Congenital Heart Disease Classification Using Phonocardiograms: A Scalable Screening Tool for Diverse Environments

Abdul Jabbar, Ethan Grooby, Jack Crozier, Alexander Gallon, Vivian Pham, Khawza I Ahmad, Raqibul Mostafa, Md Hassanuzzaman, Ahsan H. Khandoker, Faezeh Marzbanrad · 2025

Read Congenital Heart Disease Classification Using Phonocardiograms: A Scalable Screening Tool for Diverse Environments by Abdul Jabbar, Ethan Grooby, Jack Crozier, Alexander Gallon, Vivian Pham, Khawza I Ahmad, Raqibul Mostafa, Md Hassanuzzaman, Ahsan H. Khandoker, Faezeh Marzbanrad (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-22773.txt.

Introduction

Abdul Jabbar and co-authors (2025) train a model that hears congenital heart disease in a child's chest sounds. These birth defects demand early detection, yet specialist screening is rare where many children are born. A phonocardiogram, the recorded sound of the heartbeat, is cheap to capture with a digital stethoscope. The team's deep learning model reads those recordings. On its main dataset from Bangladesh it reached 94.1 percent accuracy, catching 92.7 percent of cases while clearing 96.3 percent of healthy children.

The model held up on two public challenge datasets from other populations, a sign it is not overfit to one clinic. It kept accuracy above 85 percent using a single spot on the chest, where a full exam uses several. That robustness matters for the intended use: a health worker with one stethoscope position, in a clinic without a cardiologist, screening children who would otherwise go unheard.

Catalog record

CREATORAbdul Jabbar, Ethan Grooby, Jack Crozier, Alexander Gallon, Vivian Pham, Khawza I Ahmad, Raqibul Mostafa, Md Hassanuzzaman, Ahsan H. Khandoker, Faezeh Marzbanrad
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE36 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-22773

