ECG Heartbeat Classification Using CNN
Mayank Chourasia, Anurag Thakur, Shresth Gupta, Anurag Singh
Abstract
Electrocardiogram(ECG) is a valuable clinical signal, which is widely used to identify the cardiovascular diseases. However, it remains a cumbersome process to manually evaluate the ECG signals because of smaller variations in its physiological features in normal and abnormal cases that too when there are a huge number of cardiac patients to examine. In such a scenario, automatic classification of ECG signals can provide an ease to the doctors to make a correct diagnosis of a particular disease. This work proposes a classification model to classify the ECG in five different classes based on their morphological features. Instead of using manually designed features as most of the existing ECG classification works do, we have extracted data-driven non-linear features using convolutional neural network. The 1D-CNN model architecture is based on three convolutional, max pooling and dense layers which automatically extracts distinguishable nonlinear features from the ECG signals and automatically classify them into five different classes: Non-ectopic beats (Normal Beat), Supraventricular ectopic beats, Ventricular ectopic beats, Fusion Beats and Unknown Beats. The proposed algorithm was assessed using open-source database of MIT-BIH, which is based on 47 subjects. After 5-fold cross-validation, the presented algorithm achieves an accuracy of 97.36% and f1 score of 99.83%. It is a simple yet fast performing model that is implementable on e-healthcare-based devices for remote heart diagnosis of patients.