Phonocardiogram segmentation by using an hybrid RBF-HMM model

Abstract

This paper is concerned to the segmentation of heart sounds by using Radial-Basis Functions for acoustical modelling, combined with a Hidden Markov Model for heart sounds sequence modelling. The idea behind the use of RBF’s is to take advantage of the local approximations using exponentially decaying localized nonlinearities achieved by the Gaussian function, which increases the clustering power relatively to MLP’s. This neural model can be advantageous over the global approximations to nonlinear input-output mappings provided by Multilayer Perceptrons (MLP’s), especially when non-stationary processes need to be accurately modelled. The above described RBF’s properties combined with the non-stationary statistical properties of Hidden Markov Models can help in the detection of the T-wave which is fundamental for the detection of the second heart sound. The feature vectors are based on a MFCC based representation obtained from a spectral normalisation procedure, which showed better performance than the MFCC representation alone, in an Isolated Speech Recognition framework. Experimental results were evaluated on data collected from five different subjects, using CardioLab system and a Dash family patient monitor. The ECG leads I, II and III and an electronic stethoscope signal were sampled at 977 samples per second.