Lightweight Neural Networks for Automatic Classification of ECG Signals

Currently, most existing models for the automatic classification of electrocardiogram (ECG) signals demand high computation and memory capacity and thus are not applicable for edge communication scenarios with resources constrained devices. In this paper, we propose a lightweight two-dimensional convolutional neural network architecture (Ghost-VGG-16) for the accurate classification of five typical heart rates. Specifically, the proposed model follows the basic network structure of Visual Geometry Group(VGG)-16 and adopts the Ghost Module instead of the traditional convolution layer to generate more feature maps in terms of inexpensive linear operation, thus substantially reducing the parameters and calculations of the network model. To further simplify the pre-processing procedure, ECG signals are initially segmented into heartbeats and each heartbeat is then converted into a two-dimensional grey-scale image as the input data, which avoids noise filtering and feature extraction procedure. Experimental results show that the proposed method achieves an average accuracy of 99.74% at a lower computation and memory cost, which makes it more suitable for edge computing networks with performance-constrained edge devices.