Automatic detection of auditory brainstem responses using feature vectors

The efficiency of several methods for the automatic detection of near threshold auditory brainstem responses (ABRs), based on feature vectors, was assessed on a sample of 374 recordings obtained from neonates. Features are quantitative descriptors of different aspects of the response commonly taken into consideration by expert ABR judges, such as: similarity with a template, reliability of the response, and intrinsic morphological characteristics. The classification methods used were: comparisons of individual features with a threshold, linear discriminant functions based on combination of features, artificial neural networks (ANNs) based on the raw ABR, and ANNs based on combinations of features. The network used was a modified version of the Fuzzy ART Map model. The accuracy of the classification into normal and abnormal (as previously scored by a panel of experts) was assessed with methods from signal detection theory. Our results showed that the approaches based on feature vectors (discriminant function and ANN) performed more efficiently than the ANN with raw data, or the individual features.