Improving Evoked Response Audiometry With Special Reference To The Use Of Machine Scoring

Abstract

The use of visual scoring of evoked response (ER) records to detect auditory threshold is reviewed, and it is concluded that current ER audiometry is not a reliable clinical tool because of substantial error compounded by subjectivity of score. The mathematics underlying ER detection is reviewed. The critical part played by the signal-to-noise ratio (f) and techniques for enhancing f are discussed. The accepted techniques of averaging, segregation by EEG classification before averaging, and filtering are considered, along with two newly proposed techiques, fast periodic stimulation and correction of records by predictors averaging. These hold promise of enhancing the effective f by a factor of about 3 and reducing the lenght of test runs by an order of magnitude. Self-normalization is introduced in order to define scores which are sensitive primarily to f rather than to these scores is relatively invariant from subject to subject since it depends on f alone. This simplifies normative studies, particularly for the no-ER condition for which f = 0, since estimates of type-II error, false positives, are independent of the subject. Five machine scores are defined: (1) normalized peak-to-trough amplitudes; (2) Normalized rectified mean amplitudes; (3) the ratio of power of the ER to that of the background EEG; (4) the correlation coefficients between subsets in the same run; and (5) a combined measure based on (3) and (4). The sensitivity of each of these measures and their dependence on f are considered. From the explorations thus far, it is concluded that machine scoring techiques are available which are at least as sensitive and reliable as visual scoring and which can be implemented by relatively simple, inexpensive special purpose computers. Furthermore, since machine scoring lends itself to a superior test strategy, it is concluded that machine scoring can greatly increase the sensitivity and reliability of ER audiometry. These conclusions apply without taking into account the vast improvements possible if the newly proposed techniques (fast periodic stimulation and correction by prediction) for enhancing f should prove successful.