Prediction of Atrial Fibrillation using Deep Learning techniques

Atrial fibrillation (AF) is one of the most prevalent arrhythmias encountered in cardiological practice; it is a leading cause of stroke and heart failure and has a growing financial impact on the healthcare system. The aim of this study is to test the efficacy of deep learning in predicting AF. Early diagnosis can improve patients quality of life, reduce hospitalizations and minimize the financial impact on the health care system. Two methods for automatically predicting AF based on Convolutional Neural Networks (CNN) are proposed. The first model is characterized by a one-dimensional (1D) CNN network that takes, directly, electrocardiogram (ECG) signals as input, while the second model consists of a two dimensional (2D) CNN network that inputs image data corresponding to the numerical values of the most important events associated with each ECG waveform. The 1D-CNN model achieved an accuracy of 71.69%, while the 2D-CNN model reached 76.19%. The second method, in addition of being a novel approach to classification problem, reached a good score in predicting the outcome of atrial fibrillation.