Introduction: The 12-lead electrocardiogram (ECG) is globally used as the mainstay diagnostic modality in the field of cardiology, as it can be used to diagnose a large variety of cardiac pathologies. Most built-in diagnostic algorithms in ECG devices require considerable manual correction, as the ECG interpretation is often incomplete or incorrect. We propose a deep learning-based ECG diagnosis algorithm that is able to recognize multiple abnormalities with high accuracy.
Methods: We designed a Resnet-like 1-D convolutional neural network (CNN), that distinguishes between nine different ECG-features, namely: normal sinus rhythm, atrial fibrillation, first-degree atrioventricular block, left and right bundle branch block, premature atrial and ventricular complexes, and ST-segment elevation and depression. To train the algorithm, we used publicly available training data from the PhysioNet challenge 2020. These training data consisted of 6,877 (3,699 male and 3,178 female) 12-lead ECG recordings lasting from 6 to 60 seconds per ECG, all sampled at 500 Hz. The data were randomly split into a training and a validation set, consisting of 6189 (90%) and 688 (10%) ECGs respectively. The resulting CNN was independently tested in the PhysioNet challenge on previously unknown data.
Results: On the validation set the CNN achieved an F2-score of 0.74, a G2-score of 0.52, and an accuracy of 0.95. On the test data by the challenge our algorithm achieved an F2-score of 0.74, a G2-score of 0.52, and a geometric mean of 0.62.
Conclusion: The developed algorithm is able to distinguish between multiple heart rhythms with a high accuracy of 95%.