Artificial intelligence-enabled electrocardiogram to distinguish cavotricuspid isthmus dependence from other atrial tachycardia mechanisms

Figures

• 

  Figure 1. Electrocardiogram class examples. Example 12-lead electrocardiograms of the classes used to train the neural network. An electrocardiogram showing (A) non-cavotricuspid isthmus dependent atrial tachycardia (B) cavotricuspid isthmus-dependent atrial flutter.

• 

  Figure 2. Neural network architecture. Selected model architecture is shown. Our final network was a modified version of one used by an Attia et al.16 It is comprised of a total of seven convolutional layers with two fully connected layers. AT, non-cavotricuspid isthmus dependent atrial tachycardia; AFL, CTI dependent atrial flutter; Conv, convolutional layer; ReLU, rectified linear unit; F, number of filters; S, kernel size; MP, maxpool factor.

• 

  Figure 3. Study flowchart. Data from 288 patients were collected, 141 non-cavotricuspid isthmus-dependent atrial tachycardia cases and 147 cavotricuspid isthmus-dependent atrial flutter cases. Twelve-lead electrocardiograms were segmented into 5 second recordings. For the training dataset, multiple recordings were used per patient, and the test set used only one recording per patients. An ensemble model was created from the 10 models through cross validation. Model performance was evaluated on the unseen test set.

• 

  Figure 4. Model performance. (A) Confusion matrix demonstrating the accuracy of the model in classifying between the two groups. Cavotricuspid isthmus-dependent atrial flutter and non-cavotricuspid isthmus-dependent atrial tachycardia cases were correctly classified at 89.7% and 82.1%, respectively (B) Receiver operating characteristic curves for model performance; diagnosis of cavotricuspid isthmus-dependent atrial flutter was considered a positive case. The area under the receiver operating characteristic was 0.94.

• 

  Figure 5. Saliency maps of each electrocardiogram class. Saliency maps showing the most prominent sections of electrocardiogram recordings in influencing the output of the neural network for the electrocardiogram classes of (A) cavotricuspid isthmus-dependent atrial flutter, (B) Non-cavotricuspid isthmus-dependent atrial tachycardia. Darker red indicates higher saliency. The atrial activity in each case appears to be salient; this is consistent with both human interpretation and the underlying mechanism of the tachycardias.