Automated Coronary Artery and Congestive Heart Failure Disease Detection Using Nonlinear Features and On-Line Sequential Extreme Learning Machine

Abstract

Cardiovascular diseases (CVDs) are major reason of mortality in the world population, and the numeral of cases is up surging every year. The mortality rate due to coronary artery disease (CAD) and congestive heart failure (CHF) is higher than any other type of CVDs. The vast majority of cardiovascular disease fatalities occur in middle and low-income nations, including Ethiopia. According to the latest WHO data published in 2020, coronary heart disease deaths in Ethiopia reached 47,712 or 7.81% of total deaths. The age adjusted death rate is 112.44 per 100,000 of population and it ranks Ethiopia @112 in the world. Arrhythmias are a marker of the heart's abnormal activity and are linked to an increased risk of sudden cardiac death (SCD), one of the most crucial tests in cutting-edge cardiology. Many of the symptoms of CVDs can be relieved or avoided if they are detected and treated early. Consequently, the improvement of investigative procedures might improve the health of many individuals. As a result, the goal of this research is to automate the identification of CVDs. In this proposal, automated algorithms have been developed for the categorization of heart disease to identify cardiac arrhythmias using Heart Rate Variability (HRV) standard database and self recorded data.For investigation and detection of CHF and CAD, twelve non-linear attributes like correlation dimension (CD), detrended fluctuation analysis (DFA) variants DFA-α1 and DFA-α2, Bubble Entropy (BBEn), sample entropy (SampEn), dispersion entropy (DISEn), Lempel–Ziv complexity (LZ), sinai entropy (SIEn), improved multiscale permutation entropy (IMPE), hurst exponent (HE), permutation entropy (PE), approximate entropy (ApEn) and standard deviation (SD1/SD2) were retrieved from HRV signal. A feature reduction technique known as generalized discriminat analysis (GDA) has been used to reduce the dimension of these attributes. The reduced attributes have been normalized between 1 and -1 and then fed to online sequential extreme learning machine (OSELM) for classification and detection of CHF and CAD. For analysis of these cardiac diseases, the HRV signal was obtained from self recoded and standard database of MIT BIH of participants with normal sinus rhythm (NSR). The St. Petersburg Institute of Cardiological Technics data source provided the CHF and CAD databases used in this study. The numerical results have shown that GDA with Gaussian kernel function and OSELM with sine activation function achieved accuracy (AC) of 99.34% and sensitivity (SE) of 99.32% for NSR-CAD group, and AC and SE of 100% were achieved for NSR-CHF group. In addition, it has been observed that the algorithm's classification performance was improved with fewer blocks of data, and its generalization performance was excellent for detecting CHF and CAD.Additionally, the 1-norm extreme learning machine (1-NELM ) binary classifier suggested in the preceding part of the study was expanded for various data groups in this section, and the findings were elaborated utilizing ranking methods such as Fisher Wilcoxon, Entropy, Bhattacharya and receiver operating characteristicand reduction strategies such as GDA. For this study, the proposed method was validated using self recorded and standard database of MIT-BIH and SPICT. The considered dataset were grouped as CAD-CHF, YOUNG-CAD, YOUNG-CHF, ELDERLY-CAD and ELDERLY-CHF subjects. To diagnose disorders of the heart, such as CHF and CAD, linear discriminant analysis (LDA) and GDA are used as feature reduction techniques in conjunction with the 1-NELM binary classifier. The activation functions used with the classifier are Sigmoid, Hardlim, and RBF whereas LDA and GDA are used with the kernel functions such as Gaussian and RBF. Various nonlinear features were generated from HRV data and utilized in training and validation of proposed algorithm. The analyses were carried out numerically through the combination of database sets YNG-ELY, YNG-CAD and ELY-CAD subjects. The numerical results have shown that ROC with GDA and 1-NELM approach achieved an accuracy of 99.76±0.14, 99.87±0.12 and 100±0 for YNG-CAD, YNG-ELY and ELY-CAD groups respectively. The Fisher with GDA and 1-NELM; and Bhattacharya with GDA and 1-NELM approach achieved an accuracy of 100±0 for all considered datasets. The proposed method also achieved very good generalization performance with the smallest 1-Norm Root Mean Square Error (1-NRMSE) and less execution validation time as compared to support vector machine (SVM) and probabilistic neural network (PNN).Finally, this work presented an approach to detect health of CHF subject which is based on multiresolution wavelet packet (MRWP) decomposition method, attributes ranking approach, kernel principle component analysis (KPCA) and 1-norm linear programming extreme learning machine (1-NLPELM). For this investigation, the heart rate variability (HRV) signal has been decomposed up to 5-level using MRWP decomposition method. The sixty three log root mean square (LRMS) attributes were extracted from the decomposed HRV signal. The top ten attributes are selected by ranking approaches. The ten ranked attributes were then mapped to one new feature by KPCA and fed to 1-NLPELM. The simulation results demonstrated that Bhatacharya+KPCA with 1-NLPELM approach achieved an accuracy of 98.44±1.4%, 99.13±1.85% for NSR-CHF and ELY-CHF respectively. Out of all ranking methods,combined with KPCA+1-NLPELM provided the highest degree of accuracy for all datasets. In addition, the proposed method has also achieved very good generalization performance and less execution time as compared to 1-NLPELM, KPCA+PNN, KPCA+SVM, PNN and SVM.