Skip to content
Open access

Explainable 1D-CNN for Automated Detection of Valvular Heart Diseases Using PCG Signals

2026 · IEEE Access · Vol 14, pp. 164537-164551 · 0 citations · 38 references

Abstract

Valvular heart diseases (VHDs) produce characteristic alterations in phonocardiogram (PCG) signals that can support automated screening. However, the limited interpretability of deep-learning models restricts their clinical acceptance. This study proposes an explainable one-dimensional convolutional neural network (1D-CNN) for classifying VHDs directly from preprocessed PCG signals and employs Gradient-weighted Class Activation Mapping (Grad-CAM) to identify the temporal regions contributing to each prediction. Two classification settings were investigated: five-class classification using three-cycle PCG segments and nine-class classification using physician-identified single-cycle pathological heart-sound patterns. The five-class model achieved an accuracy of 84.42%, with macro-averaged sensitivity, specificity, precision, and F1-score of 85.38%, 96.19%, 85.24%, and 84.76%, respectively. The nine-class experiment achieved an accuracy of 89.75%, with macro-averaged sensitivity, specificity, precision, and F1-score of 93.86%, 98.77%, 83.81%, and 85.95%, respectively. Grad-CAM analysis showed that the network assigned importance to clinically meaningful events, including systolic and diastolic murmurs, systolic clicks, opening snaps, and the third heart sound. These observations demonstrate the potential of integrating waveform-based classification with explainable AI for clinically interpretable PCG analysis.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.