A comparison of artificial intelligence–enhanced electrocardiography approaches for the prediction of time to mortality using electrocardiogram images

Figures

• 

  Figure 1. (A) Overview of the current standard approach for artificial intelligence–enhanced electrocardiogram using natively digital electrocardiogram signals and two novel approaches used for application of artificial intelligence–enhanced electrocardiogram to images of electrocardiograms. (B) Depiction of synchronous 10-s electrocardiograms (typical for natively digital electrocardiograms) and asynchronous 2.5-s electrocardiograms (typical of electrocardiogram images).

• 

  Figure 2. (A) Example images demonstrating electrocardiogram image resolutions evaluated. (B) Image 2D convolutional neural networks were trained at each resolution, to predict time to mortality. Performance as measured by the C-index was assessed for each model. Error bars indicating 95% confidence interval are too small to be seen; values are reported in Supplementary material online, Table S1.

• 

  Figure 3. Saliency maps were plotted to demonstrate regions of the electrocardiograms most important for model predictions. Electrocardiogram signal and in particular leads aVR and V1 are highlighted in these example electrocardiograms.

• 

  Figure 4. Artificial intelligence–enhanced electrocardiogram models were trained to predict mortality using natively digital electrocardiograms, 2D image electrocardiograms, and digitized electrocardiograms. Models were trained and evaluated in different five configurations as shown on the x-axis. The natively digital 1D convolutional neural network trained on 10-s natively digital electrocardiograms and applied on 2.5-s digitized electrocardiograms performed the worst, with the other four models having broadly similar performance. Importantly, the model trained on 2.5-s asynchronous natively digital electrocardiograms performed equivalently to the model trained on 2.5-s asynchronous digitized electrocardiograms, when both were evaluated on digitized electrocardiograms. Performance as measured by the C-index was assessed in the hold-out test set for each model. Error bars indicating 95% confidence interval are too small to be seen; values are reported in Supplementary material online, Table S2.

• 

  Figure 5. Natively digital 1D convolutional neural network and image 2D convolutional neural network models were externally validated in the CODE and SaMi-Trop cohorts. The no-grid 2D electrocardiogram image format was evaluated as an additional image format. Performance as measured by the C-index was assessed for each model. Error bars indicating 95% confidence interval are too small to be seen in CODE; values are reported in Supplementary material online, Table S4.