A Review of Automated Diagnosis of ECG Arrhythmia Using Deep Learning Methods

Introduction

Heart disorders, also known as Cardiovascular Diseases (CDs), are the leading causes of increased mortality rates. They are caused by a lack of blood in the coronary artery, which also supplies blood to the heart. CDs cause arrhythmia, or irregular heartbeats, and depending on the severity of the arrhythmia issue, sudden death can occur (Dahe & Sodini, 2014). The electrocardiogram/electrokardiogram (ECG/EKG) depicts the electrical activity of the human heart, and the ECG pulse morphologies reveal particular types of arrhythmia based on various heart diseases. Arrhythmia identification from the ECG signal that is precise and efficient can save billions of lives and save healthcare costs throughout the world (Chen et al., 2017). This prompted us to conduct a detailed ECG analysis review and respond in the form of a method model based on many phases in order to further describe and categorize the flow and relevance of each stage of ECG signal analysis. With the massive effects of an efficient ECG signal analysis on public health and the economy, providing a worldview of hardware and software components, as well as real-time data collection using wearable and portable technologies to analyze an ECG signal as a function of the stage-based method, is another motivating factor that led us to conduct this research (Luz et al., 2016).

Figure 1.

Arrhythmia detection process (a) Machine learning (b) Deep learning framework (Parvanesh et al., 2019)

The conventional CDs diagnostic paradigm relies on the medical history and clinical evaluations of each patient. Such observations are analyzed using a set of quantitative medical characteristics to identify individuals according to medical disease taxonomy. In many conditions, the conventional rule-based diagnostic approach is inefficient because of handling the large amounts of data sources, and it requires extensive analysis and specialized knowledge to reach appropriate diagnostic performance. The challenge will become increasingly acute in areas where medical specialists and clinical equipment are few, particularly in developing nations. Due to its simplicity and modest cost, analyzing an Electrocardiogram is the more used method for identifying cardiac arrhythmia. A large size of ECG data collected every day, at residence and in hospitals, could limit human operators/technicians from reviewing the data (Hannun et al., 2019). As a result, numerous strategies for totally automated arrhythmia identification or event selection for subsequent validation by human specialists have been presented. Conventional machine learning to deep learning and their combinations are among the machine learning-based approaches for ECG observation and interpretation (Parvaneh & Rubin, 2018). In conventional machine learning (ML) algorithms Fig. 1, input such as ECG and RR series data is identified to a collection of features such as morphological properties or any ECG characteristic, the presence of a P wave, vector cardiogram, entropy measures, vector cardiogram, Poincare section-based features, coefficients derived with wavelets and heartbeat interval (Parvanesh et al., 2019; Zhang et al., 2014; Parashar & Agrawal, 2020) and then a classification model is used to classify the data as a neural network, learning vector quantization, support vector machine (Parashar & Agrawal, 2021; Llamedo & Martínez, 2010). Expertise knowledge is used to identify features that are commonly depictions of cardiac arrhythmia. In recent years, conventional machine learning algorithms have made substantial progress. They depend on hand-crafted feature identification, which needs significant domain expertise and signal data preprocessing. Furthermore, due to the wide variation in wave shape across subjects and the background noise, computerized interpretation is hard to obtain high accuracy.