Automated diagnosis methods for heart diseases using flexible analytic wavelet transform

Abstract

Cardiovascular diseases (CVDs) are the major cause of death globally. The global contribution of the CVDs is 30% of all deaths every year. The electrocardiogram (ECG) is the most popular non-invasive tool for detecting the heart diseases. The cardiologists perform the manual inspection of the ECG recordings for the detection of heart diseases. The accuracy (Ar) of the diagnosis depends upon skills and experience of the cardiologist. Lack of expertise may result in an inaccurate diagnosis. Therefore, the automated decision making methods are needed to help the doctors during the diagnosis and also can reduce their workload. Hence, we have proposed computer-aided automated methods for the diagnosis of di erent types of heart disorders in this thesis work. These methodologies are described below: Coronary artery disease (CAD) causes maximum death among all types of heart disorders. Therefore, we have proposed a new technique which can detect CAD automatically. In this method, the heart rate variability (HRV) signals are decomposed to sub-band signals using exible analytic wavelet transform (FAWT). Two parameters namely; K-nearest neighbour (K-NN) entropy estimator and fuzzy entropy (FEnt) are computed from the decomposed sub-band signals. Various ranking methods are also used for optimising the features. The proposed methodology has shown better performance using entropy ranking technique. The least squaressupport vector machine (LS-SVM) with Morlet wavelet and radial basis function (RBF) kernels yielded the highest classi cation Ar on the studied dataset. An automated method for the detection of CAD using ECG signals is also proposed. First, the ECG signals of normal and CAD subjects are segmented into beats. The cross information potential (CIP) is computed from the detail coe - cients obtained by FAWT based decomposition of ECG beats. For CAD subjects, mean values of CIP parameters are found higher than normal subjects. Further, the features are fed to LS-SVM classi er. We have observed signi cant improvement in the classi cation Ar up to the fourth level of decomposition. At the fth level of decomposition, no signi cant improvement is noticed in the classi cation Ar as compared to the fourth level of decomposition. Hence, the ECG beats are analyzed up to the fth level of decomposition. The Ar of classi cation is slightly higher for Morlet wavelet kernel (99.60%) than RBF kernel (99.56%).Myocardial infarction (MI) is a condition which can cause the death of the heart muscles. Therefore, we have developed a method for automated identi cation of MI ECG signals using FAWT. First, the segmentation of ECG signals into beats is performed. Then, FAWT is applied to each ECG beat to decompose them into sub-band signals. Sample entropy (SEnt) is extracted from each sub-band signal. We have achieved the highest classi cation Ar of 99.31% when SEnt is fed to the LS-SVM classi er. We have also incorporated Wilcoxon and Bhattacharya ranking methods and observed no improvement in the performance. A method for automated diagnosis of congestive heart failure (CHF) is proposed in this work. In this methodology, HRV signals of three di erent lengths (500, 1000, and 2000 samples) are analyzed using FAWT based decomposition. The accumulated fuzzy entropy (AFEnt) and accumulated permutation entropy (APEnt) are computed from the di erent combination of the obtained sub-band signals and ranked using the Bhattacharyya ranking method. Further, these ranked features are applied to the LS-SVM classi er. The proposed system has obtained the Ar of 98.21%, 98.01%, and 97.71%, for the 500, 1000, and 2000-sample length of HRV signals. Atrial brillation (AF) represents a condition of abnormal heart rhythm. Hence, a new approach for the detection of AF using FAWT is developed. First, the small segments of 1000 ECG samples are extracted from the long duration ECG signals. Then, the permutation entropy (PEnt) and log energy entropy (LEE) features are computed from di erent sub-band signals which are obtained using FAWT. The classi cation Ar of LEE features is observed better as compared to PEnt features. We have obtained Ar, sensitivity (Ss), and speci city (Sc) of 96.84%, 95.8%, and 97.6% respectively using LEE features with random forest (RF) classi er.