MMMuLNet: Multimodal Mutual Deep learning Framework for the Pediatric Congenital Heart Disease Detection

Congenital heart disease (CHD) is the most common birth defect and a major cause of infant morbidity and mortality worldwide. While echocardiography remains the diagnostic gold standard, its cost and reliance on expert interpretation limit availability in low- and middle-income countries (LMICs). This underscores the need for scalable, affordable, and robust approaches that can operate reliably under real-world signal degradation and low-quality recording conditions. In this study, we present a multimodal framework for pediatric CHD detection that integrates phonocardiograms (PCG), electrocardiograms (ECG), and clinical symptoms. Features from PCG and ECG were extracted using a pretrained foundation audio model, while key symptoms were selected using SHAP (SHapley Additive exPlanations) and encoded as structured embeddings. These modalities were combined in a fusion network with modality dropout to handle missing inputs, and mutual learning was applied to promote knowledge sharing across unimodal and multimodal branches. The framework was validated on a dataset of 751 pediatric patients in Bangladesh, comprising 3,435 synchronized PCG-ECG recordings with expert-confirmed labels. In 10-fold patient-wise cross validation, PCG alone achieved 90% accuracy, ECG alone 79%, and symptoms alone 73%. Combining PCG and ECG improved accuracy to 94%, while the full multimodal system reached 94.5% accuracy, 94.2% sensitivity, 95.6% specificity, and an AUROC of 96%. Importantly, even under degraded signal conditions when both PCG and ECG were of low quality, the multimodal framework maintained an accuracy of 88% and an AUROC of 87%, demonstrating superior performance compared to the unimodal models. These findings demonstrate that integrating PCG, ECG, and symptoms enables accurate, resilient CHD screening, offering a practical pathway for scalable early detection in LMICs.