Abstract 15255: Auscultation and ECG for Detection of Valvular Pathologies and Ventricular Dysfunction: Can Machine Learning Help?

Abstract

Introduction: Simple testing methods such as auscultation and electrocardiogram (ECG) that are widely used may offer easy opportunity for early diagnosis of cardiac pathologies and lead to prompt treatment. Hypothesis: Combination of multiple simple testing using machine learning (ML) will achieve diagnostic efficiency on valvular pathologies and ventricular dysfunction. Methods: Auscultation at three locations (second right sternal border, Erb, apex) and 12-lead electrocardiograms (ECG) were collected in 1052 patients undergoing echocardiography. Independent cases (N=103) were also enrolled for clinical validation. The cardiovascular diseases targeted for detection were severe aortic stenosis (AS), severe mitral regurgitation (MR), and left ventricular dysfunction (LVD) with ejection fraction ≤40%. Best neural networks were obtained by four-fold cross-validation training using heart sounds and ECG (12-lead, limb-lead and precordial-lead). The output from each model was integrated into one diagnostic prediction by stacking technique. The model performance was verified in the clinical validation cohort by area under the receiver-operator-characteristics curve (AUC). Results: The best performance was achieved by stacking auscultation of all three locations and 12-lead ECG for all three pathologies (AUC = 0.931 for AS, 0.801 for MR, and 0.736 for LVD). The contribution analyses demonstrated diverse patterns of diagnostic contributions of the auscultatory locations and ECG types depending on the target diseases. For AS detection, the heart sound played a greater role compared to the limb- and precordial ECG. Auscultatory and electrocardiographic data contributed similarly for MR detection. For LVD detection, the 12-lead ECG contributed largely followed by the heart sound at Erb. Conclusions: Severe AS, severe MR, and left ventricular dysfunction, which are currently only definitively diagnosed by imaging, can be detected with high efficiency by ML models combining auscultation and electrocardiography with synergistic effect. The contribution of each testing data differs depending on the target diseases. Our data suggests that single-modal AI may not necessarily diagnose multiple diseases.