A novel deep neural network heartbeats classifier for heart health monitoring

Abstract

The electrocardiogram (ECG) is a very useful diagnostic tool to examine the functioning of the heart and to detect myocardial infarction (MI) and arrhythmias. It contains the records of the electrical signal of the heart and it is an investigation tool to check the heart's rhythm and thereby analyze heartbeats. Automatic detection of arrhythmia is possible by analyzing a patient's abnormal heartbeats and has become a major research area in recent years, as the manual examination of heart activity is time-consuming and prone to errors. Nowadays, the deployment of artificial intelligence (AI) - based algorithms to predict abnormal heartbeats categorized into five classes namely, non-ectopic (N), supra ventricular ectopic (S), ventricular ectopic (V), fusion (F) and unknown beats (Q) has drawn more attention in detecting arrhythmias. The use of intuitive hand-crafted features with shallow feature learning architectures is one of the key drawbacks of machine learning (ML) techniques. So, we present a novel deep neural network heartbeat classifier to extract and classify the heartbeat signals. The novel one-dimensional convolution neural network (1D CNN) model is developed by modifying the LENET architecture for the classification of heat beats (MIT-BIH Arrhythmia Database) and has attained an accuracy of 97.37%. This model's performance is also enhanced by the implementation of smote oversampling technique and gained an accuracy of 98.41%. Finally, the proposed model's performance is compared with other pre-existing models and various oversampling methods are deployed for analysis.

Keywords

• Convolutional neural networks
• Electrocardiogram
• MIT-BIH dataset
• Deep learning
• SMOTE
• Heartbeats