AI Helps Doctors Diagnose Deadly Conditions

Two new artificial intelligence programs are helping doctors diagnose life-threatening heart infections and treat open wounds.

The research could help save hospitals millions of dollars a year in needless tests and help save patients lives while avoiding painful and invasive procedures.

"These are very serious infections with high mortality rates of between 30 to 50 percent," said M. Rizwan Sohail, a doctor at the Mayo Clinic in Rochester, Minn. who developed software to identify patients with heart infections. Sohail points out that diagnosing endocarditis "is an invasive procedure. We want to diagnose the infection without having to stick a probe down a person's food pipe."

The formal name for a picture of someone's heart taken with a camera-equipped probe insert into a sedated patient's throat is a transesophageal endocardiogram. The process is invasive and expensive. One 30-minute transesophageal endocardiogram can cost up to $2,000. The procedure is so technically demanding that many hospitals don't even have the right equipment to perform it.

Instead of inserting tubes, the Mayo doctors input data. First doctors trained the software by entering the heart rate, blood pressure, white blood cell count, the presence of a pacemaker or other implanted device, body temperature and perceived body temperature of 187 patients into the computer. They also included the final diagnosis for those patients.

The computer algorithm then searched for patterns in the data linking symptoms to diagnosis.

"Initially the computer made a lot of mistakes," said Sohail. "But over time the mechanism develops and becomes smarter, to the point where the network makes good decisions in a timely matter and with a good level of confidence in that decision."

In 50 percent of the cases the software can make a prediction with 99.99 percent accuracy in less than four seconds.

"Anything less and we felt we couldn't trust the diagnosis," said Sohail.

For the remaining cases the software was accurate 80 percent of the time or better.

The next step for the Mayo researchers is to use the software on 200 cases where the computer doesn't know the final diagnosis.

Heart infections are difficult to diagnose but can often be treated with antibiotics taken over a week or so. Open wounds that refuse to heal after weeks or months of treatment, known as ischemic wounds, are easy to diagnose but frustratingly difficult to treat.

Chandan Sen, a professor and doctor at The Ohio State University in Columbus, along with his colleagues, has developed a mathematical algorithm that predicts when an ischemic wound will close and what complications are likely to arise during closure.

"Current models address wounds that will close anyway," said Sen, who co-authored a paper that appeared in a recent issue of the Proceedings of the National Academy of Science. "We wanted to model wounds that won't close."

Six and a half million Americans suffer ischemic wounds each year. They often occur in older patients, as foot ulcers in diabetic patients, or in patients who are already hospitalized for other conditions.

To help treat ischemic wounds, Sen and his colleagues developed a program that processes patient data like oxygen concentration, growth factors, presence of white blood cells, and fibroblast density. Using this data the computer creates a 3D model of the wound, looks at how quickly the wound is healing, and estimates when the wound will close.

According to the model, a normal wound will close in about 13 days. After 20 days however, only 25 percent of ischemic wounds heal. The numbers match what actually happens to patients, but so far have only been used in theory -- the model has not yet been tested on human patients.

Whether used for wound healing or heart infections, artificial intelligence programs will not replace human doctors anytime soon.

"These neural networks cannot see patients, they cannot examine them to find signs of infection or symptoms," said Sohail. "Most cases are pretty straightforward. These programs will mostly assist clinicians in the confusing cases."