Classification of Arrhythmia Signals Using Hybrid Convolutional Neural Network (CNN) Model

Abstract

Arrhythmia is a fatal cardiovascular disease that presents an excessively fast heartbeat, excessively slow heartbeat or an irregular heartbeat rhythm. Atrial fibrillation (AF) is a common type of arrhythmia that can be diagnosed using an electrocardiogram (ECG) pattern. Identification of arrhythmia through ECG can be very challenging because the process is highly dependent on experts and very time consuming. The use of deep learning in automatically assisting the detection of arrhythmia using one-dimensional (1-D) input is proposed in this study. Deep learning is preferred over standard neural networks because it facilitates training in an end-to-end manner and directly trains the classification system with raw signals. This study aims to investigate the performance of 1-D convolutional neural network (CNN) for arrhythmia classification and improve its performance by introducing a hybrid approach based on long short-term memory (LSTM) approach. Experimental data are obtained from PhysioNet CinC Challenge 2017 database. ECG signals are preprocessed via filtering, QRS detection, segmentation and median wave selection. One-dimensional CNN, hybrid CNN–long short-term memory (CNN–LSTM) and hybrid CNN–bidirectional LSTM (CNN–biLSTM) models are developed and evaluated in this study to classify ECG signals into (1) normal rhythm, (2) AF rhythm, (3) other rhythms and (4) noisy signal. Accuracies of the 1-D CNN, hybrid CNN–LSTM and 1-D hybrid CNN–biLSTM models were 91.67%, 82.33% and 94.67%, respectively. The experimental results showed that the proposed CNN models can aid in atrial fibrillation (AF) diagnosis for healthcare advancement.