Beyond Heart Murmur Detection: Automatic Murmur Grading From Phonocardiogram.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Heart murmurs are abnormal heart sounds graded by intensity, correlating with patient health.
• Accurate murmur grading is crucial but challenging in low-resource settings due to expert limitations.

Purpose of the Study:

• To develop and validate an AI-driven method for estimating pediatric heart murmur grades (absent, soft, loud) from phonocardiograms (PCGs).
• To assess the feasibility of this approach for pre-screening in resource-limited rural environments.

Main Methods:

• An ensemble of 15 convolutional residual neural networks with channel-wise attention was used to classify Mel spectrograms of PCGs.
• A decision rule integrated classifications from multiple auscultation locations for patient-level grading.
• The model was rigorously cross-validated on a large dataset (3456 PCGs from 1007 patients) and tested on a hidden set (1538 PCGs from 442 patients).

Main Results:

• Cross-validation achieved an 86.3% unweighted average sensitivity and 81.6% F1-score for patient-level murmur grading.
• Sensitivities for absent, soft, and loud murmurs were 90.7%, 75.8%, and 92.3%, respectively.
• The hidden test set yielded an 80.4% unweighted average sensitivity and 75.8% F1-score.

Conclusions:

• The AI method offers a viable solution for algorithmic pre-screening of pediatric heart murmurs.
• This technology can help overcome high expert screening costs in low-resource settings.
• The approach advances beyond simple murmur detection to intensity characterization, potentially improving clinical outcome classification.