A systematic review on deep learning in ventricular septal defect screening, diagnosis, and management

Background: Ventricular septal defect (VSD) is the most common congenital heart defect (CHD), affecting 2-6/1000 live births and representing nearly 30% of pediatric CHD cases. Despite advances in imaging, disparities in diagnosis persist in low-resource settings. Early detection and management are essential to prevent complications and improve outcomes. Objective: To systematically evaluate deep learning (DL) applications for screening, diagnosis, and management of VSD in pediatric populations. Methods: A systematic MEDLINE/PubMed search was conducted in January 2025 for studies published between 2000 and 2025 using terms including "deep learning," "artificial intelligence," and "ventricular septal defect." Studies were included if they were original, peer-reviewed, in English, involved pediatric populations, and reported on DL applications in VSD care. Two authors independently screened and reviewed studies. Quality was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 tool. Results: Fourteen studies met the inclusion criteria, encompassing 1,434,840 subjects. Due to heterogeneity, no meta-analysis was performed. Convolutional neural networks (CNNs) applied to prenatal echocardiography reached up to 100% sensitivity and 98% accuracy for VSD detection. DL models analyzing phonocardiograms showed >90% accuracy, comparable to expert auscultation. Chest X-ray-based CNNs differentiated between atrial and VSDs with an area under the curve (AUC) of 0.91. Risk stratification models predicted spontaneous VSD closure (AUC 0.98) and pulmonary hypertension risk (AUC 0.81). Conclusions: DL models, especially CNNs, show promise for enhancing VSD diagnosis and management in children. However, most studies are retrospective, lack external validation, and use inconsistent outcomes. Broader clinical integration requires prospective, multicenter trials and improved interpretability.