CardioNet: An Efficient ECG Arrhythmia Classification System Using Transfer Learning

Abstract The electrocardiogram (ECG) is a noninvasive test used extensively to monitor and diagnose cardiac arrhythmia. Existing automated arrhythmia classification methods hardly achieve acceptable performance in detecting different heart conditions, especially under imbalanced datasets. This paper presents a novel method of heartbeat classification from ECG using deep learning. An automated system named ‘CardioNet’ is devised that employs the principle of transfer learning for faster and robust classification of heartbeats for arrhythmia detection. It uses pre-trained architecture of DenseNet that is trained on ImageNet dataset of millions images. The weights obtained during training of DenseNet are used to fine-tune CardioNet learning on the ECG dataset, resulting a unique system providing faster training and testing. The ECG dataset is prepared using augmentation process to provide a comprehensive learning of heartbeat morphology in the presence of intraclass variations. Two benchmark datasets of ECG recordings e.g., MIT-BIH arrhythmia and PTB are used to classify 29 types of heartbeats for arrhythmia classification. The proposed CardioNet system achieves higher classification accuracy of 98.92% outperforming other methods and shows robustness to different irregular heartbeats or arrhythmias.