Electrocardiogram Based Heartbeat Detection Using Deep Learning

Abstract

Cardiovascular diseases remain a leading cause of death worldwide, which results in important requirements for early and accurate arrhythmia diagnosis. This work proposes a novel design of automated heartbeat detection, which consists of a convolutional neural network and three-channel images using the electrocardiogram (ECG) signals. A combination of various preprocessing is applied for the elimination of interferences of the ECG signals such as band-pass filtering and wavelet transform for R-peak identification using a sliding window. Multimodal image fusion method is used to construct three-channel images from different grayscale images, which are transformed from the heartbeats by three transformation techniques namely Gramian angular field, Markov transition field, and Recurrence plot. Grid-search based optimization method in combination with 5-fold cross validation procedure are implemented for selection of the optimal hyper-parameters of the CNN models using the input three-channel images. The proposed algorithm including CNN models and MIF images is estimated the detection performance using 5-fold cross validation, which produces average accuracy of 99.63%, precision of 99.41%, recall of 99.52%, and F1-score of 99.64%. The relatively high performance of the proposed algorithm confirms the effectiveness for the arrhythmia recognition on the ECG signals.