Robustness of Residual Network in Predicting PR Interval Trained Using Noisy Labels

The PR interval represents the time required from the electrical impulse to advance from the atrium to AV node and His-Purkinje system until the ventricular myocardium begins to depolarize. PR interval prolongation has been associated with significant increases in atrial fibrillation, heart failure and mortality. Over the past years, multiple deep learning models have been proposed to interpret electrocardiogram $(ECG)$ signals. Despite initial success, these models are often trained and validated using datasets that contain partially incorrect labels. These “noisy” labels exist because of the way the annotated data was collected and pose challenges for model training and validation. As a result, a residual neural network (ResNet), trained on noisy data, was proposed to estimate $PR$ intervals. $In$ addition, an electrophysiologist performed a blinded manual adjudication on a stratified sample to validate the accuracy of both the model and the noisy labels. The conclusion is that a ResNet trained on noisy data can correctly estimate $PR$ intervals and outperforms the noisy labels it was trained on.