s next in line for a shot, “you don't want it to be overly onerous and overly specified.”
“In my view, the overarching goal is taking care of as many high-risk people as possible,” Adalja said, “without having any vaccines ending up in the trash.”
One final hurdle would need to be overcome before a singular algorithm could be applied to vaccine distribution nationally: We would have to agree on what we value most.
Does saving lives matter ahead of other goals? Or do we also care about ensuring certain people don’t get ill at all, even if they’re at low risk for dying, and even if moving them ahead in line causes more vulnerable people to die?
“There is a nontrivial question of what the overriding objective of a framework that might be used here is,” said Harald Schmidt, a professor at the University of Pennsylvania who studies medical ethics, in an email. “Different sequences are rational choices if your primary focus is averting deaths versus preventing spread of infection.
“In many cases, the sequences simply pose dilemmas, and there is no clear ‘right’ sequence that everyone would find acceptable.”
Threading the needle
Despite the reluctance to use it, the technology remains available to guide medical priorities far more precisely than officials have done since COVID-19 vaccines became available. More leaders private and public are showing interest in threading the needle between life-saving allocation formulas and decisions that are easier to explain to the general public.
President Joe Biden's administration is making strides in using data to target specific populations for COVID-19 vaccinations. It recently announced that the Vaccine Community Connectors program will enlist health insurers to use enrollee data and socioeconomic indicators of risk to find and “enable the vaccination of 2 million seniors age 65-plus in America’s most at-risk, vulnerable and underserved communities,” according to America’s Health Insurance Plans, which is organizing the effort.
David Allen, a spokesperson for the trade group, said insurers signing on to the pilot initiative include Anthem, Blue Shield of California, Cigna, Kaiser Permanente and United.
What might work in the next pandemic would be a proliferation of risk scoring algorithms rather than one master plan for the country.
Foraker at Washington University in St. Louis is working on a decision support algorithm for the cardiovascular health of cancer survivors across six electronic health record systems.
“It’s a huge lift scaling the system from one health system to another let alone a whole community of health systems,” she said.
When an algorithm is tuned to a precise population and geography, Foraker said, it makes clearing the data and policy hurdles a lot easier.
“Because the question is so focused, you can get clinical buy-in if you have a health care provider that really wants this tool implemented with their patients,” she said. “That’s half the battle.”
Shawn Murphy, a professor at Harvard Medical School and chief research information officer at Mass General Brigham, said stitching together networks of data sources is increasingly taking place.
“No place is going to have enough data on every single patient,” said Murphy, who worked with Estiri developing the hospital’s current COVID-19 risk algorithm.
But, Murphy said, health insurers, government programs like Medicare and major medical centers like his are increasingly pooling patient data. “That does, at the end of the day, give you a very good computational base” for algorithms to run on.
Like Foraker, Murphy said trust is key. A successful algorithm – like one that could be used to allocate vaccines based on one’s vulnerability to COVID-19 – needs to be clear and explainable to doctors and the public alike.
“It’s really critical for people to be able to understand a model,” Murphy said. “'Why are they getting a vaccine and I’m not?' If your answer is a super complicated model, that becomes a big political issue. But if you can say, ‘This person is over 85. and you’re 50, and that gives him a 10 times greater risk than you have,’ that’s really important.”
This article originally appeared on USA TODAY: COVID vaccine: What if medical records decided your place in line?