Automated Detection of Poor-Quality Digital Heart Sound Recordings Using Noise Injection

In pediatric and critical-care settings, environmental noise (people talking near the child, ICU equipment, and babies crying or moving) often makes auscultation difficult with both standard and digital stethoscopes. To improve the reliability of computer-based cardiac diagnostics, this paper proposes a novel approach for assessing the quality of digitally recorded heart sounds using deep learning and artificially generated noisy datasets. We combined the CirCor DigiScope phonocardiogram dataset (5,272 heart sound recordings) with Google’s AudioSet to simulate diverse noisy environments. A deep learning model was trained on spectrogram representations of the audio, employing causal convolutions to preserve temporal sequence, integrating attention mechanisms, and applying SwiGLU activation in deeper layers to enhance feature learning. The approach was evaluated using 10-fold stratified cross-validation. Experimental results show that the proposed method achieves 95.8% ± 0.68 AUROC, 81.8% ± 15.71 F1-score, 88.7% ± 24.7 sensitivity, and 97.8% ± 0.5 specificity. The developed model not only demonstrates robust capability in distinguishing high-quality heart sound recordings from noise-impaired ones but also provides a promising reference for enhancing digital auscultation systems in clinical practice.