Assisted auscultation : creation and visualization of high dimensional feature spaces for the detection of mitral regurgitation

Abstract

Cardiac auscultation, listening to the heart using a stethoscope, often constitutes the first step in detection of common heart problems. Unfortunately, primary care physicians, who perform this initial screening, often lack the experience to correctly evaluate what they hear. False referrals are frequent, costing hundreds of dollars and hours of time for many patients. We report on a system we have built to aid medical practitioners in diagnosing Mitral Regurgitation (MR) based on heart sounds. Our work builds on the "prototypical beat" introduced by Syed in [17] to extract two different feature sets characterizing systolic acoustic activity. One feature set is derived from current medical knowledge. The other is based on unsupervised learning of systolic shapes, using component Analysis. Our system employs self-organizing maps (SOMs) to depict the distribution of patients in each feature space as labels within a two-dimensional colored grid. A user screens new patients by viewing their projections onto the SOM, and determining whether they are closer in space, and thus more similar, to patients with or without MR. We evaluated our system on 46 patients. Using a combination of the two feature sets, SOM-based diagnosis classified patients with accuracy similar to that of a cardiologist.