This work compares the performance of new proposals for 1D and 2D convolutional networks in classifying cardiac arrhythmias into 2 classes, 5 classes and 8 classes. The database used is MIT-BIH, a very unbalanced dataset. The 1D networks use a set of 1600 samples of the ECG signal, while the 2D networks use gray-level images of 40 × 40 pixels. In order to expand and balance the MIT-BIH dataset, this work used a data augmentation technique based on shifting the position of the beat containing the arrhythmia in sets of 1600 ECG signal samples. The best results obtained on the test set for accuracy were 99.14% for the binary classification, 96.47% for the 5-class classification and 98.36% for the 8-class classification.