Automatic estimation of canine heart murmur with electronic stethoscope and software

Abstract

Heart sounds carry information about the health of heart structures and about the dynamics of blood circulation. Auscultation is the activity dedicated to the analysis of sounds produced spontaneously by organs — which includes heart sounds. Despite being important to detect diseases early and to improve the use of resources in the health sector, both human and animal, the practice of auscultation is still largely subjective and dependent on factors such as experience, technique and auditory ability of the professional, as well as environmental noise and friction. The present work represents an effort to increase tooling for veterinary medicine practice by proposing a new method to automatically estimate the degree of canine heart murmur on a scale of 0 to 6 from a record of heart sounds marked with the position of the main heartbeats. The method includes the generation of features from this record (through specially adapted algorithms) and the use of machine learning techniques to make the estimation. The proposed method was developed from a set of records of 56 patients, each of which had 9 heart cycles processed. Of the 4032 rows of data resulting from the study, only 0.35% of them could not have the degree of murmur adequately estimated. An analysis of the error profile allows one to conclude that the estimated and the actual degree of murmur of the patient never reach a difference of 2 degrees or more, and that further techniques can be researched to reduce the error even more or to increase the generalization capacity of the developed method.