The graphical recordings of electrical stimuli generated by heart muscle cells are known as an electrocardiogram (ECG). In cardiology, ECG is widely used to detect different cardiovascular diseases among which arrhythmias are the most common. Irregular heart cycles are collectively known as arrhythmias and may produce sudden cardiac arrest. Many times, arrhythmia evolves over an extended period. Hence, it requires an artificial-intelligence-enabled continuous ECG monitoring system that can detect irregular heart cycles automatically. In this regard, this chapter presents a methodological analysis of machine-learning and deep-learning-based arrhythmia detection techniques. Focusing on the state of the art, a deep-learning-based technique is implemented which recognizes normal heartbeat and seven different classes of arrhythmias. This technique uses a convolutional neural network as a classification tool. The performance of this technique is evaluated using ECG records of the MIT-BIH arrhythmia database. This technique performs well in terms of different classification metrics.
TopIntroduction
As per the report of the world health organization (WHO), cardiovascular diseases (CVDs) are the leading cause of death globally. In this report, it is stated that 17.9 million people die annually due to CVDs which constitutes an estimated 32% of all deaths worldwide (World Health Organization, Cardiovascular Diseases). For the diagnosis of CVDs, electrocardiogram (ECG) signals are widely used by cardiologists. It provides a non-invasive diagnostic method that is most preferable to medical experts. Abnormal heart rhythms generate irregular ECG waveforms which are collectively known as arrhythmias. There are different types of arrhythmias that evolve during an extended period of time. Early diagnosis of arrhythmias requires continuous monitoring of cardiac patients. For this purpose, Holter monitors are used which continuously records ECG signals for 24 to 48 hours (Jha & Kolekar, 2020, 2021). Manual analysis of these long-term ECG records is difficult for cardiologists. Hence, artificial intelligence (AI) enabled continuous ECG monitoring systems are essential to reduce the burden of cardiologists in arrhythmia detection. In these systems, machine learning/ deep learning-based methods are used to recognize normal heart rhythm and arrhythmias using ECG signal analysis. In Figure 1, the ECG waveform of a normal heart cycle is shown which has different components: P-wave, QRS-complex, T-wave, PR-Segment, QT-Segment, and ST-segment. Morphology of ECG waveforms changes in the case of arrhythmias which corresponds to abnormal cardiac activities. There are different types of arrhythmias which are broadly categorized into five classes as per the recommendation by the Association for the Advancement of Medical Instrumentation (AAMI).
These five classes of ECG beat as per AAMI are non-ectopic beat, supra-ventricular ectopic beat, ventricular ectopic beat, fusion beat, and an unknown beat. The class of non-ectopic beats includes normal (N), left-bundled branch block (LBBB), and right-bundled branch block (RBBB) ECG beats. The class of
Figure 1. ECG waveform of a normal heart cycle
supra-ventricular ectopic beat includes atrial-premature contraction (APC) beats while the class of ventricular ectopic beat includes premature ventricular contraction (PVC) beats. The fusion of normal and paced beat (FNPB) comes in the class of fusion beats while the fusion of normal and paced beat (FNPB) and paced (P) beats are associated in the class of unknown beat. The automated arrhythmia detection technique distinguishes different classes of arrhythmias based on different features of ECG beats which include time-domain features, frequency domain features, and statistical features.
In past, many arrhythmia detection techniques have been developed which are based on machine learning-based methods. Generally, machine learning-based methods use a support vector machine (SVM) (Khalaf et al., 2015; Qin et al., 2017), artificial neural network (ANN) (Li et al., 2017; Melin et al., 2014), probabilistic neural network (PNN) (Gutiérrez-Gnecchi et al., 2017), and ensemble classification tools (Mondéjar-Guerra et al., 2019; Rajesh & Dhuli, 2018). These machine-learning-based methods utilize different kinds of features which are used as inputs to the training model of the classifier. These features of ECG beats are extracted separately and among these features, efficient features are selected. The features extraction and selection steps are important in machine-learning-based methods which affect the classification performance. In recent few years, many deep-learning-based methods have been developed which are based on convolutional neural network (CNN), recurrent neural network (RNN), and long short-term memory network (LSTM). These are different architectures of deep-neural networks which extract features of ECG beats as weights of the network. Hence, features extraction and selection steps are not required separately in deep-learning-based methods. General methodologies of machine-learning and deep-learning-based techniques are discussed in the next section.