Classification of cardiac arrhythmias based on overlapped ECG image with lightweight neural network

Abstract

This dissertation presents the classification of different types of cardiac arrhythmias, including Atrial Fibrillation (AF), Normal Sinus Rhythm (NSR), Premature Atrial Contraction (PAC), and Premature Ventricular Contraction (PVC), using a lightweight neural network. A novel ECG data transformation method, referred to as transforming into overlapped ECG images, has been developed and utilized as input images for our neural network. During the transformation process, individual heartbeats within a 30-second time frame are cropped based on heart rate calculations. These resulting beats are then overlapped with respect to the x and y axis limits, generating distinctive images representing different arrhythmia types examined in this study. The lightweight neural networks employed in this research have been designed to be deployable on low-resource mobile devices due to their reduced network architecture parameters,. The performance of the developed neural network is evaluated using the long-term atrial fibrillation (LTAF) database from PhysioNet, which is clinically certified. This database comprises a total of 84 records, which consist of 24-hour duration for each record. The effectiveness of the proposed approaches is analyzed using confusion metrics, yielding an accuracy rate exceeding 98%. Furthermore, a demonstration Android application has been developed based on the trained model of the lightweight neural network, providing proof of concept for the potential applications of deploying lightweight deep learning neural networks on mobile phones.