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

Figures

• 

  Fig. 1. Flowchart illustrating the process of classifying infants and children into five groups based on the characteristics of congenital heart disease (CHD) and hemodynamic significance: normal heart structure (NOR), non-significant left heart disease (LHA), significant left heart disease (LHB), non-significant right heart disease (RHA), and significant right heart disease (RHB)

• 

  Fig. 2. Examples of 12-lead ECG and corresponding continuous wavelet transform time-frequency diagrams (upper limit of 2 × 105): (2 A) infant with normal heart structure with (2B) corresponding time-frequency diagram; (2 C) infant with right heart disease (tetralogy of Fallot) with (2D) corresponding time-frequency diagram, where leads V2 and V3 exhibit signals of the highst intensity; (2E) infant with left heart disease (patent ductus arteriosus) with (2 F) corresponding time-frequency diagram where leads V3 - V5 exhibit signals of highest intensity

• 

  Fig. 3. Process of training and classification: 10-second ECGs were first exported from MUSE in XML format for conversion into CSV format within a Python environment. The CSV files were imported into MATLAB to undergo continuous wavelet transform. Oversampling was performed by segmenting the initial ECG segments into fie 2-second intervals. The dataset was then split into a training set (80% of the ECG data) and a test set (the remaining 20%). Model training and 5-fold cross-validation were conducted using three pre-trained architectures: ResNet- 18, InceptionResNet-V2, and NasNetMobile. Performance was evaluated in terms of accuracy, sensitivity, specificity, F1 score, and area under the ROC curve (AUC)

• 

  Fig. 4. Heart defect and age distributions of patients. The cohort comprised examined 1,035 patients aged 0 to 5, including 234 (22%) with normal heart structure (NOR), 100 patients (10%) with non-significant right heart disease (RHA), 291 patients (28%) with significant right heart disease (RHB), 141 patients (14%) with non-significant left heart disease (LHA), and 269 patients (26%) with significant left heart disease (LHB). The age distribution was as follows: 0 years old (416 patients; 40.2%), 1 year (249 patients; 24.1%), 2 years (164 patients; 15.8%), 3 years (99 patients; 9.6%), and 4 years (107 patients; 10.3%)

• 

  Fig. 5. Boxplots illustrating 5-fold cross-validation results in predicting clinically significant CHD. The Nasnet- 18 model demonstrated the best detection performance across all metrics. InceptionResNetV2 also demonstrates strong performance, however it presented variability in sensitivity and F1-scores