Artificial intelligence can distinguish overlapping inflammatory conditions with total accuracy, according toat the annual meeting of the American College of Rheumatology.
Texas pediatricians faced a conundrum during the pandemic. Endemic typhus, a flea-borne tropical infection common to the region, is nearly indistinguishable from multisystem inflammatory syndrome in children (MIS-C), a rare condition set in motion by SARS-CoV-2 infection. Children with either ailment had seemingly identical symptoms: fever, rash, gastrointestinal issues, and in need of swift treatment. A diagnosis of endemic typhus can take 4-6 days to confirm.
Tiphanie Vogel, MD, PhD, a pediatric rheumatologist at Texas Children’s Hospital, Houston, and colleagues sought to create a tool to hasten diagnosis and, ideally, treatment. To do so, they incorporated machine learning and clinical factors available within the first 6 hours of the onset of symptoms.
The team analyzed 49 demographic, clinical, and laboratory measures from the medical records of 133 children with MIS-C and 87 with endemic typhus. Using deep learning, they narrowed the model to 30 essential features that became the backbone of AI-MET, a two-phase clinical-decision support system.
Phase 1 uses 17 clinical factors and can be performed on paper. If a patient’s score in phase 1 is not determinative, clinicians proceed to phase 2, which uses an additional 13 weighted factors and machine learning.
In testing, the two-part tool classified each of the 220 test patients perfectly. And it diagnosed a second group of 111 patients with MIS-C with 99% (110/111) accuracy.
Of note, “that first step classifies [a patient] correctly half of the time,” Dr. Vogel said, so the second, AI phase of the tool was necessary for only half of cases. Dr. Vogel said that’s a good sign; it means that the tool is useful in settings where AI may not always be feasible, like in a busy ED.
Melissa Mizesko, MD, a pediatric rheumatologist at Driscoll Children’s Hospital in Corpus Christi, Tex., said that the new tool could help clinicians streamline care. When cases of MIS-C peaked in Texas, clinicians often would start sick children onand treat for MIS-C at the same time, then wait to see whether the antibiotic brought the fever down.
“This [new tool] is helpful if you live in a part of the country that has typhus,” said Jane Burns, MD, director of theResearch Center at the University of California, San Diego, who helped develop a similar AI-based tool to distinguish MIS-C from Kawasaki disease. But she encouraged the researchers to expand their testing to include other conditions. Although the AI model Dr. Vogel’s group developed can pinpoint MIS-C or endemic typhus, what if a child has neither condition? “It’s not often you’re dealing with a diagnosis between just two specific diseases,” Dr. Burns said.
Dr. Vogel is also interested in making AI-MET more efficient. “This go-round we prioritized perfect accuracy,” she said. But 30 clinical factors, with 17 of them recorded and calculated by hand, is a lot. “Could we still get this to be very accurate, maybe not perfect, with less inputs?”
In addition to refining AI-MET, which Texas Children’s eventually hopes to make available to other institutions, Dr. Vogel and associates are also considering other use cases for AI. Lupus is one option. “Maybe with machine learning we could identify clues at diagnosis that would help recommend targeted treatment,” she said
Dr. Vogel disclosed potential conflicts of interest with Moderna, Novartis, Pfizer, and SOBI. Dr. Burns and Dr. Mizesko disclosed no relevant conflicts of interest.
A version of this article appeared on.