Deep learning of heart-sound signals for efficient prediction of obstructive coronary artery disease

Background: Due to the limitations of current methods for detecting obstructive coronary artery disease (CAD), many individuals are mistakenly or unnecessarily referred for coronary angiography (CAG). Objectives: Our goal is to create a comprehensive database of heart sounds in CAD and develop accurate deep learning algorithms to efficiently detect obstructive CAD based on heart sound signals. This will enable effective screening before undergoing CAG. Methods: We included 320 subjects suspected of CAD who underwent CAG. We employed advanced filtering techniques and state-of-the-art deep learning models (VGG-16, 1D CNN, and ResNet18) to analyze the heart sound signals and identify obstructive CAD (defined as at least one ≥50 % stenosis). To assess the performance of our models, we prospectively recruited an additional 80 subjects for testing. Results: In the test set, VGG-16 exhibited the highest performance with an area under the ROC curve (AUC) of 0.834 (95 % CI, 0.736–0.930), while ResNet-18 and CNN-7 achieved AUCs of only 0.755 (95 % CI, 0.614–0.819) and 0.652 (95 % CI, 0.554–0.770) respectively. VGG-16 demonstrated a sensitivity of 80.4 % and specificity of 86.2 % in the test set. The combined diagnostic model of VGG and DF scores achieved an AUC of 0.915 (95 % CI: 0.855–0.974), and the AUC for VGG combined with PTP scores was 0.908 (95 % CI: 0.845–0.971). The sensitivity and specificity of VGG-16 exceeded 0.85 in patients with coronary artery occlusion and those with 3 vascular lesions. Conclusions: Our deep learning model, based on heart sounds, offers a non-invasive and efficient screening method for obstructive CAD. It is expected to significantly reduce the number of unnecessary referrals for downstream screening.