The study by Youness et al. explores the use of deep-learning models to accurately identify Still’s murmur, the most common innocent pediatric heart murmur, during routine clinical care. Researchers collected pediatric phonocardiograms (PCGs) using the StethAid digital stethoscope platform across four major medical centers, in addition to incorporating data from the Littmann 4100 dataset. A previously validated heart sound segmentation algorithm (NR-HSS) located the primary heart sounds (S1 and S2) to segment PCGs into periodic heart cycles. These cycles were converted to grayscale spectrograms for training and testing of two types of deep-learning models: convolutional neural networks (ConvNets) and transformer-based architectures. The results demonstrated that both ConvNets and transformers achieved high diagnostic performance. Reported sensitivities ranged from 90.7% to 100% and specificities from 75% to 98.2%, with the strongest models, such as ResNet18, DeiT, and AST, meeting or exceeding the clinically acceptable benchmarks of greater than 90% sensitivity and specificity. ​​ConvNets generally showed slightly higher sensitivity, while transformer-based models performed more consistently overall and achieved stronger specificity. Additionally, the Littmann dataset improved transformer accuracy by at least 2%, highlighting the benefit of domain-specific transfer learning. In their discussion, the authors highlight the potential of AI supported auscultation in supporting providers, while also acknowledging the need for future studies that can validate their findings on independent datasets.