Development and Implementation of an Efficient Deep Residual Network for ECG Classification

Abstract

In this paper, we propose an efficient deep residual network-based method to classify electrocardiogram signals using Web/edge devices with improved accuracy and CPU latency. The method classifies different irregular heartbeats according to the Association for the Advancement of Medical Instrumentation standard. We designed, trained, and validated an improved deep residual network using the MIT-BIH dataset. The dataset comprises five distinct classes: normal beat, supraventricular premature beat, premature ventricular contraction, the fusion of ventricular and normal beat, and unclassifiable beat. After training and validating our model, we incorporate quantization techniques using the TensorFlow library to reduce the size of the model. Due to a great reduction in model size using the techniques of quantization, we develop a Web application for online ECG monitoring. The model size is optimized to reduce it by more than 90% of the original size. The results obtained after the quantization technique in our designed convolutional neural network demonstrate an acceptable classification performance. The proposed method seems suitable for deployment in smartphones and Web applications. © 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.