A Deep Learning Based Assisted Tool for Atrial Fibrillation Detection Using Electrocardiogram

Abstract

Atrial fibrillation (AF) is a disorder related to the heart. Irregularity of RR intervals and lack of P wave are the two main indicators of AF. Detection of AF using Electrocardiogram (ECG) remains one of the real challenges in the field of medical science. In this paper, we propose Discrete Wavelet Transform based method coupled with Deep Learning methods such as 2 layer Long Short Term Memory (LSTM) along with Gradient Recurrent Unit (GRU), 2 layer Bidirectional Long Short Term Memory (BiLSTM) along with Gradient Recurrent Unit (GRU) are used separately to classify the ECG signal into 3 classes namely: Normal, AF and other rhythms. Physionet challenge 2017 dataset is used for the study purpose. The results of LSTM and BiLSTM are compared with Support Vector Machine (SVM). The result indicated that LSTM provided improved performance compared to BiLSTM and SVM methods. The class specific accuracy of normal, AF and other rhythm are 96.92%, 97.36% and 96.39% respectively and Area Under the Curve (AUC) is 0.982. The overall accuracy of LSTM network is obtained as 96.94%. The developed technology has immense applications in medical devices.