A University of Minnesota computer algorithm is so accurate at identifying COVID-19 infections, just by examining chest X-rays, that it is being made available to 450 health systems worldwide.
U researchers aren’t sure what the algorithm detects in X-rays that distinguishes patients with COVID-19, but after testing it on thousands of images, they know it works.
“The COVID positive X-rays really sort of isolated themselves from the COVID negative X-rays,” said Dr. Christopher Tignanelli, an assistant professor at the U medical school and a critical care surgeon.
The algorithm was developed into a clinical tool by M Health Fairview, the partnership between the U and Fairview Health Services, and Epic, the Wisconsin-based provider of electronic health records. Doctors with M Health Fairview are being trained on how to use the results to guide patient care, and the tool will soon be offered for free to other hospitals with Epic record-keeping systems.
While diagnostic testing for COVID-19 is broadly available, the creators of the algorithm said there are numerous ways it could help amid the pandemic, which has caused 100,200 lab-confirmed infections and 2,049 deaths in Minnesota alone. That includes 13 COVID-19 deaths and 1,066 infections reported Thursday by the Minnesota Department of Health.
Many patients admitted to hospitals for reasons other than COVID-19 receive diagnostic tests right away, but then receive multiple chest X-rays that could identify infections that emerge later on. Diagnostic tests have varying degrees of accuracy, making the analysis of chest X-rays — which hospitalized patients with respiratory symptoms often receive anyway — a handy double-check.
X-ray analysis could be a useful backstop if supply problems leave communities short of tests, said Dr. Genevieve Melton-Meaux, M Health Fairview’s chief analytics and care innovation officer.
In addition to the tool being accurate, she said “it is definitely telling us there are changes in the body as a result of COVID in the lungs.”
The innovation was announced amid signs that the pandemic is worsening in Minnesota. The positivity rate of diagnostic testing is back above 5%, a threshold that suggests to state health officials a rising spread of the coronavirus that causes COVID-19.
The state also reported that at least 356 Minnesotans were admitted to hospitals for COVID-19 in the seven-day period ending Sunday. That is the highest seven-day total since June 1.
It also comes amid increased scrutiny of the reliance on PCR molecular diagnostic tests, which provide qualitative yes/no answers about the presence of the virus in people’s nasal or throat swab samples.
A letter in the New England Journal of Medicine this week called for a switch to newly developed rapid antigen testing, which isn’t as sensitive as PCR testing but might catch more people when they are at infectious stages of COVID-19 illness.
M Health Fairview leaders stressed that the X-ray algorithm isn’t envisioned to replace diagnostic tests.
U researchers this winter were preparing to test and create an algorithm through machine learning that would automatically detect rib fractures.
When the pandemic reached Minnesota in March, their research pivoted to the potential use of this technology for COVID-19.
The algorithm was trained this spring by analyzing 67,000 chest X-rays from people known not to have COVID-19, and 17,000 X-rays from people who had tested positive. The comparison included visually similar X-rays of people with influenza, pneumonia or other lung diseases. Then it was applied real time in July to all chest X-rays in the M Health Fairview system.
The result is that it is 80% accurate at finding COVID-19 infections and 62% accurate at weeding out unrelated health issues.
“We are characterizing [the result] as high likelihood of negative and high likelihood of positive,” Melton-Meaux said.
Creators offered several reasons why it was less accurate at assessing the likelihood of people not having COVID-19. Some people were scanned early in their infections, when X-rays weren’t as conclusive, and only later developed symptoms and tested positive, Tignanelli said.
While the algorithm learned from a variety of X-ray images, including from people sick this winter with influenza and pneumonia, Tignanelli said it will be important to monitor its progress and accuracy amid the coming flu season.
“It’s technically built into the training of the algorithm, but only time will tell,” he said.
M Health Fairview officials said this is one of the first applications of this type of artificial intelligence research in a “decision support” tool to guide doctors on patient care.
The algorithm was developed with a “black box” AI method, meaning that researchers fed data into the analysis, but then had no immediate feedback on how it generated its results.
Tignanelli said research is ongoing to examine the process and understand the visual qualities on X-rays that identify COVID-19 infections.
“We teach an algorithm how to run something,” he said. “Now we ask the algorithm how to teach back to us.”