Heart Murmur Detection from Phonocardiogram Recordings: The George B. Moody PhysioNet Challenge 2022

Matthew Reyna1, Yashar Kiarashinejad1, Andoni Elola2, Jorge Oliveira3, Francesco Renna4, Annie Gu1, Nadi Sadr1, Erick Andres Perez Alday1, Sandra Mattos5, Miguel Coimbra4, Reza Sameni1, Ali Bahrami Rad1, Gari Clifford6
1Emory University, 2University of the Basque Country, 3Instituto de Telecomunicações, 4INESC TEC, Faculdade de Ciências da Universidade do Porto, 5Real Hospital Português, 6Emory University and Georgia Institute of Technology


Abstract

The George B. Moody PhysioNet Challenge 2022 focuses on the detection of abnormal heart function from phonocardiogram (PCG) recordings in a pediatric population. While the echocardiogram is a standard diagnostic screening tool for detecting abnormal cardiac function and structure, cardiac auscultation is a much more accessible approach that can help to identify patients with heart murmurs for follow-up diagnostic screening and treatment in resource-constrained environments.

For this Challenge, we asked participants to design working, open-source algorithms to use PCG recordings to identify heart murmurs and the eventual clinical outcomes that were determined by a full diagnostic screening. The Challenge algorithms perform prescreening of easily collected heart sound recordings and recommend expert referrals for confirmatory echocardiograms and treatment.

This Challenge provides several innovations. First, we sourced data from a pediatric population in rural Brazil with 5272 PCG recordings from 1568 patients. Second, we required the Challenge participants to submit the complete code for their training and running their models, improving the reproducibility and utility of the diagnostic algorithms. Third, we devised a cost-based evaluation metric that captures the costs of screening, treatment, and diagnostic errors, allowing us to investigate the benefits of algorithmic prescreening and facilitate the development of more clinically relevant algorithms.

To date, 81 teams have submitted a total of 294 algorithms, including 167 algorithms that successfully ran in our reproducible and containerized cloud computing environment. These algorithms represent a diversity of approaches from both academia and industry for detecting abnormal cardiac function from PCGs.