An End-to-End framework for automatic detection of Atrial Fibrillation using Deep Residual Learning

Electrocardiogram (ECG) inspection is performed by expert cardiologists for diagnosing cardiac diseases such as atrial fibrillation that is ubiquitous in 1-2% of the population worldwide. Prolonged presence of atrial fibrillation tends to form blood clots that travel to the brain through blood stream and cause stroke that inevitably leads to death, making its detection of utmost priority. In the past, people have developed temporal and morphological features to tackle this problem but these features are prone to rhythm changes. Very recently, deep learning methods have shown remarkable performance for better ECG classification. Hence, we aim to develop an end-to-end framework for classifying different length ECG segments into four classes namely, atrial fibrillation, normal, other and noisy rhythms using a deep residual neural network thereby eliminating the need of handcrafted features. To make the model more robust towards noise, a data augmentation technique is employed. The proposed method produced an F-1 score of 0.88 +/- 0.02 on PhysioNet/Computing in Cardiology Challenge 2017 database, which is better than existing methods in the literature.