Abstract

Atrial Fibrillation (AF) is the most dangerous arrhythmia for human health. The beating in the top two chambers is irregular (asynchronous and faster) because electrical impulses suddenly start firing in the atria, overriding the natural pacemaker of the heart. Consequently, patients with AF may require periodic check-ups that involve performing a standardized twelve-lead Electrocardiogram (ECG) exam. Therefore, to assist doctors with AF detection, conventional algorithms have been developed based on the irregularity of the R-R segment or detecting the P-wave absence. In addition, Convolutional Neural Networks (CNN) have been employed to detect the AF, showing promising results. However, the training of these networks requires large amounts of ECG data. Furthermore, the careful design and tuning of a deep neural network model consumes a considerable amount of time and computational resources. In this work, we employ the Inductive Transfer Learning (ITL) method and a ResNet18. This network is pre-trained with the public icentia11k database in a beat classification job. The fine-tuning stage was performed by merging the PhysioNet data challenges (2020, 2021) to configure a single dataset. We obtain an accuracy greater than 89% using a single II-lead. We conclude that by employing the ITL technique, a ResNet18, and a single II-lead is possible to classify the AF and AFL arrhythmias. Therefore, this work provides significant insight into the ITL and the reliability of employing a single lead to classify the AF. This technique may be useful in real-time AF classification when employing a single ECG lead in a mobile device.