A Light Weight Deep Learning based Technique for Patient-Specific ECG Beat Classification

Abstract

An electrocardiogram (ECG) is an important medical tool in diagnosing different cardiac disorders. In general, the length of long-term ECG records is 24–48 hours, and it is not easy to analyse these records manually. Therefore, an intelligent computer-based automated tool is required to analyse these records. Deep learning methods have recently improved in identifying arrhythmias in complex ECG readings. Efficient usage of training data, complexity, and performance are the major challenges with deep learning algorithms. A combination of convolutional neural network (CNN) and bi-directional LSTMs (Bi-LSTMs) is proposed in this work for efficient training and classification. Additional training is made possible by bidirectional LSTMs (Bi-LSTMs) since they traverse the input data twice. Bi-LSTMs outperform standard unidirectional LSTMs because of their fixed sequence-to-sequence prediction and increased training capacity. The proposed work is validated on a standard, publicly available physionet arrhythmia database. The time required to detect the beat type by the proposed method is 0.52 (S). Experimental results reveal that the proposed method performs better than the techniques in the literature.