VA Piloting AI to Predict Mortality Rates of COVID-19 Patients

Andrii Vodolazhskyi/Shutterstock

The tool has been in development since the spring and is now being piloted ahead of a larger rollout, VA officials said.

As the COVID-19 pandemic continues to wreak havoc across the U.S., the Veterans Affairs Department is piloting a new artificial intelligence tool to quickly predict a patient’s prognosis and recommend next steps.

During tense medical crises, as hospital beds fill up and minutes and seconds mean the difference between life and death, doctors, nurses and other care providers need to assess new patients quickly to determine the right course of action with the resources available.

“The problem is, given a positive detection of COVID viral test—if someone has a positive test—what is the prognosis? Will the patient need hospitalization? Are they at risk of death, etc.? What will be their needs?” Gil Alterovitz, director of AI at the VA and with the National Artificial Intelligence Institute there, said Thursday at the Genius Machines 2020 Virtual Summit hosted by Nextgov and Defense One. “The approach was to develop an applied AI model for that using both clinical and nonclinical information.”

VA officials started working on the tool shortly after the outbreak in early spring. Working with clinicians at the Washington DC VA Medical Center, the team pulled data from more than 11,000 patients from across the VA to analyze how more than 30 features—such as heart rate, blood pressure, preexisting health issues, etc.—affected recovery and mortality rates.

A VA official offered an example of how it works: A patient with a preexisting kidney disorder comes to a VA clinic and is diagnosed with COVID-19. A standard prognosis might give that person a high mortality rating due to the underlying conditions. However, the algorithm gave that person a low mortality risk score, discovering other factors that showed the patient was likely to recover.

Alterovitz noted the tool features explainable AI so users can deconstruct how it arrived at its conclusions, helping clinicians to trust the results and spot recurring trends in this novel outbreak.

The tool is still in the pilot phase and has yet to be used to treat a patient directly, according to the VA. If the pilot is successful, the tool could be rolled out across the department quickly, as it has already been coded into a dashboard that can be accessed with a link or added to existing diagnostic tools as a module.

“It’s being used already operationally,” Alterovitz said. “And we’re going to be now working with other medical sites to see how this can be used across the country in different settings, as well.”