Digitization and Linkage of PDF Formatted 12-Lead Electrocardiograms in Adult Congenital Heart Disease

Figures

• 

  Figure 1. Top 15 anatomic CHD diagnoses for 1409 patients. Summary of all the patients initially screened for inclusion. CHD, congenital heart disease.

• 

  Figure 2. Algorithm 1: ECG digitization pseudocode. The column output was derived from linked outpatient clinic letters. ECG, electrocardiogram.

• 

  Figure 3. Digitization and linkage of ECG PDF documents. (A) A sample ECG PDF document that uses vectorized graphic format to store ECG waveforms. (B,C) The results of the digitized leads (for leads V4 and V5). ECG, electrocardiogram.

• 

  Figure 4. ECG alignment using random 50 ECGs on lead I of patients with a Mustard procedure for transposition of the great arteries. (A) Individual nonaligned ECGs for the random 50 patients. (B) QRS aligned ECGs with the overall signal vector (outlined in bold) for the same random 50 patients. ECG, electrocardiogram.

• 

  Figure 5. Demonstration of a normal single-lead ECG trace, with P, Q, R, S, and T points. ECG, electrocardiogram.

• 

  Figure 6. Stratified patient leave-out setup. Each colour represents a different diagnosis. Each coloured circle contains all the ECGs of a single patient with the corresponding diagnosis highlighted with the colour. Each run represents a training and testing cycle and the SVM model trained from scratch in each run. For each run, 5 patients are randomly selected, and all the data for these patients are kept for testing, whereas the rest are used for the training, from run 1 to 100. ASD, atrial septal defect; ECG, electrocardiogram; PA, pulmonary atresia; SVM, support vector machine; ToF, tetralogy of Fallot.

• 

  Figure 7. Comparison of our digitization results of the vectorized graphic-based ECGs. (A) Original signals from the corresponding ECG document for leads V2 and V3. (B,D) Digitization results of Paper-ECG.30 (C,E) Digitization results of our algorithm for the same leads V2 and V3, respectively. ECG, electrocardiogram.

• 

  Figure 8. Bland-Altman plots between vendor and extracted values. PR interval, QRS duration, and ventricular rate are calculated, and Bland-Altman plots are plotted in (A), (B), and (C), respectively. Plots depict the agreement between the vendor calculated values on original signals and user calculated values on extracted signals. It also shows that there is no bias as the mean difference between 2 measurements is not consistently positive or negative.

• 

  Figure 9. Classification performance of the SVM model. (A) The average area under the curve (AUC) scores of all 100 runs for each class are displayed separately. (B) Confusion matrix of all 100 runs combined is displayed to highlight model’s prediction vs true label. (C) Calibration plots for each class are displayed separately. ASD, atrial septal defect; PA, pulmonary atresia; ToF, tetralogy of Fallot.

• 

  Central Illustration. Overall workflow from digitization to utilization in research. Selected set of ECG PDF documents are placed in an analysis file. Then, digitization is applied to extract all the leads and all the extracted data is stored for later use. Alignment and validation steps are also performed to ensure that there are no problems, and then all the data is matched with the corresponding output labels. After all these steps, it is now fully digitized and stored in a format that can be utilised in research.