Deep learning-based evaluation of the severity of mitral regurgitation in canine myxomatous mitral valve disease patients using digital stethoscope recordings.

BACKGROUND

Myxomatous mitral valve disease (MMVD) represents the most prevalent cardiac disorder in dogs, frequently resulting in mitral regurgitation (MR) and congestive heart failure. Although echocardiography is the gold standard for diagnosis, it is an expensive tool that involves significant clinical training to ensure consistent application. Deep learning models offer an innovative approach to assessing MR using digital stethoscopic recordings, enabling early screening and precise prediction. Thus, in this study, we evaluated the effectiveness of a convolutional neural network 6 (CNN6) in providing an objective alternative to traditional methods for assessing MR. This study, conducted at the Seoul National University Veterinary Medicine Teaching Hospital, included 460 dogs with MMVD, classified according to the American College of Veterinary Internal Medicine guidelines. Phonocardiogram signals were recorded using digital stethoscopes and analyzed using the deep models CNN6, patch-mix audio spectrogram transformer (PaSST), and residual neural network (ResNET38), which were trained to categorize MR severity into mild, moderate, and severe based on MINE score. Performance metrics were calculated to evaluate model effectiveness.

RESULTS

The CNN6-Fbank model achieved an accuracy of 94.12% [95% confidence interval (CI): 94.11-93.12], specificity of 97.30% (95% CI: 97.30-97.34), sensitivity of 94.12% (95% CI: 93.74-94.50), precision of 92.63% (95% CI: 92.29-92.97), and F1 score of 93.32% (95% CI: 93.05-93.59), outperforming the PaSST and ResNet38 models overall and demonstrating robust performance across most metrics.

CONCLUSIONS

Deep learning models, particularly CNN6, can effectively assess MR severity in dogs with MMVD using digital stethoscope recordings. This approach provides a rapid, noninvasive, and reliable adjunct to echocardiography, potentially enhancing diagnosis and outcomes. Future studies should focus on broader clinical validation and real-time application of this technology.