Classification of Atrial Fibrillation and Cardiac Arrhythmias by a CNN-BiLSTM Hybrid Model with DWT Preprocessing

Abstract

This paper presents a novel approach for the classification of atrial fibrillation and cardiac arrhythmias using a hybrid model combining a convolutional neural network (CNN) and a bidirectional long-short-term memory (BiLSTM) recurrent network. Data pre-processing was carried out using discrete wavelet transform (DWT) to remove unnecessary frequency bands. Two databases, denoted St Petersburg INCART 12-lead Arrhythmia Database and MIT-BIH Atrial Fibrillation Database, were used to train and evaluate the model. The results indicate exceptional performance with a test accuracy of 99.84%. The five classes considered are normal class, supraventricular ectopic beats (SVEB) class, ventricular ectopic beats (VEB) class, beat fusion, and atrial fibrillation class, thus demonstrating the effectiveness of the proposed model in the accurate classification of cardiac arrhythmias.