GenECG: a synthetic image-based ECG dataset to augment artificial intelligence-enhanced algorithm development

Figures

• 

  Figure 1. Synthetic ECG images recreated from the PTB-XL dataset. (A, B) Imperfection-free images. (C, D) The same images following the application of image degradation techniques to make it appear as though the images have been photographed. A, C have been recreated from 00074_hr_1R.dat. B, D have been recreated from 00067_hr_1R.dat.

• 

  Figure 2. Example of a synthetic ECG used in the initial Turing test with a summary of the qualitative feedback provided by healthcare professionals.

• 

  Figure 3. Confidence levels in ECG classification for Turing test round 3. (A, B) The number of responses for each option for (A) real-world ECGs and (B) synthetically created ECGs. (C) An area under the curve-receiver operating characteristic (AUC-ROC) curve examining the impact of confidence level on ability to correctly identify an ECG image as ‘real-world’ or ‘synthetically created.’

• 

  Figure 4. Performance of ECG-Dx image-based algorithm on abnormal images obtained from the PTB-XL dataset. Diagnoses along the y-axis represent PTB-XL labelled diagnoses. The label ‘unable to interpret’ corresponds to an output of ‘none’ from the image-based algorithm.

• 

  Figure 5. Area under the curve-receiver operating characteristic score curves for ECG images on the algorithm initially developed by Bridge et al.19 (A) The performance of the model on 400 test images obtained from a real-world image-based dataset,10 (B) the performance of the trained model on 43 test synthetic ECG images containing visual imperfections, (C) the performance of the trained model on the same synthetic ECG images following fine-tuning, (D) the performance of the fine-tuned model on the 400 test images from the real-world image-based. dataset, (E) the performance of both the initially developed synthetic model (orange) and fine-tuned model (blue) on 79 real-world ECG images containing visual imperfections. AUC, area under the curve; FPR, false positive rate; ROC, receiver operating characteristic; TPR, true positive rate.