Can Deep Learning Models Differentiate Atrial Fibrillation from Atrial Flutter?

Abstract

This study investigates the use of standard 12-lead ECG records from six largest PhysioNet CinC Challenge 2021 databases and a private database to differentiate Atrial Fibrillation from Atrial Flutter. Image-based and one-dimensional-based Deep Learning models were considered to perform the classification using different 1D and 2D Convolutional Neural Network architectures. For 1D CNNs, LiteVGG-11 was the architecture that best performed, with Acc 77.91 +/- 1.73%, AUROC 87.17 +/- 1.29%, F1 76.59 +/- 1.90%, Spe 71.69 +/- 4.73% and Se 86.53 +/- 5.33%. For 2D CNNs, EfficientNet-B2 was the architecture that best performed, with Acc 75.20 +/- 3.38%, AUROC 85.50 +/- 1.14%, F1 71.59 +/- 3.66%, Spe 74.76 +/- 13.85% and Se 75.74 +/- 13.85%. We showed that discrimination of AFib and AFlut is not trivial, and that one-dimensional signals had better performance than two-dimensional ones. Moreover, we demonstrate that the performance of our models only on the PhysioNet CinC Challenge 2021 databases are considerably lower than only on our private dataset.