Automated detection of pediatric congenital heart disease from phonocardiograms using deep and handcrafted feature fusion

A team of researchers led by Abdul Jabbar has developed a novel AI model that can automatically detect pediatric congenital heart disease (CHD) from digital stethoscope recordings. The method, detailed in a paper published on arXiv (2604.24767) and in Computers in Biology and Medicine, uses a deep feature fusion approach that combines deep learning features with handcrafted acoustic features extracted from phonocardiograms (PCGs). This hybrid technique aims to address the limitations of echocardiography, the current gold standard, which is costly and requires skilled clinicians, leading to delayed diagnoses in low-resource settings.

The model was trained and tested on PCG recordings from 751 pediatric subjects in Bangladesh, ranging from 1 month to 16 years old, with recordings taken at four standard auscultation locations (mitral, aortic, pulmonary, and tricuspid valves). The results are impressive: the model achieved 92% accuracy, 91% sensitivity, 91% specificity, a 96% AUROC, and a 92% F1-score, based on a patient-wise split of 70% training, 20% validation, and 10% testing. This performance demonstrates the potential for a cost-effective, real-time remote screening tool using only a digital stethoscope, which could significantly improve early detection rates in underserved communities.