Classification of Spectral Patterns Obtained from Eustachian Tube Sonometry

Abstract

Spectral patterns of sound transmission through the Eustachian tube (ET) have been obtained in a series of experiments designed to identify ET dysfunction. Previous studies of ET function using sonometry have relied on heuristic and somewhat arbitrary methods in interpreting the data. In this study, an automated classification algorithm was developed to separate these sonograms into three distinct groups. From a total of 150 sample spectra, 75 were used in the formation of learning sets. The remaining spectra were classified into these three groups using standard Bayesian techniques. Both parametric and nonparametric methods were applied in estimating conditional probability density functions. Results of classification are compared with an independent test of ET function. Agreement between our classifier and the results of the independent test was as good as 97.3 percent. The results of this study indicate that an automated classification procedure can effectively distinguish among the three major types of sonograms obtained from ET sonometry.