Abstract: Background: A heart murmur is a key clinical feature of canine degenerative valve disease (DVD). Recent machine learning algorithms, combined with electronic stethoscopes, have shown promise at detecting heart murmurs in people, but further evidence is needed of their application to canine heart sounds.
Hypothesis/Objectives: To train and evaluate a machine learning algorithm to detect and grade heart murmurs in dogs, using recordings from an electronic stethoscope.
Animals: 470 client-owned dogs (266 DVD, 132 other cardiac diseases, 72 normal) undergoing routine echocardiography at four referral centres.
Methods: Diagnostic method development and assessment. For each dog, a board-certified cardiologist or cardiology resident used an electronic stethoscope to grade murmurs and make 15-second recordings at three auscultation sites (left apex, left base, right side). A recurrent neural network, originally developed for human murmur detection, was then retrained to predict the cardiologist’s murmur grade from the sound recording.
Results: On an unseen half of the data set, the neural network detected a murmur of any grade with an area under the receiver operating characteristic of 0.907 (95% CI 0.887–0.927), with an operating sensitivity of 89.3% (95% CI 85.1%–93.5%) and specificity of 74.9% (95% CI 66.7%–83.0%). The individual Levine murmur grade was predicted with a mean absolute error of 0.648 (95% CI 0.602–0.694).
Conclusions and Clinical Importance: The algorithm demonstrates good agreement with cardiologists, potentially enabling a non-expert user to perform accurate murmur screening and grading.
