A Multi-Center Trained Residual Neural Network for Robust Classification of Atrial High-Rate Episodes in Remotely Monitored Pacemakers and Defibrillators

Figures

• 

  Figure A1. The precision–recall curves (PRC) and average precision (AP) values of the one-dimensional residual network models (1D-ResNet10) for each class. For each fold of the fourfold cross validation, the model performance on the test set is shown in (a–d). The performance of the ensemble model on the external Saint Etienne test set is shown in (e).

• 

  Figure 1. Examples of atrial high-rate episodes found on the Biotronik remote monitoring website containing (a) true atrial tachycardia (AT/AF), (b) noise, and (c) far-field oversensing (FFO). The atrial (A) and ventricular (V) intracardiac electrograms are displayed in seconds ( Temp (secondes) ) before the device detects (Prédétection) an atrial monitoring episode (Classification Monit. A). Depicted atrial and ventricular markers (in French) on the marker channels are ‘TA’ for atrial tachycardia marker, ‘AS’ for atrial sensing marker, ‘Vp’ for ventricular pacing marker, and ‘Vs’ for ventricular sensing marker. Most markers are accompanied by the interval (in milliseconds) to the previous marker.

• 

  Figure 2. The model structure of the one-dimensional residual neural network (1D-ResNet10) used in this study.

• 

  Figure 3. The amount of atrial tachycardia and atrial fibrillation (AT/AF), noise, and far-field oversensing (FFO) episodes across the datasets of different centers.

• 

  Figure 4. The Grad CAM++ map of the four models from the fourfold cross validation (a–d) and the ensemble model (e) for an AT/AF episode predicted as noise. The Grad CAM++ map is shown for the normalized cardiac electrograms of the atrial (A-EGM) and ventricular (V-EGM) channel. Models 2 and 4 focus on atrial and ventricular activations with similar morphology, possibly to look for non-conducted atrial activations. Models 1 and 3 focus both on high-frequent atrial morphologies and classify the episode as noise. The ensemble model Grad CAM++ map shows the important parts of the model for all four models. Furthermore, the high probability for the noise prediction of model 3 is reduced by the probabilities of the other three models, which led to a probability of 59% for the noise prediction of the ensemble model.

• 

  Figure 5. Examples of the Grad CAM++ map of the ensemble model for three correctly classified episodes of (a) AT/AF, (b) noise, and (c) far-field oversensing (FFO). Parts of the normalized atrial and ventricular intracardiac EGMs that were most important to the model for its prediction are given a darker shade on the Grad CAM++ map.