Session S52.1

Similarity Retrieval of Heart Sounds

T Syeda-Mahmood*

IBM Almaden Research Center
San Jose, CA, USA

Heart auscultation, i.e., listening to the sounds produced by the heart, is a common practice in the screening of heart disease. Tools that retrieve similar heart sounds as those of a patient can be useful in several applications including physician training, diagnostic screening, and decision support.
Heart sounds are highly non-stationary signals of short duration. Further, the energy content of low-frequency vibrations is much higher than that of high-frequency vibrations. Thus much of the discriminating heart sound information is captured in their low frequency envelopes. Finally, it can be shown that heart sounds from different patients with the same diseases have similar visual appearance which can enable visual shape similarity measures to capture the audio similarity of the heart sounds.
Heart sound matching, however, is a difficult problem. To make the comparison meaningful, the variations in heart beat must be taken into account. This requires extraction of single heart beats from signals which are not always periodic (particularly for cardiac arrhythmias). Further, the amplitude variations due to recording levels and varying digital stethoscope qualities affect the matching. Finally, due to variability in the systolic and diastolic phases across patients, both advancement as well as pre-ponement of cardiac audio phases (sounds S1 and S3) are possible. The net effect due to these factors is to introduce a non-linear deformation along the time axis, which must be modeled. In this paper, we present a method of matching heart sounds that explicitly addresses the above problems. Specifically, we extract single heart beats from sound signals, and approximate them through audio envelopes. The morphological shape changes of audio envelopes of heart sounds for the same disease are modeled as a constrained non-rigid translation transform. Similar heart sounds are then retrieved by recovering the corresponding alignment transform using a variant of shape-based dynamic time warping. The audio shape matching algorithm was tested on a large database of heart sounds assembled from collections used for training physicians at medical schools as well as from reference data provided by Littman stethoscopes. Currently, the collection has over 670 heart sound examples for various kinds of murmurs, Mitral regurgitation, Mitral Stenosis, septal defects, Cardiomyopathy, etc. Each disease was represented by 7-15 patient samples in the database. We compared our performance on precision and recall using all of the heart sounds as queries against the popular MFCC method of audio analysis. The ROC curve shows that our method outperforms MFCC on both precision and recall. For mitral regurgitation sounds we were able to get 70.8% precision and 56.6% recall in comparison to 54% and 43% respectively using MFCC. For PDA cases, we were able to get 75% recall and 100% precision in comparison to 54% and 43% respectively using MFCC.

(Abstract Control Number: 213)