Comparison of pattern recognition methods for computer-assisted classification of spectra of heart sounds in patients with a porcine bioprosthetic valve implanted in the mitral position

The diagnostic performance of two pattern recognition methods (or classifiers) for detecting valvular degeneration was evaluated in 48 patients with a porcine bioprosthetic heart valve inserted in the mitral position. Twenty patients had a normal porcine bioprosthetic valve and 28 patients had a degenerated bioprosthetic valve. One method was based on the Gaussian-Bayes model, and the second on the nearest neighbor algorithm using three distance measurements. Eighteen diagnostic features were extracted from the sound spectrum of each patient and, for each method, a two-class supervised learning approach was used to determine the most discriminant diagnostic patterns composed of six features or less. The probability of error of the classifiers was estimated with the leave-one-out approach. The performance of each method with respect to discriminating between normal and degenerated bioprosthetic valves was verified by clinical evaluation of the valves. The best performance in evaluation of the second spectrum (98% correct classifications) was obtained with the Bayes classifier and two patterns of six features each.