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Congenital Heart Disease Classification Using Phonocardiograms: A Scalable Screening Tool for Diverse Environments Abdul Jabbar, Ethan Grooby, Jack Crozier, Alexander Gallon, Vivian Pham, Khawza I Ahmad, Raqibul Mostafa, Md Hassanuzzaman, Ahsan H. Khandoker, Faezeh Marzbanrad Abstract Congenital heart disease (CHD) is a critical condition that demands early detection, particularly in infancy and childhood. This study presents a deep learning model designed to detect CHD using phonocardiogram (PCG) signals, with a focus on its application in global health. We evaluated our model on several datasets, including the primary dataset from Bangladesh, achieving a high accuracy of 94.1%, sensitivity of 92.7%, specificity of 96.3%. The model also demonstrated robust performance on the public PhysioNet Challenge 2022 and 2016 datasets, underscoring its generalizability to diverse populations and data sources. We assessed the performance of the algorithm for single and multiple auscultation sites on the chest, demonstrating that the model maintains over 85% accuracy even when using a single location. Furthermore, our algorithm was able to achieve an accuracy of 80% on low-quality recordings, which cardiologists deemed non-diagnostic. This research suggests that an AI-driven digital stethoscope could serve as a cost-effective screening tool for CHD in resource-limited settings, enhancing clinical decision support and ultimately improving patient outcomes. K eywords Echocardiography \cdot Phonocardiography (PCG) \cdot Deep Learning, Congenital heart disease (CHD) \cdot Low- and Middle-Income Countries (LMIC) 1 Introduction Congenital heart disease (CHD), also known as congenital heart defects, occurs in approximately 1.2% of newborns globally, with one in four cases being severe [ 1 ] . At least 220,000 people die from this disease annually, with most deaths occurring within the first year of life [ 2 ] . In some extreme cases, CHD can damage the heart valves and limit blood circulation through the coronary arteries or other parts of the body, typically requiring surgery within the first year of life [ 3 ] . While CHD mortality rates have dropped dramatically in high-income nations due to advances in diagnosis and treatment, it remains a major cause of morbidity and mortality in Low- and Middle-Income Countries (LMICs) [ 4 ] . These regions often lack the resources and advanced medical care necessary for CHD diagnosis and treatment, leading to higher costs and mortality rates associated with CHD [ 5 ] . Early detection through emerging screening methods could potentially reduce mortality rates in these countries by about 58%, preventing many cases of disability and death [ 6 ] . Echocardiography is currently the gold standard for detailed pre- and post-natal screening and is used to diagnose CHD [ 7 ] . Although echocardiography is an established and reliable test, its high cost and the need for skilled specialists to operate and interpret the results make it difficult to implement in low-resource settings. Additionally, variations in professional interpretations [ 8 ] among clinicians can delay the early detection of cardiac abnormalities when relying solely on expert opinion [ 9 ] . Moreover, many patients in rural and regional areas often have to travel considerable distances to reach hospitals with the necessary resources and trained personnel [ 10 ] . Phonocardiography (PCG), a non-invasive technique for cardiac auscultation that provides insights into the heart’s mechanical function [ 11 ] , is a potential alternative [ 12 ] . Using an electronic stethoscope, PCG signals capture the activity of the mitral, tricuspid, aortic, and pulmonic valves, along with other cardiac structures, as blood flows through and around the heart [ 13 ] . Four key cardiac sounds are produced as blood passes through these valves: S1, S2, systole, and diastole. Irregular heartbeats or heart murmurs during diastole or systole may indicate CHD [ 14 ] . Specifically, a diastolic heart murmur occurs after S2 and before S1, while a systolic murmur occurs between S1 and S2 [ 15 ] . With advancements in Artificial Intelligence (AI) and computerized algorithms, alongside the growing demand for accessible, continuous, and personalized healthcare, PCG could play a significant role in improving CHD diagnosis and management, particularly in LMICs [ 16 ] . Automated assessment of PCG signals has gained increasing attention, culminating in the focus of the PhysioNet 2016 and 2022 Challenges on this topic. The objective of the PhysioNet Challenge 2016 was the automated classification of normal and abnormal adult PCG signals [ 17 ] , while the objective of the PhysioNet Challenge 2022 was to automate the identification of murmurs and clinical outcomes from pediatric PCG signals [ 18 ] . The top-performing team in the PhysioNet Challenge 2016 was Potes et al. , who utilized AdaBoost and convolutional neural networks (CNN) for classifying heart sounds, extracting 124 features from the audio data, and achieving an accuracy of 86.02%. Furthermore, the top-performing team in the PhysioNet Challenge 2022, Yujia Xu et al. , achieved the highest weighted accuracy of 85.3% in murmur detection using a hierarchical multi-scale CNN (HMS-Net). Their method involved extracting convolutional features from spectrograms at various scales and integrating them through a hierarchical structure [ 19 ] . The second-place team, Hui Lu et al. , developed a lightweight CNN and a Random Forest classifier for heart disease classification. Their approach utilized 128 mel spectrogram features from heart sounds, along with demographic information, achieving an weighted accuracy of 79.1% in murmur detection [ 20 ] . The third-place team, McDonald et al. , used a recurrent neural network combined with a hidden semi-Markov model (HSMM) for murmur detection, resulting in an weighted accuracy of 77.6% [ 21 ] . Al-Issa et al. present a cardiac diagnostic system that combines CNN and long short-term memory (LSTM) components, achieving impressive accuracy rates of 98.5% and 99.87% for non-augmented and augmented datasets, respectively, while also demonstrating strong performance on the PhysioNet/Computing in Cardiology 2016 challenge dataset [ 22 ] . Despite these advances, there are significant research gaps in the practical implementation of detecting CHD using stethoscopes in LMICs. Most studies focus on murmur detection and CVD, which are not equivalent to CHDs. Aiming to detect CHDs using PCGs recorded in Bangladesh, Hassanuzzaman et al. , proposed a transformer-based neural network, achieving 93% accuracy and 97% sensitivity [ 23 ] . While proving the feasibility of detecting CHDs from heart sounds, did not address several practical challenges relevant to real-world deployment, particularly in LMICs. It is crucial to analyze how the algorithm performs with low-quality signals, that is a common issue in LMICs, especially in pediatric populations and mobile health contexts, where PCGs are often recorded by non-experts. Morover, the sensitivity of CHD detection to the precise site of sound collection on the chest also remains unexamined. It is unclear whether all four major auscultation sites are necessary for accurate CHD detection, or if a single location or subset of these would suffice. In a mobile health context, collecting PCGs from multiple precise locations could be challenging for non-experts, impacting the usability of the method. Furthermore, it has not been explored how well the algorithm, trained on a specific population and using a particular stethoscope, would generalize to a different population or device. The transferability of algorithms developed for other tasks, such as murmur classification, to CHD detection is also yet to be explored. Addressing these gaps is

Read the complete text