Key result
Late-fusion CNN-Transformer model achieves robust heart murmur detection across diverse recording devices.
Why the study?
Robustness of AI-enabled digital stethoscopes for heart murmur detection across devices, sites, and noise requires rigorous, reproducible evaluation protocols.
Does a late-fusion 1D+2D CNN/Transformer model with acoustically targeted augmentation maintain high discrimination and calibration for heart murmur detection across unseen devices/sources?
Does a late-fusion 1D+2D CNN/Transformer model with acoustically targeted augmentation maintain high discrimination and calibration for heart murmur detection across unseen devices/sources?
A prespecified, source-held-out evaluation protocol on the public PhysioNet 2016 corpus provides a transparent foundation for benchmarking robust AI auscultation models.
Enables standardized device-robust benchmarking for murmur-detection AI; leaves open prospective clinical validation before practice adoption.
AI-enabled digital stethoscopes can standardize auscultation, but real-world success depends on robustness across devices, sites, and noise. Public, de-identified corpora permit rigorous, ethics-exempt development and transparent benchmarking. To design and describe a reproducible evaluation workflow for robust murmur detection tailored to AI digital stethoscopes, using the open PhysioNet/CinC 2016 heart-sound dataset, with a protocol that prioritizes generalization to unseen sources/devices. We prespecify training/validation splits within the public CinC-2016 training pool and construct an external test set by holding out entire sources/devices (leave-source-out). The proposed modeling pipeline fuses 1D waveforms and 2D Mel-spectrograms in a late-fusion CNN/Transformer, with acoustically targeted augmentation and class-imbalance handling. The primary outcome measure is AUROC on the external test; secondary outcomes include AUPRC, sensitivity/specificity, F1, and calibration (Brier/ECE). We contextualize the planned performance against DOI-indexed CinC-2016 benchmarks and present non-redundant figures (class prevalence, ROC-space summary, multi-metric profiles). All data are de-identified and publicly available; no ethics approval or consent was required. The official CinC-2016 corpus comprises 3,153 training recordings (764 subjects) across six sources and 1,277 hidden-test recordings (308 subjects). Published hidden-test benchmarks range from modified accuracy 0.84 to 0.86 with varying sensitivity/specificity trade-offs. Our protocol operationalizes these constraints into a transparent, device-aware evaluation that can be replicated and extended without access to the hidden test. A prespecified, source-held-out evaluation on a public corpus, paired with calibrated, decision-oriented reporting, provides a strong foundation for robust AI auscultation and prepares the ground for prospective, multi-site validation.
No takes yet. Share an insight, caveat, or question.
Milić et al. (2026) studied Heart murmur (n=1,072). Late-fusion CNN/Transformer model (1D+2D) for murmur detection vs. Published benchmarks was evaluated on AUROC on the external test (leave-source-out). The proposed late-fusion CNN/Transformer modeling pipeline and evaluation protocol provides a reproducible workflow for robust heart murmur detection across unseen devices using the PhysioNet 2016 dataset.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: