By Alan Mozes
MONDAY, March 30, 2020 (HealthDay News) -- An international team has designed a computer program that predicts with up to 80% accuracy which COVID-19 patients will develop serious respiratory disease.
Developed by U.S. and Chinese researchers, the artificial intelligence (AI) program has been tested at two hospitals in China with 53 patients who were diagnosed in January with COVID-19. The new tool is considered experimental and is now in testing.
The aim is to help doctors make the best use of limited resources, by identifying early on which patients will likely need hospital beds and which can be sent home for self-care. In theory, it could also help direct administration of aggressive treatment even in the initial absence of severe symptoms.
"Of those who have symptoms, 80% -- maybe up to 85% -- will have mild disease; around 15% to 17% will have severe disease and need to be hospitalized; and a further 3% to 5% will need intensive care, usually due to Acute Respiratory Distress Syndrome [ARDS]," said study co-author Dr. Megan Coffee.
She's a clinical assistant professor of infectious disease and immunology at NYU Grossman School of Medicine in New York City.
ARDS is a potentially deadly condition in which fluid leaks into the lungs, making breathing increasingly difficult. Coffee said at least two-thirds of COVID-19 patients who go on to need treatment in a hospital intensive care unit develop ARDS, which is the "underlying process leading to death in many of the cases."
But Coffee noted that COVID-19 starts mildly in everyone, with a cough, fever and upset stomach.
"A small percentage will go on, five to 10 days later, to develop very severe disease and some will require intubation. It's not always clear who," she said. "Sometimes someone in their 30s with no medical history has more severe disease than someone in their 70s with multiple medical problems."
So the goal, Coffee said, was to develop an artificial intelligence version of a "master clinician" -- meaning a very experienced doctor dealing with a well-known disease.
Working with researchers at two hospitals in Wenzhou, China, the NYU team devised a computer model based on what co-author Anasse Bari described as the kind of "predictive analytics" used to forecast stock market activity and voting patterns. Bari is a clinical assistant professor of computer science at NYU's Courant Institute.
They fed it relevant patient information, such as results of lung scans and blood tests, muscle ache and fever patterns, immune responses, age and gender.
To their surprise, researchers found that the factors most clinicians would likely focus on -- such as lung status, age and gender -- were not helpful in predicting outcomes.
So what was?
The most accurate predictors were slight elevations in a liver enzyme called alanine aminotransferase (ALT); deep muscle aches; and higher levels of hemoglobin, the protein that facilitates blood transport of oxygen throughout the body.
"That's the value of this approach," Coffee said, "to look for what we, as clinicians, might not notice."
While the program needs to be validated on larger populations, she said it would be easy to roll out if future testing finds similar accuracy.
The tool could prove "very useful," said Dr. Maria Luisa Alcaide, a fellow with the Infectious Diseases Society of America who reviewed the findings.
"What's happening with COVID-19 is that cases have significantly increased to the point where some hospitals' ICUs are overwhelmed," she noted. "And for reasons that are not well-understood, not everyone who gets very sick fits the profile of an older person with underlying conditions."
The better doctors are able to predict those who will, the more carefully they could track them and their care, said Alcaide, who is also an associate professor of infectious diseases at the University of Miami.
But the method has been tested only in a very small sample of patients, and hundreds of thousands are now infected, she pointed out.
"It's unlikely that this small sample is representative of all COVID-19 patients," Alcaide said. Some of these predictors may turn out to be important. But we just don't know yet. Other markers may turn out to be more important. So this really needs to be validated with more patients."
The findings are reported online in the March 30 issue of the journal Computers, Materials & Continua.