Artificial Intelligence Helps Diagnose Cardiac Infections

Researchers reported that "teachable software" designed to mimic the human brain may help them identify cardiac infections without an invasive exam.

The study's findings were presented by investigators from the Mayo Clinic (Rochester, MN, USA) in September 2009 at the Interscience Conference on Antimicrobial Agents and Chemotherapy (ICAAC) in San Francisco, CA, USA.

Endocarditis, an infection involving the valves and sometimes chambers of the heart, can be a problem in patients with implanted medical devices. It is serious and can be deadly. The mortality rate can be as high as one in five, even with aggressive treatment and removal of the device. With further complications, the mortality could be over 60%. Diagnosis typically requires transesophageal echocardiography, an invasive procedure that also has risks. It involves use of an endoscope and insertion of a probe down the esophagus.

The software program, called an artificial neural network because it mimics the brain's cognitive function and reacts differently to situations depending on its accumulated experience, is made of programs that mimic interconnected artificial neurons. That knowledge or training is provided by the researchers by affecting the connection weights of the artificial neurons; similar to how a person would train a dog by repetition, reinforcing behavior at each instance. In this case, the neural net underwent three separate "trainings" to learn how to evaluate the symptoms it would be considering.

"If, through this novel method, we can help determine a percentage of endocarditis diagnoses with a high rate of accuracy, we hope to save a significant number of patients from the discomfort, risk, and expense of the standard diagnostic procedure," said M. Rizwan Sohail, M.D., a Mayo Clinic infectious diseases specialist and leader of the study.

The team examined 189 Mayo patients with device-related endocarditis diagnosed between 1991 and 2003. The artificial neural network was tested retrospectively on the data from these cases. When tested on cases with known diagnosis of endocarditis, the best-trained artificial neural network was correct most of the time (72 of 73 implant-related infections, and 12 of 13 endocarditis cases) with a confidence level greater than 99%.

Researchers reported that, when used on an overall sample that included both known and unknown cases, the neural net accurately excluded endocarditis in at least half of the cases, thereby eliminating half the cohort from a needless invasive procedure.

Related Links:
Mayo Clinic