Prediction of significant congenital heart disease in infants and children using continuous wavelet transform and deep convolutional neural network with 12-lead electrocardiogram.

BACKGROUND

Congenital heart disease (CHD) affects approximately 1% of newborns and is a leading cause of mortality in early childhood. Despite the importance of early detection, current screening methods, such as pulse oximetry and auscultation, have notable limitations, particularly in identifying non-cyanotic CHD. (AI)-assisted electrocardiography (ECG) analysis offers a cost-effective alternative to conventional CHD detection. However, most existing models have been trained on older children, limiting their generalizability to infants and young children. This study developed an AI model trained on real-world ECG data for the detection of hemodynamically significant CHD in children under five years of age.

METHODS

ECG data was retrospectively collected from 1,035 patients under five years old at Chang Gung Memorial Hospital, Taoyuan, Taiwan (2013-2020). Based on ECG findings, patients were categorized into the following groups: normal heart structure (NOR), non-significant right heart disease (RHA), significant right heart disease (RHB), non-significant left heart disease (LHA), and significant left heart disease (LHB). ECG signals underwent preprocessing using continuous wavelet transformation and segmentation into 2-s intervals for data augmentation. Transfer learning was applied using three pre-trained deep learning models: ResNet- 18, InceptionResNet-V2, and NasNetMobile. Model performance was evaluated in terms of accuracy, sensitivity, specificity, F1 score, and area under the receiver operating characteristic curve (AUC).

RESULTS

Among the tested models, the model based on ResNet-18 demonstrated the best overall performance in predicting clinically significant CHD, achieving accuracy of 73.9%, an F1 score of 75.8%, and an AUC of 81.0% in differentiating significant from non-significant CHD. InceptionResNet-V2 performed well in detecting left heart disease but was computationally intensive. The proposed AI model significantly outperformed conventional ECG interpretation by pediatric cardiologists (accuracy 67.1%, sensitivity 71.6%).

CONCLUSIONS

This study highlights the potential of AI-assisted ECG analysis for CHD screening in young children. The ResNet-18-based model outperformed conventional ECG evaluation, suggesting its feasibility as a supplementary tool for early CHD detection. Future studies should focus on multi-center validation, inclusion of more CHD subtypes, and integration with other screening modalities to improve diagnostic accuracy and clinical applicability.