Classifying heartbeats from electrocardiogram signals using a siamese convolutional neural network

Abstract

The Electrocardiogram (ECG) is a low-cost exam commonly used to diagnose abnormalities in the cardiac cycle such as arrhythmias and problems in the heart’s muscle. With the advance of machine learning (ML) techniques in recent years, the automatic classification of ECG signals garnered interest in the scientific community. However, the process of annotating large and diverse datasets to support the training of ML techniques is still very time-consuming and error-prone. Thus, ML techniques whose training does not require a large, well-annotated datasets are becoming even more prominent. This means that underrepresented data in ECG datasets, like rare cardiologic disturbs can still be properly identified and classified. In this work, the use of Siamese Convolutional Neural Networks, popular in imaging classification problems, to classify 12-Lead ECG heartbeats is investigated. The early results indicate accuracy of up to 95% in a public dataset by using models composed of different combinations of similarity and loss functions. The class by class classification results are also compared with those of similar methods found in the literature, obtaining metrics on par and even exceeding them in the classification of some classes.