We work with part of the IRIDIA AF v1 database.

Here we represent two portions of an ecg, one of which is in Atrail-Fibrillation state and the other in normal state.

To do detection task, we use Residual Convolutional NN to analyze 8192-samples ECGs ( 1 = atrial fibrillation).

We get around 0.95 test accuracy.

We use two techniques to augment an ECG:

• Flipping
• Permutating

We experiment GAN with several architecture for the generator/discriminator. The best results (not so good) are obtained using a BiLSTM generator and a CNN discriminator.

Here we represent the losses for the GAN and an example of generated ECG after 100 epochs.