Comprehensive Summary
This study investigated the ability of AI to detect pediatric congenital heart disease (CHD) from phonocardiogram recordings (PCGs). A total of 3,004 PCGs were collected from 751 pediatric patients, aged 1 month to 16 years, at four auscultation sites using the FDA-cleared Eko DUO digital stethoscope in Bangladesh between 2021 and 2023. Recordings were labeled as CHD or non-CHD based on cardiologist-confirmed diagnosis through echocardiography, CT, ECG, and auscultation. Preprocessing included noise removal with a Butterworth band-pass filter and z-score normalization. Feature extraction was performed through handcrafted feature extraction (heart rate variability, brightness, timbre, and frequency), Mel-frequency cepstral coefficients (MFCCs) feature extraction (transforming the data into logarithmic frequencies reflecting human hearing for AI analysis), and deep feature extraction of 4 layers with a 1D convolutional neural network. Features were then classified using a fully connected neural layer. To address data imbalance (63% CHD, 37% non-CHD), the team implemented weighted loss functions. Predictions from multiple auscultation sites were aggregated at the patient level, with averaging outperforming majority voting and “at least one positive” approaches. Results showed high accuracy (92%), sensitivity (91%), specificity (92%), and AUROC (96.4%), with five-fold cross-validation confirming consistent performance (91% accuracy, 93% AUROC). A feature contribution analysis showed handcrafted and MFCC features contributed complementary diagnostic value, and comparisons with prior literature confirmed this approach surpassed earlier machine learning and deep learning models applied to PCGs.
Outcomes and Implications
Congenital heart disease is a leading cause of child mortality, especially in low- and middle-income countries where echocardiography is not widely available. The findings suggest that AI-enhanced digital auscultation could provide an affordable and scalable frontline screening method, allowing for earlier identification of children who should be referred for echocardiography. However, some factors limit the generalizability of the results, including that the dataset was drawn from a single country and device, no external or subgroup validation was performed, and CHD subtype classification was not explored. The authors recommend future directions such as incorporating ECG data, building multi-device datasets, implementing signal-quality checks, and developing interpretable AI models.