Page 99 - The-5th-MCAIT2021-eProceeding

Impact of Bidirectional LSTM Layer Variation on Cardiac

Arrhythmia Detection Performance

b
a
a
Shahab Ul Hassan , Mohd Soperi Mohd Zahid *, Khaleel Husain
a Department of Computer & Information Sciences, Universiti Teknologi PETRONAS, Seri Iskandar, Malaysia
b Institute of Health and Analytics, Universiti Teknologi PETRONAS, Seri Iskandar, Malaysia
* Email: msoperi.mzahid@utp.edu.my

Abstract

Cardiac arrhythmia is responsible for significant fatalities and is one of the critical cardiovascular diseases. Prediction of
arrhythmia at the right time can save lives and is of considerable importance to current researchers and clinicians. Deep
learning models are recently being used to analyze the electrocardiogram (ECG) signal and arrhythmia prediction.
Bidirectional long short-term memory (BiLSTM) algorithms are commonly used deep learning techniques for arrhythmia
classification. However, there is a lack of study that analyses the impact of BiLSTM layer variation on the performance of
the deep learning model for arrhythmia prediction. In this paper, the performance of BiLSTM algorithms for arrhythmia
classification with varying numbers of BiLSTM layers is analyzed. Specifically, the MIT-BIH arrhythmia dataset is used,
and the performance is measured in terms of accuracy, recall, precision, and specificity. Evaluating the performance of these
algorithms will aid in the creation of a useful BiLSTM model for arrhythmia prediction that uses the optimum number of
BiLSTM layers.

Keywords: Arrhythmia; BiLSTM; ECG; Accuracy; Specificity; Precision; Recall.

1. Introduction

Cardiovascular diseases (CVDs) are a major threat to human health because of their high morbidity and
mortality rates. The electrocardiogram (ECG) is a safe and effective, transthoracic diagnostic technique that
offers a wealth of details on diagnosing and treating cardiovascular diseases (Mukhopadhyay et al., 2012). It is
commonly used in tracking the operation of heartbeats. Arrhythmias normally occur in a particular situation
and are represented in the ECG by an irregular, sluggish, or a sign of a quick erratic heartbeat. The detection
and ECG arrhythmias signal classification may provide valuable medical diagnosis information based on the
different forms of arrhythmias related patterns (Lay-Ekuakille et al., 2013). Manually analyzing ECG features
is time-consuming and repetitive, and hence it is important to build an automated ECG analysis algorithm.
Several cardiac arrhythmia prediction algorithms have been proposed in the last few years.
Recently, ECG signal analysis has been successfully applied using deep learning-based approaches.
Deep learning architectures are usually made up of restricted Boltzmann machines, Deep belief networks
(DBNs) and deep Boltzmann machines (DBMs), convolutional neural networks (CNNs), and recurrent neural
networks (RNNs). The long short-term memory (LSTM) network is a type of RNN widely used to analyze
time-series data. It can efficiently preserve historical information and acquire long-term text dependency
information. ECG arrhythmia has been detected in the past using LSTM (Lay-Ekuakille et al., 2013, Hassan et
al., 2020, Husain et al., 2021, Oh et al., 2018, Hou et al., 2019). The bidirectional long short-term memory
(BiLSTM) algorithm is an extension of the LSTM network that uses a two-way learning approach. In this paper,
the performance of the BiLSTM-based model for ECG arrhythmia classification for the varying number of Bi-

E- Proceedings of The 5th International Multi-Conference on Artificial Intelligence Technology (MCAIT 2021) [86]
Artificial Intelligence in the 4th Industrial Revolution