A Lightweight Neural Network to Detect Arrhythmias

摘要

A lightweight deep learning algorithm called LTE Network was devised to automatically detect arrhythmias from original electrocardiograms (ECG) with small model size without sacrificing noticeable accuracy. The algorithm is based on a cascaded architecture that uses point-depthwise convolutions, which combine a pointwise convolution with a depth wise convolution to build a nine-layer lightweight convolutional neural network. Furthermore, we use an optimized loss function and Adam optimizer which minimize classification errors and alleviate vanishing gradient problem in the learning process. The experiments are conducted in original datasets of ECG signals coming from MIT-BIH ECG databases. It is contrasted with AlexNet and MobileNet, and the results confirm that the LTE Network outperform others on accuracy and efficiency.