Neural Networks Applied to Retrocochlear Diagnosis

Methodologies have been developed, based on insights from signal detection theory, to evaluate quantitatively the diagnostic performance of tests. Several studies have demonstrated that, in fact, performance of a test battery can be inferior to the best of the tests it includes. These studies have been quite persuasive in damping enthusiasm for the test battery approach. Because the results of all tests in a battery were weighted equally in these studies, it is not surprising that an individual test with good sensitivity and specificity is more effective diagnostically than a combination of tests with poorer sensitivity and specificity. The authors of many of these studies were well aware of the limitations of this approach. In the present study, neural networks were applied to evaluate audiological tests used to predict retrocochlear pathology by differentially weighting the results of the tests in the battery. This technique avoids some of the limitations of previous approaches. Of the audiological tests evaluated in the present analysis, the superiority of the auditory brainstem evoked response (ABR) in predicting retrocochlear disease was again demonstrated. However, the results also demonstrated that identification accuracy could be improved by combining the ABR with other tests (in this case contralateral acoustic reflex at 2000 Hz, ipsilateral acoustic reflex at 2000 Hz, tone decay, and word recognition score). Further, it was demonstrated that performance could be improved over that obtained using dichotomous test measures (i.e., positive or negative presence of pathology) by using raw test measures in conjunction with ABR.