A Framework for Segmentation and Classification of Arrhythmia using Novel Bidirectional LSTM Network

Abstract

Segmentation of ECG to obtain significant and relevant features has been a significant and inevitable step in achieving dimensionality reduction in automated heart disease diagnosis systems. Reduction in mortality due to the cardiovascular issues rate can only be achieved through accurate and rapid classification. Nonstationarity and high variability augment the complexity of the detection process in the domains of time and frequency. These difficulties are further enhanced due to imbalanced and vague datasets. In this paper, we propose to use a deep learning module to tackle the imbalance in the datasets by applying recurrent neural networks using long short-term memory layers (LSTM) to classify ECG into two classes. It has been seen that LSTM networks can effectively extract sequential timing information in the input ECG samples. To remove the imbalance in the datasets, oversampling and focal loss-based weight balancing techniques have been used, which eventually enhance the accuracy of classification. The proposed approach, an LSTM network with oversampling technique, provides an accuracy of 99.54%, which is considerably better than the traditional approaches that yield an accuracy of around 98%. Moreover, this method is insensitive to the quality of the ECG signal due to the fuzzification process followed in the initial stages of processing the dataset. Deployment of the proposed method for bio-signal telemetry or pharmaceutical research to assist physicians in their work is the most promising advancement in this domain.