Detection of shockable ventricular cardiac arrhythmias from ECG signals using FFREWT filter-bank and deep convolutional neural network

Abstract

Among various life-threatening cardiac disorders, ventricular tachycardia (VT) and ventricular fibrillation (VF) are shockable ventricular cardiac arrhythmias (SVCA) which require immediate defibrillation therapy for the survival of patients. Timely and accurate detection of rapid VT or VF episodes using ECG signals is extremely important before initiating external defibrillator (AED) and implantable cardioverter-defibrillator (ICD) therapies. In this paper, a novel approach for the detection of SVCA using ECG signals is proposed. The fixed frequency range empirical wavelet transform (EWT) (FFREWT) filter-bank is introduced for the multiscale analysis of ECG signals. The modes evaluated using FFREWT of ECG signals are used as input to a deep convolutional neural network (CNN) for the detection of SVCA. The architecture of the proposed deep CNN comprises of four convolution, two pooling, and four dense layers. The ECG signals from various public databases are used to evaluate the proposed FFREWT domain deep CNN approach. The results show that the proposed approach has obtained an accuracy of 99.036%, 99.800%, and 81.250% for the classification of shockable vs non-shockable, VF vs Non-VF, and VT vs VF, respectively using 8 s ECG frames with 10-fold cross-validation (CV) strategy. Our proposed approach has obtained an average accuracy value of 97.592% using 8 s ECG frames with subject-specific CV. The hardware implementation of the proposed SVCA detection approach can be done using an Internet of things (IoT) driven patient monitoring system.

Introduction

Smart healthcare systems are the collection of medical devices, sensors, services, and applications that connect and communicate through the internet [1], [2]. Such frameworks have helped to provide quality healthcare and handle the constant increase in the demand for healthcare services [3], [4]. One such application is the development of a smart healthcare system for the accurate and real-time detection of life-threatening cardiac ailments [5], [6]. Rapid ventricular tachycardia (VT) and ventricular fibrillation (VF) are shockable ventricular cardiac arrhythmias (SVCA) whose treatments require immediate defibrillation therapy [7], [8]. VT is based on the presence of abnormal electrical activity in the ventricles of the heart, which causes the heart to beat faster than the usual and it is unable to pump sufficient blood to the lungs and entire body [9]. On the other hand, VF is a disorganized electrical activity generated in the ventricle of the heart. As a result, the heart fails to pump blood to other parts of the body which might lead to sudden cardiac arrest [10]. It has been found that approximately three hundred thousand deaths in the United States per year are due to SVCA [11]. The treatment of patients with SVCA is more complex than using an automated external defibrillator (AED), and it is performed according to clinical guidelines [12]. Typically, a high energy shock therapy is given to the patients with VF to restore their heart activity to normal sinus rhythm. Usually, a $200$ J energy-based shock is given by an AED device to the patients having VF episodes. [13]. Similarly, for implantable cardioverter-defibrillator (ICD) devices, a 20 J energy-based shock therapy is used [13]. In ICD and AED devices, the automated detection of VT and VF episodes from ECG signals is the primary task to initiate the shock therapy [14]. The extraction of features from the ECG signal and the use of machine learning approaches are the important steps in the development of automated detection of VT and VF ailments [8]. The accurate and timely detection of SVCA will be helpful in the initiation of shock therapy in automated patient monitoring.

In the last few decades, various feature extraction and classification approaches have been proposed for the detection and classification of SVCA episodes using ECG signals [15], [16]. The features namely, complexity measure (CPLX) [17], threshold crossing intervals (TCI) [13], VF filter leakage measure (VFF) [18], and the auto-correlation function (ACF) [19] coupled with various machine learning based classifiers have been widely used for the automated detection of VT and VT ailments using the ECG signals. Similar works include, the extraction of features based on the spectral algorithm (SPEC) [20], phase space representation (PSR) [21], and wavelet transforms [22], [23], [24], [25]. Moreover, a few authors have developed automated SVCA detection algorithms based on the combinations of aforementioned features with machine-learning techniques [26], [27] [28], [29], [30]. Furthermore, the time–frequency analysis based method has also been used for the detection of SVCA using ECG signal [31]. In recent years, authors have used the signal-driven multiscale analysis approaches such as empirical mode decomposition (EMD) and variational mode decomposition (VMD) for the decomposition of ECG signal into various modes, and extracted features from those modes for the detection of VT and VF episodes [8], [32], [33], [34], [35]. These multiscale analysis methods have demonstrated higher classification performance for the detection of SVCA. However, the high computational complexity of these aforementioned multiscale analysis methods make them infeasible for smart healthcare applications.

In recent years, deep learning approaches have been applied for the automated detection of cardiac pathologies using ECG signals [36], [37], [38]. The convolutional neural network (CNN) based deep learning approaches have also been used for the detection of SVCA using ECG signals [16], [39]. The advantage of deep learning-based approaches is that they do not rely on handcrafted features and automatically extract learnable features directly from the raw ECG signals [37]. The empirical wavelet transform (EWT) is a multiresolution signal processing approach used for the extraction of modes from the ECG signal [37], [40], [41]. In EWT, the adaptive wavelet functions are used to design the filter-bank for the evaluation of modes [42]. The filter-bank in EWT is derived based on the segregation of the Fourier spectrum of the non-stationary signals by detecting the boundary points. The Fourier–Bessel domain EWT with fixed order ranges has been used for the analysis of multi-lead ECG signal for the detection of myocardial infarction [37]. However, the Fourier–Bessel domain EWT is found to be computationally expensive compared to the EWT of ECG signals [37], [43]. For the decomposition of non-stationary signals, the EWT filter-bank can be developed using fixed frequency ranges. The novelty of this paper is the development of an automated SVCA detection approach using fixed frequency range EWT (FFREWT) filter-bank. The important contributions of this paper are as follows:

(i) The FFREWT filter-bank is introduced for the decomposition of ECG signal into various modes.

(ii) A novel deep CNN architecture is proposed for the detection of SVCA using various modes of ECG signal.

(iii) The performance of the proposed multiscale deep CNN approach is compared with existing SVCA detection techniques.

The remaining sections of this paper are organized as follows. In Section 2, a description of the proposed SVCA detection approach is given. The results obtained and the discussion of the results are presented in Section 3. Finally, the conclusion of this paper is summarized in Section 4.