Computer-Aided Identification of Lyme Disease Rashes

Published in Insight - Insight Sept/Oct 2019

Every year, 300,000 people in the United States are infected with Lyme disease, a potentially serious illness caused by the bite of a deer tick or western blacklegged tick carrying Borrelia burgdorferi bacteria. If left untreated early on, the disease’s side effects can progress from fevers and headaches to facial palsy, irregular heartbeats, cognitive difficulties and arthritis. 

Because the blood test for diagnosing Lyme disease isn’t always accurate during the earliest stages of infection, doctors also look for a rash that occurs in roughly 80% of infected patients. While the most well-known Lyme disease rash is shaped like a bull’s-eye, most resemble an unremarkable reddish-blue circular or oval blotch.

“The classic reaction is thinking it’s a spider bite because it doesn’t look like a Target department store logo,” says infectious disease specialist John Aucott, director of the Johns Hopkins Lyme Disease Research Center. “The rash actually goes away without treatment — but the infection doesn’t, and more than half of those people have later manifestations that are even harder to treat.” 

Along with researchers from the Johns Hopkins University Applied Physics Laboratory, including principal scientist Philippe Burlina, Aucott is helping develop an algorithm using deep learning that can analyze photos of rashes to make preliminary Lyme disease diagnoses. Over the next few years, the team plans to train an artificial intelligence system that could be the basis of an app to help people determine whether they should see a doctor about a tick bite, and software that will work with electronic medical record systems to help physicians better diagnose Lyme disease rashes. In both cases, users will simply upload a photograph of a rash for the algorithm to make its prediagnosis, prior to confirmation by a physician.

A $100,000 Johns Hopkins University Applied Physics Laboratory research and development award funded the yearlong pilot study, during which the team uploaded 2,000 rash photos from patients and the internet. Doctors then curated the photos to develop deep learning models to distinguish Lyme disease rashes from other types of rashes. Aucott says the algorithm currently has an 87% accuracy rate for diagnosing the rashes. 

The Lyme Disease Research Foundation is funding the project’s next phase, during which 50,000 more photos — including those of over a dozen types of skin disease that could serve as confusers — will be uploaded to improve the algorithm.