ECG Beat Classification Based on Deep Bidirectional Long Short-Term Memory Recurrent Neural Network

Abstract

Accurate electrocardiogram (ECG) heartbeats detection and classification are crucial for the diagnosis and treatment of heart disease. This paper proposed a simple and accurate prediction framework for the heartbeat classification based on wavelet transform and deep bidirectional long short-term memory (DBLSTM) recurrent neural network. The framework consists of three main steps. Firstly, in this work, we use biorthogonal wavelet transform to remove high-frequency noise and baseline drift; secondly, detecting the R wave peak by the modular maximum and minimum values of the binary spline wavelet transform and then extracting the QRS wave group on this basis; and finally, we propose to use DBLSTM to extract heartbeat feature automatically and classify ECG heartbeats. We have used the MIT-BIH arrhythmia database to obtain ECG time sequence data of five different heartbeat types. Experiments reveal that the accuracy rate reached 99.43% in the unbalance sample set. The prime contribution of this paper is the introduction of DBLSTM automatic extraction of ECG beats features, which significantly improves the prediction accuracy of our method.