Predicting audiometric status from distortion product otoacoustic emissions using multivariate analyses

Objectives: 1) To determine whether multivariate statistical approaches imp rove the classification of normal and impaired ears based on distortion pro duct otoacoustic emission (DPOAE) measurements, in comparison with the resu lts obtained with more traditional single-variable applications of clinical decision theory. 2) To determine how well the multivariate predictors, der ived from analysis on a training group, generalized to a validation group. 3) To provide a way to apply the multivariate approaches clinically. Design: Areas under the relative operating characteristic (ROC) curve and c umulative distributions derived from DPOAE, DPOAE/Noise, discriminant funct ion (DF) scores and Zest function (LF) scores were used to compare univaria te and multivariate predictors of audiometric status. DPOAE and Noise ampli tudes for 8 f(2), frequencies were input to a discriminant analysis and to a logistic regression. These analyses generated a DF and LF, respectively, composed of a linear combination of selected variables. The DF and LF score s were the input variables to the decision theory analyses. For comparison purposes, DPOAE test performance was also evaluated using only one variable (DPOAE or DPOAE/Noise when f(2) = audiometric frequency). Analyses were ba sed on data from over 1200 ears of 806 subjects, ranging in age from 1.3 to 96 yr, with thresholds ranging from -5 to >120 dB HL. For statistical purp oses, normal hearing was defined as thresholds of 20 dE HL or better, For t he multivariate analyses, the database was randomly divided into two groups of equal size. One group served as the "training" set, which was used to g enerate the DFs and LFs. The other group served as a "validation" set to de termine the robustness of the DF and LF solutions. Results: For all test frequencies, multivariate analyses yielded greater ar eas under the ROC curve than univariate analyses, and greater specificities at fixed sensitivities. Within the multivariate techniques, discriminant a nalysis and logistic regression yielded similar results and both yielded ro bust solutions that generalized well to the validation groups. The improvem ent in test performance with multivariate analyses was greatest for conditi ons in which the single predictor variable resulted in the poorest performa nce. Conclusions: A more accurate determination of auditory status at a specific frequency can be obtained by combining multiple predictor variables. Altho ugh the DF and;LF multivariate approaches resulted in the greatest separati on between normal and impaired distributions, overlap still exists, which s uggests that there would be value in continued efforts to improve DPOAE tes t performance.