# Beyond Heart Murmur Detection: Automatic Murmur Grading from Phonocardiogram

**Authors:** Andoni Elola, Elisabete Aramendi, Jorge Oliveira, Francesco Renna, Miguel T. Coimbra, Matthew A. Reyna, Reza Sameni, Gari D. Clifford, Ali Bahrami Rad  
**arXiv:** [2209.13385](https://arxiv.org/abs/2209.13385)  
**Date:** 2022-09-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2209.13385v2/

## Key points

- Objective: Develop an algorithm to automatically grade the intensity of heart murmurs in pediatric patients from low-resource rural areas.
- Phonocardiograms (PCGs) recorded from multiple auscultation locations on a large population of pediatric patients.
- Mel spectrogram representation used for each PCG recording.
- Ensemble of 15 convolutional residual neural networks with channel-wise attention mechanisms employed for classification.
- Decision rule proposed to determine final murmur grade based on estimated labels from available recordings.
- Cross-validation performance: Unweighted average sensitivity of 86.3% and F1-score of 81.6% for patient-level murmur gradings.
- Sensitivities (and F1-scores) for absent, soft, and loud murmurs were determined to be:
- Absent: 90.7% (93.6%)
- Soft: 75.8% (66.8%)
- Loud: 92.3% (84.2%)
- Test set performance: Unweighted average sensitivity of 80.4% and F1-score of 75.8%
- Potential approach for algorithmic pre-screening in low-resource settings
- Provides characterization of murmur intensity, contributing to improved classification of clinical outcomes and patient outcomes in resource-constrained environments.
- Assists healthcare providers in identifying patients who require further evaluation and intervention
- Improves patient outcomes in resource-constrained environments

## Summary

Objective: The objective of this study is to develop an algorithm that can automatically grade the intensity of heart murmurs in pediatric patients from a low-resource rural area. Murmurs are abnormal heart sounds that can provide important information about a patient's clinical condition, and their grading is strongly correlated with the severity of the underlying condition. Methods: The researchers used phonocardiograms (PCGs) recorded from multiple auscultation locations on a large population of pediatric patients. They employed a Mel spectrogram representation of each PCG recording and fed it into an ensemble of 15 convolutional residual neural networks with channel-wise attention mechanisms. This ensemble was trained to classify each PCG recording based on its murmur grade. To determine the final murmur grade for each patient, a decision rule was proposed that considered all estimated labels from available recordings. The proposed method was cross-validated using a dataset consisting of 3456 PCG recordings from 1007 patients, employing a stratified ten-fold cross-validation approach. Additionally, the method was tested on a hidden test set comprising 1538 PCG recordings from 442 patients. Results: The overall performance of the algorithm in cross-validation for patient-level murmur gradings was evaluated using the unweighted average of sensitivities and F1-scores, which were found to be 86.3% and 81.6%, respectively. The sensitivities (and F1-scores) for absent, soft, and loud murmurs were determined to be 90.7% (93.6%), 75.8% (66.8%), and 92.3% (84.2%), respectively. On the test set, the algorithm achieved an unweighted average sensitivity of 80.4% and an F1-score of 75.8%. Conclusions: This study presents a potential approach for algorithmic pre-screening in low-resource settings where expert screening costs may be high. The proposed method goes beyond the detection of murmurs and provides a characterization of their intensity, which can contribute to enhanced classification of clinical outcomes and improved patient outcomes in these resource constrained environments.. Significance: The development of an algorithm that can automatically grade the intensity of heart murmurs has significant implications for low-resource settings as it can assist healthcare providers in identifying patients who require further evaluation and intervention potentially improving patient outcomes in these resource constrained environments . Objective: This study aims to develop an algorithm that can automatically grade the intensity of heart murmurs in pediatric patients from low resource rural areas . Murmurs are abnormal heart sounds which provide important information about a patient's clinical condition , with their grading being strongly correlated with severity levels . Methods: Phonocardiograms (PCGs) were recorded from multiple locations on pediatric patients . A Mel spectrogram representation was used for each PCG recording , then fed into an ensemble consisting 15 convolutional residual neural networks with channel wise attention mechanisms . An estimation label was assigned to each PCG recording based on its murmur grade , then all estimated labels were taken into consideration when determining final murmur grades per patient . A stratified ten fold cross validation approach was employed using 3456 PCG recordings from 1007 patients , followed by testing on 1538 PCGs from 442 patients . Results: Unweighted average sensitivity & F1 scores during cross validation yielded 86 . 3 % & 81 . 6 % respectively , while sensitivities & F1 scores for absent , soft & loud murmurs were 90 . 7 % & 93 . 6 % ; 75 . 8 % & 66 .