AI-Enabled Digital Stethoscope for Early Detection of Congenital Heart Disease: Prospective Study

Early detection of congenital heart disease (CHD) is critical, particularly in low- and middle-income countries where access to congenital cardiac services is limited. While auscultation remains a first-line screening tool, its accuracy is operator dependent. Advances in digital stethoscopes and artificial intelligence (AI) may enhance diagnostic performance. To evaluate the accuracy of an AI-based algorithm integrated with a digital stethoscope for detecting CHD in children and young adults, using echocardiography as the reference standard. In this prospective single centre study, heart sounds were recorded from 721 subjects (382 with CHD, 339 controls) using the digital stethoscope at four standard auscultatory sites. Of these, 301 recordings were used for algorithm training and 420 for blinded forward testing, which was divided temporally with the data collected in first 5 months used for training and the data collected after that used for forward testing. Recordings were denoised, transformed into mel-spectrograms, and analyzed using a convolutional neural network (CovNN) to classify CHD probability. A threshold probability of > 0.5 across > 2 sites was used to predict CHD. Performance was compared against echocardiography. In forward testing, the AI algorithm correctly identified 222 of 233 CHD cases and 171 of 187 controls, yielding a sensitivity of 95.2% and specificity of 91.4%. The algorithm remained effective in noisy environments and in uncooperative children, where conventional auscultation is challenging. An AI algorithm integrated with a digital stethoscope demonstrated high sensitivity and specificity for CHD detection in children and young adults. By leveraging mel-spectrogram features, it overcomes key limitations of isolated time-domain or frequency domain analysis and shows promise as a scalable, non-invasive screening tool for early CHD detection, particularly in resource-limited settings.