Automated diagnosis of congestive heart failure using dual tree complex wavelet transform and statistical features extracted from 2 s of ECG signals

Highlights

• Identification of CHF using 2 s electrocardiogram signals is proposed.

• Features are extracted from dual tree complex wavelet transform coefficients.

• Extracted features are ranked using different ranking methods.

• Achieved 99.86% accuracy, 99.78% sensitivity and 99.94% specificity using KNN classifier.

Abstract

Identification of alarming features in the electrocardiogram (ECG) signal is extremely significant for the prediction of congestive heart failure (CHF). ECG signal analysis carried out using computer-aided techniques can speed up the diagnosis process and aid in the proper management of CHF patients. Therefore, in this work, dual tree complex wavelets transform (DTCWT)-based methodology is proposed for an automated identification of ECG signals exhibiting CHF from normal. In the experiment, we have performed a DTCWT on ECG segments of 2 s duration up to six levels to obtain the coefficients. From these DTCWT coefficients, statistical features are extracted and ranked using Bhattacharyya, entropy, minimum redundancy maximum relevance (mRMR), receiver-operating characteristics (ROC), Wilcoxon, t-test and reliefF methods. Ranked features are subjected to k-nearest neighbor (KNN) and decision tree (DT) classifiers for automated differentiation of CHF and normal ECG signals. We have achieved 99.86% accuracy, 99.78% sensitivity and 99.94% specificity in the identification of CHF affected ECG signals using 45 features. The proposed method is able to detect CHF patients accurately using only 2 s of ECG signal length and hence providing sufficient time for the clinicians to further investigate on the severity of CHF and treatments.

Introduction

Failure of heart in pumping the required amount of cardiac output due to impaired ventricles is one of the alarming symptoms of congestive heart failure (CHF) [60]. The main cause of CHF is the structural or functional cardiac disorder that impairs the left ventricle (LV) due to multiple clinical syndromes [49]. Structural and functional cardiac abnormalities such as LV hypertrophy (LVH) and LV systolic dysfunction affect the LV contraction ability in filling blood thereby reducing the cardiac output, the LV ejection fraction (LVEF) [11], [60], [9]. Thus, heart fails due to the structural and/or functional disorder which causes vascular congestion in meeting the blood circulation requirement of the body, leading to CHF [12]. It is estimated that approximately five million people are getting affected with CHF in the USA and predominantly in elderly patients of more than 65 years of age [39]. The magnitude of the problem cannot be precisely assessed. However, the projected cost of CHF in the US was 27.9 billion dollars in 2005 [39] and 33.2 billion in 2007 [50]. In the UK, CHF consumes almost 2% of National Health Service budget, most of the cost being linked to hospital admission [55]. Therefore, early identification of CHF can avoid further structural or functional damage and save life.

Electrocardiogram (ECG) is the primary noninvasive test routinely used by the clinicians for capturing the signs of CHF in the patients [44]. In addition to diagnosis of CHF, ECG is useful for the prognosis and management of patients diagnosed with CHF [28], [51]. Changes in the 12 lead ECG are generally not specific to CHF; however, certain ECG patterns particularly during severe CHF (with ischemic and non-ischemic cardiomyopathy) include intraventricular conduction delays, low amplitude QRS complexes because of multiple previous myocardial infarctions (MIs) and ventricular aneurysms [29]. In patient with CHF together with cardiomyopathy, a distinctive ECG pattern (the ECG-CHF triad), which is characterized by low voltages in the limb leads, high QRS voltages in the precordial leads, and an R/S ratio <1.0 (or slow R wave progression) in lead V4, have been observed [10], [21], [28]. Moreover, majority of CHF patients implant pacemakers, hence, a paced ECG patterns are often encountered [28]. Recently few studies reported of ST-segment depression and inversion of T-wave in the left sides leads, called as LV “strain pattern” in patients with CHF and LVH [28], [41]. In addition, decrease in QRS complexes [30], and P waves amplitude values [31], shortening of QRS duration [32], and QTc intervals [33] and peripheral edema (PEED) have been experimented and published by various researchers.

Visual observation of these ECG morphology variations and manual interpretations of characteristics are extremely significant in capturing the clues of CHF. Thus, the reliability of the manual ECG interpretation depends on the operator's skill and experience, even though studies claim not much specialized training is required to interpret [19], [53]. In addition, it is time consuming and tedious to handle the enormous ECG signals and its overlapping characteristics, thus, prone to errors. This necessitates the use of computer-aided techniques in order to tackle the study of enormous ECG signals [63].

Assessment of heart rate variability (HRV) [13], [15], [20], [24], [27], [3], [37], [38], [40], [45], [46], [54], [56], [58], [61] and ECG [14], [16], [17], [26], [35], [36], [42] signals and its characteristics using various computer-aided techniques have been explored by many researchers in order to detect the patient having CHF. Different time-domain [20], [45], [46], [61], frequency-domain [45], [46], [61], energy [15], [17], entropies [14], [17], [27], [3], [40], Poincare plot [20], [40], [56], [61], bispectrum [61], and fractal scaling exponents [16] features have been evaluated using different methods. The variations of those HRV and ECG signal features during normal and CHF conditions have been examined. Even though there are many contributions in the detection of CHF, techniques for accurate and early detection of CHF is still progress.

Thus, in view of an early detection of patients having CHF, this paper proposes a new algorithm using dual tree complex wavelet transform (DTCWT) method using short-term (2 s) ECG signal. In this work, ECG signal analysis using DTCWT combined with statistical features is developed, which is capable of capturing the CHF patients and classify them from normal ones in 2 s of ECG segment. The block diagram of proposed algorithm is displayed in Fig. 1. Six levels of DTCWT is performed on the normal and CHF ECG signals of 2 s duration and five different statistical features such as maximum ($M{x}_i^k$), minimum ($M{n}_i^k$), mean ($M_i^k$), standard deviation ($St{d}_i^k$) and median ($M{d}_i^k$) are extracted from the 24 DTCWT coefficients obtained. The extracted features are ranked using seven different ranking methods (Bhattacharyya, entropy, minimum redundancy maximum relevance (mRMR), receiver-operating characteristics (ROC), Wilcoxon, t-test and reliefF) and fed to decision tree (DT) and K-nearest neighbor (KNN) classifiers for automated classification.