Comparison of spectral techniques for computer-assisted classification of spectra of heart sounds in patients with porcine bioprosthetic valves

Abstract

The diagnostic performance of two spectral techniques (the fast Fourier transform, FFT, and autoregressive modelling, ARM) combined with four windowing functions (rectangular, Hanning, Hamming, and sine-cosine) and two classifiers (Bayes and nearest neighbour) to detect valvular degeneration was evaluated in a group of 95 patients. Forty-seven patients had a porcine bioprosthetic valve inserted in the aortic position and 48 patients had a porcine bioprosthetic valve inserted in the mitral position. Among the aortic valves, 24 were normal and 23 were degenerated whereas among the mitral valves, 19 were normal and 29 were degenerated. The aortic and mitral valves were analysed separately. For each type of valve, 21 features were extracted from the spectra of the valve closure sounds to train and test the performance of four pattern recognition systems by using the leave-one-out method. The discriminant properties of all feature combinations between two and five among the 21 features selected were evaluated. Results show that the FFT combined to the nearest neighbour classifier provided the best performances: 87 per cent of correct classifications (CCs) for aortic valves when using the Hanning or the Hamming window and 94 per cent of CCs for mitral valves when using the rectangular window. The best performances obtained with the ARM were 81 per cent of CCs for the aortic valves (nearest neighbour classifier and the Hanning or the Hamming window) and 92 per cent of CCs for the mitral valves (nearest neighbour classifier and the Hamming window or the Bayes classifier and the Hanning or the Hamming window).