Pulse
Hearing the Beat
A Duluth researcher hopes artificial intelligence will help students and physicians fine-tune their ear for heart murmurs.
Nothing symbolizes the medical profession more than the stethoscope. But not all physicians are as proficient in the use of that iconic tool as they should be. That’s why Glenn Nordehn, D.O., an assistant professor at the University of Minnesota Medical School, Duluth campus, is using artificial intelligence to help physicians improve their ability to distinguish heart sounds that indicate disease from those that do not.
Nordehn has been collaborating with the university’s department of electrical and computer engineering for the last three years on the development of a system that can recognize heart murmurs and extra heart sounds. “We’re coming up with a system that will not replace the physician or auscultation but will help provide them with additional information,” he says.
Nordehn’s system works in combination with an electronic stethoscope that records heart sounds, which can be downloaded to a computer. The sounds are then translated into visual images (spectrograms, wavelets, and phono- cardiograms) and identified as innocent or pathological.
Nordehn got the idea for the project while teaching medical students to recognize heart murmurs. A student asked him why physicians grade heart murmurs on a scale of one to six, with six being the loudest and most severe. Nordehn discovered the system dated back to the 1920s, which made him think that modern computing could improve on it.
His team has collected more than 100 recordings of human heart sounds so far. Nordehn needs more in order to refine his system, which currently provides a diagnosis and a “certainty coefficient” indicating the extent to which the computer believes it made the right call.
Nordehn has started testing his system with medical students and physicians in hope that it helps them better learn auscultation skills.
Tuning the Ear
Diagnosing problematic heart sounds can be challenging not just for medical students but also for practicing physicians both because the sounds can be difficult to hear and because it’s difficult to distinguish problematic murmurs from those that don’t indicate disease. Researchers in New Zealand examined records of patients at Wellington Hospital in Wellington from 1984 through 1991 who had nonspecific systolic murmurs. Echocardiograms showed that 76 percent of these patients, who were between the ages of 1 and 50 years, had no indication of disease.
Experience helps physicians overcome these challenges. But there has been concern in recent years that a decline in auscultation skills has contributed to over-referral of patients for echocardiograms, which can cost hundreds of dollars.
Steven S. Saliterman, M.D., an internist in private practice in St. Louis Park with an interest in computers and biomedical engineering, says he believes that in the future computerized auscultation could help reduce unnecessary referrals. “The physician’s skill at auscultation is not as good as we would like and a computer assist could help guide a diagnosis,” Saliterman says.
The Sound of Success?
Enlisting technology to improve the accuracy of auscultation isn’t a new idea. A number of academic centers and for-profit companies have been developing their own aids for hearing abnormal heart sounds.
One leader in computer-assisted auscultation is Princeton, New Jersey-based Zargis Medical, which already has an FDA-approved system called Cardioscan. The system records heart sounds using an electronic stethoscope and guides the user through a standard auscultatory protocol with voice prompts and a graphical user interface. It displays recorded sounds visually, and detects and identifies individual heart beats, recognizes diastolic and systolic murmurs, provides information about the murmurs that helps distinguish those that are pathological, and determines the probability with which the detected murmurs have characteristics consistent with pathological ones.
Zargis used a database of several thousand heart sound recordings it collected as well as one from the Johns Hopkins University School of Medicine that has more than 6,000 heart sounds, many of which have corresponding echocardiograms, to develop Cardioscan, “training” the computer by exposing it to a variety of heart sounds.
In October, Zargis presented results of a study to the American Academy of Pediatrics. Researchers tested the ability of a group of primary care physicians to evaluate 100 heart sounds from asymptomatic patients and found that with the help of Cardioscan, the physicians were able to reduce unnecessary referrals by an average of 41 percent and reduce the number of missed pathological murmurs by an average of 46 percent.
Another company BioSignetics Corporation in Exeter, New Hampshire, sells software that presents heart sounds as sonograms on a computer as well as a product that uses a more elaborate mechanism for presenting heart sounds visually.
Nordehn, who has preliminary data showing promise in his system’s ability to identify heart murmurs and other problematic sounds, hopes his system will one day be available at little or no cost.
Despite their promise, computers aren’t infallible when it comes to identifying abnormal heart sounds. For one thing, computers need to learn to recognize sounds that can be slightly different for each person. “You can think of this as a speech-recognition problem with the heart speaking and repeating the same word over and over again,” says Raymond Watrous, Ph.D., chief technology officer and founder of Zargis. The difficulty lies in the fact that each heart pronounces the same “word” slightly differently and that the same heart repeats the word in a slightly different way each time.
It’s a challenge Watrous believes computer-based algorithms will ultimately overcome. He believes that computer algorithms will continue to improve in order to more consistently identify pathological heart sounds so physicians can make more accurate diagnoses and treatment decisions.
Regardless of how skilled computers become at identifying abnormal heart sounds, they still can’t ask about a patient’s family history and symptoms. For that reason, Nordehn believes physicians won’t find themselves sidelined by auscultation software. “A computer making a diagnosis just based on sound is a fool. You still have to talk to the patient,” he says.
—Scott D. Smith