Artificial intelligence could help ‘fine-tune’ vaccine priority lists, predict mortality, study reports

The CDC has recommended a phased approach to vaccine distribution, which states and communities can modify as they see fit.AHMAD GHARABLI/AFP via Getty Images

Much of the debate around vaccine prioritization hinges on one question: Who faces the greatest risk of dying if they become infected with COVID-19? Thus far, it is a question without a definitive answer.

Age is one way to gauge risk, with the Centers for Disease Control and Prevention recommending that people aged 75 and older be among the first members of the general public to have access to the vaccine. But in the next phase of distribution, as the CDC tries to factor in underlying medical conditions, the calculation becomes much more complex.

Artificial intelligence, when applied to standard patient medical records, can help untangle that web, a new study by Massachusetts General Hospital and Harvard researchers found.

Using only information known before a patient’s COVID-19 infection — diagnosed health problems, medications, and basic demographic information — researchers were able to identify factors that predict a heightened risk of COVID-related death. Age emerged as the most important predictor, immediately followed by a history of pneumonia, a condition not currently listed in the CDC’s prioritization plan.

“If we can predict [mortality] so well, based off of all these features that happen before individuals even get sick, this can really be applied in ways that I think are novel for an algorithm like this,” said Dr. Zachary Strasser, one of the study’s lead researchers, along with Hossein Estiri, an assistant professor of medicine at MGH and Harvard. “We can really think about who needs to get prioritized for limited resources, because these are the people that are probably going to do worse.”

The CDC has recommended a phased approach to vaccine distribution, which states and communities can modify as they see fit. Health care workers, long-term-care facility residents, front-line essential workers, and people 75 and older were first in line.

From there, the CDC recommends vaccinating all people over 65 and younger people with underlying medical conditions known to increase risk of serious COVID-19 infection or death. While the CDC maintains a list of such conditions, its website emphasizes that it is incomplete and subject to change.

Strasser and Estiri say their research could help refine those recommendations.

The study, published Thursday in Nature, is based on analysis of electronic health records for 16,709 COVID-19 patients — data shared securely through Mass General Brigham’s Center for COVID Innovation.

Using artificial intelligence, the team narrowed tens of thousands of potential variables to 46 clinical conditions and a handful of demographic features likely to impact mortality. Of these, 18 features were consistently associated with increased risk of mortality.

Age proved to be most significant.

But when zeroing in on people of similar ages, the importance of age within a given group varied.

Mortality risk varied fairly widely for people ages 65 to 85. But a 45-year-old and 65-year-old, for example, would not see as much variation in their risk based on age alone. Age had an extremely limited effect for the oldest group, those aged 85 and older.

The second-most important factor, a history of pneumonia, was “shocking” to researchers, Strasser said. Since pneumonia is not usually classified as a chronic condition, it is rarely considered in epidemiological models, making this one of the first studies to tie past pneumonia infection to COVID-19 mortality.

Other important medical conditions were type 2 diabetes with complications, heart failure, and chronic kidney disease.

Another interesting finding was that risk associated with these conditions varied by age.

Diabetes complications were most important for those aged 45 to 65. Pneumonia was a stronger predictor of risk for 65- to 85-year-olds. Kidney disease and hypertensive emergencies increased the odds of mortality for those older than 85.

The study repeated findings that women have better odds of beating COVID-19 than men. Race and ethnicity, once other variables were accounted for, did not rank among the 18 variables most closely associated with risk.

The study and its practical applications have limitations, its researchers acknowledge and other experts said.

The model’s integrity depends in part on the accuracy of patient medical records and diagnoses. And, given privacy and ethics concerns, it would be difficult to use a predictive model of this kind to determine the exact order in which individuals should receive vaccines.

“It’s just more evidence that the CDC could take to fine-tune their guidance about priority groups,” said Dr. David Hamer, an infectious disease expert at Boston Medical Center and Boston University who was not involved in the study.

Hamer said the study’s key findings should be used to update an understanding of COVID-19 mortality — in particular, the risk associated with a history of pneumonia and the suggestion that race and ethnicity on their own are not significant factors, once enough health conditions and other variables are taken into account.

Hamer said the model did not identify some known risk factors, including weight and body mass index. Still, he said, the study demonstrates that artificial intelligence can help identify new high-risk comorbidities, especially in cases where a condition is rare or research is limited, and predictive models might be the best tools.

Researchers said their model could have applications beyond vaccination planning.

Though built using complex computational methods, the model is easily adaptable, they said, and produces results quickly.

“We are facing a virus that is changing. It’s mutating. ... As the new mutations comes out, we can update our models. We can study long-haulers,” Estiri said, referring to those who recover from the initial symptoms of COVID-19 but face some long-term problems. “We’ve shown that we can quickly leverage this data to predict mortality.”