Time-frequency analysis of native and prosthetic heart valve sounds

In the past, a number of researchers have applied various spectral estimation techniques in an attempt to analyse recorded heart sounds. The majority of these studies have used spectral estimation algorithms such as the Fourier transform and various autoregressive modelling techniques. Despite the definite potential these techniques have shown for the diagnosis of valvular heart disease, they are limited by their assumption of signal stationarity and lack of relation to present stethoscope-based medical evaluation procedures. A solution to these limitations can be achieved by analysing the recorded sounds in the time-frequency domain rather than in the frequency-domain or time-domain independently. The research detailed in this thesis investigates the application of time-frequency techniques to the description and analysis of recorded heart sounds. Time-frequency is further investigated as a tool for the description of heart sounds in an attempt to diagnose valvular heart disease. Data used in the study was recorded from 100 subjects in four main valve populations. The four populations investigated were subjects with native heart valves, Carpentier-Edwards bioprosthetic heart valves, Bjork-Shiley metallic prosthetic heart valves and subjects before and after surgery for heart valve replacement. Prior to the analysis of these data sets, an investigation was performed into the suitability of various time-frequency techniques to the analysis of heart sounds. By comparing the shorttime Fourier transform, wavelet transform, Wigner distribution and the Choi-Williams distribution it was found that the Choi-Williams distribution provides definite advantages over the other techniques due to its high resolution and reduced interference properties. Applying the Choi-Williams distribution to typical examples of each data set demonstrated that timefrequency offers definite potential as a heart sound descriptor. Typical results also demonstrate that time-frequency can be used as an aid to understanding the origins of heart sounds. The work is concluded with the development of a classification scheme designed to diagnose valve condition in the native and bioprosthetic valve populations. Classification was performed using morphological features extracted using the Choi-Williams distribution and via an optimised feature set extracted using the discrete wavelet transform. Classification results suggest that the time-frequency analysis of recorded heart sounds can be used successfully for the identification of valve condition and position of dysfunction.