Pattern recognition of normal, lightly and heavily calcified bioprosthetic valves implanted in the mitral position

The performances of two pattern recognition methods for detecting and quantifying valvular degeneration by spectral analysis of the first heart sound were evaluated in 48 patients with a porcine valve in the mitral position. Twenty patients had normal valves, 13 had lightly calcified valves, and 15 had heavily calcified valves. The pattern recognition methods were designed to classify the valve status into three subclasses: normal, degenerated with light calcification (scale 0-2), and degenerated with heavy calcification (scale 3-8). One method was based on a three-class supervised learning approach, while the other was based on a two-node decision rule using a two-class supervised learning approach at each node. The performance of the classifiers based on the Bayes rule and the nearest-neighbor algorithm was expressed in percentage of correct classifications (%CC). Results show that the two-node decision rule provides higher performance (85%CC) than the direct three-class classifiers (75%CC)