Computer-aided Detection of Congestive Heart Failure based on Nonlinear HRV Features

The aim of this study is to present an automated system for accurate detection of congestive heart failure (CHF) patients. Considering the intrinsic nonlinear characteristics of HRV signal, its fractal behaviour and dynamical complexity in case of CHF and normal subjects are assessed. To this end, multi-fractal analysis and wavelet packet based complexity analysis is performed. A total of 538 features are extracted, from which the top ranked 15 features are selected based on Mann-Whitney U test and correlation coefficient measures. The classification performance of SVM classifier for multiple kernel functions, namely linear, polynomial, radial basis function (RBF), tanh, Sigmoid, and rectified linear unit (ReLU) functions are examined. On cross-validation using leave-one-subject-out technique, it is found that ReLU based SVM classifier have performed the best with an accuracy of 97.72%, sensitivity of 97.93%, specificity of 97.33%, and Mathews correlation coefficient of 0.949.