Assessment Methods of Neural Network Classification Applied to Otoneurological Data

Neural networks are applied, among other things, to classify medical data to various disease classes. It is a great advance if data relevance and complicated relations between input data and output values (classifications) can thoroughly be analyzed. Nevertheless, basic techniques for neural network algorithms commonly disallow such chances. We developed methods to analyze such relevance and relations in detail and to outline the entire variable space formed by variables of a dataset and tested these methods with otoneurological data of vertiginous patients. We also treat the problem of using neural network classification in the case of a relatively small dataset that has a biased class distribution, as is usual for medical datasets. To overcome the problem we designed a set of multilayer perceptron neural networks instead of employing only one. This approach improved our classification results at least 10 % compared to our previous results with a single network. Our methods are of general type.