High Accuracy Arrhythmia Classification using Transfer Learning with Fine-Tuning

Abstract

A heart arrhythmia is an irregular heartbeat that occurs when the electrical signals synchronizing the heart’s beat do not work properly. Some arrhythmias are life-threatening and can result in heart failure and stroke. An electrocardiogram (ECG) is an important tool widely used to diagnose arrhythmia as it is a non-invasive method to assess signals and electrical activities of the heart. Deep learning techniques have been successfully used in the automatic detection and classification of arrhythmias. The insufficiency of ECG data is one of the major obstacles in training a deep neural network (DNN) for arrhythmia classification without over-fitting. In this paper, we propose a two-step methodology to detect and classify multi-class arrhythmia with high accuracy. The first step is the data pre-processing, and the second step is the detection and classification of arrhythmia using transfer learning with fine-tuning. In data pre-processing, high and low frequency noise in the data is eliminated using a low pass filter and baseline wander filter respectively and feature extraction is achieved using Daubechies Wavelet Transform. To implement transfer learning with fine-tuning we used a pre-trained EfficientNet B7 model. In this paper, we present multi-class classifications of arrhythmia such as 17-class, 15-class, 13-class, and 12-class classifications using 2-Dimensional (2-D) heartbeat images. The proposed approach achieved the highest average accuracy of 99.23% for 13-class classification. The performance of the proposed method is evaluated in terms of Accuracy, Recall, Precision, F1 Score, and Confusion Matrix.