A Novel Approach to Classify Electrocardiogram Signals Using Deep Neural Networks

Abstract

Atrial brillation (AF) is an abnormal heart rhythm that takes place when electrical impulses re o from multiple places in the atria (the top chambers of the heart) in a disorganized way. This causes the atria to twitch and results in an irregular heartbeat or pulse. Atrial brillation is a major cause of stroke. Atrial Fibrillation is usually screened manually with the help of Electrocardiodiagram (ECG) reading. Manually reading ECG is usually a tedious and time-consuming task, which is laden with human errors. Therefore, an automated process is quintessential. However, discerning anomaly in heart function using an e cient automated process has been a challenging task for quite some time. In this paper we propose two intricate Neural Network architectures for the classi cation amongst four types of heart condition-Normal, Atrial Fibrillation, Noisy Sinus Rhythm and Alternative Rhythm, using a dataset from PhysioNet/2017 challenge. Volunteers in PhysioNet/2017 challenge dataset came from diverse background and had a wide window of variation in their physical attributes, making the dataset su ciently reliable. Also, the size of this dataset exceeded any other before it on this topic, which further adds to the comprehensiveness of this dataset.Initially, a preprocessing is done on the dataset to make it more robust and push the accuracy to its edge. We then trained a Deep Neural Network, which combined feature extraction layers of CNN with long-short term memory (LSTM). Our method reached a summit accuracy of 91.19%.The second model just applied a CNN model and the resulting output reached an accuracy of 84.3%