Wavelet-Based Relevance Vector Machines for Identification of Diseased Patterns in Plethysmographic Observations in Wrist Pulse

Abstract

The morphology of the radial pulse enshrouds several patterns which correlate to normal and diseased conditions in the human body. This study explores the use of the Impedance Cardiovasography technique to bare the existence of eight such morphological contours from plethysmography observations on the radial pulse. Wavelet-based parallel Relevance Vector Machine (mRVM) architecture with Gaussian kernel achieves the highest accuracy of 87.27% as compared with Cauchy and spline kernels and also with principal components of the morphological patterns in the higher order space using an assortment of similar kernels. The results of the genotype are ably supported by several statistical parameters, including the Matthews Correlation Coefficient (MCC), Generalized Correlation Coefficient (GCC) and Kappa Coefficient, which provide the basis for the better performance of wavelet mRVM in comparison to the PCA technique. The higher sparsity achieved due to the wavelet features due to the reduction in the hyperparameters by 30% seals the fate of wavelets as the ideal classifier for radial pulse data. The morphological patterns are observed in normal subjects and those with heart, liver and lung diseases.