Computerized method homes in on heart murmurs

Methodology more accurate than human exam

The adaptation of a computer-based technology to the analysis of heart murmurs in children may improve the chance of more accurate diagnoses in the future, according to a study conducted by Denver-based researchers. The study involved the use of an artificial neural network, or ANN, in conjunction with a stethoscope. In 69 patients (37 with abnormal heart murmurs and 32 with innocent murmurs), they achieved 100% sensitivity (the ability to identify an abnormal heart murmur) and 100% specificity (the ability to identify an innocent heart murmur).

An ANN is a computer program that can recognize complex patterns, and previously has been used successfully in the evaluation of electrocardiogram signals.

"My thought was, if we trained a device with the best possible information, we would get advice at least as good as the best physician," says Curt G. DeGroff, MD, a pediatric cardiologist at The Children’s Hospital and assistant professor in the department of pediatrics at the University of Colorado Health Sciences Center in Denver.

The researchers used heart-sound recordings from the patients to train the network; the mathematical signature for each child, as well as the patterns for innocent and abnormal murmurs, are different. Using a mathematical model, they converted the recordings into the energy-per-unit of frequency interval to take advantage of the computer’s pattern-recognition capabilities. Then they fed samples of the recordings back to the ANN model and adjusted frequency range and sensitivity to differentiate between the abnormal and innocent murmurs.

It’s extremely important to correctly diagnose heart murmurs in children, DeGroff notes. When a patient visits a primary care physician and is noted to have a heart murmur, one of two things could go wrong: The physician could interpret a murmur that signifies a problem as an innocent heart murmur and not refer the patient to a specialist; or the physician may listen to an innocent murmur, think it’s pathologic, and send the patient to a specialist for further evaluation, he adds.

"It may take days to weeks to see a specialist, and some amount of anxiety is put on the family during that time period. Now, in addition to a history and auscultation, the primary care physician may end up ordering a chest X-ray or an EKG, but these are not very sensitive or specific. We hope this new method might provide a better sensitivity and specificity," he says.

DeGroff is quick to emphasize that the human factor remains essential. "The doctor doesn’t just come in and listen to the heart," he says. "There’s the whole spectrum of physical exam we are not claiming to replace. It will not, for example, replace an echocardiogram. But it will assist the physician in making the determination of whether the murmur is innocent or pathological, and whether to send the patient to a specialist."

DeGroff also points out that 69 patients is not a large enough group to definitively evaluate the ANN/stethoscope method. "I suspect we need 500 or more, and we’re working on that right now. But theoretically, we can make this as good as the best clinician in auscultation."

DeGroff sees possible applications for this diagnostic approach in hospitals, physicians’ offices, and school nurses’ offices. "We could make it accurate enough in that setting," he notes. "We also have future plans to extend this to use in adults, but right now we’re focusing on pediatric patients."

The bottom line, he says, is that the use of ANN opens up the possibility for better screenings in many areas. "This has been my first opportunity to apply neural networks in cardiology, but I think there are a lot more applications in cardiology and even beyond that," DeGroff observes. "People are looking at analyzing images as well as sounds; a group here at the University of Colorado is looking at histologic samples of prostate cells; you might be able to train the ANN to recognize visual differences of how cells look. Another group is looking at stool samples. It may have applications for pathologic samples, images, sounds, and even clinical decisions, such as the complex nature of trying to predict when a patient is beginning to undergo organ rejection. It holds a lot of promise."

Need more information?

For more information, contact:

Curt G. DeGroff, MD, UCHSC — The Children’s Hospital, 1056 East 19th Ave. B100, Denver, CO 80218. Telephone: (303) 861-6821.