INVESTIGATION OF THE EFFICIENCY OF USING PNN FOR SOLVING THE PROBLEMS OF MEDICAL DIAGNOSTICS IN THE CONDITIONS OF THE ANALYSIS OF SMALL DATA OF HIGH DIMENSION

Abstract

Parkinson's disease is one of the illnesses that cause certain difficulties at the stage of diagnosis. Recently, there has been a tendency to increase the popularity of the use of artificial intelligence methods as an auxiliary diagnostic tool. A probabilistic neural network (PNN) has become widely used for solving problems in the field of medicine. Despite the rather high efficiency of its use for certain tasks, some aspects of its functioning remain insufficiently researched in practice. Existing scientific works do not pay due attention to the issue of using the optimal distance as a measure of similarity between objects. Calculating the distance between the current data vector and each reference sample vector is the first step in implementing a PNN. The classification accuracy of a neural network of this type depends on its efficiency. In this work, the effectiveness of using different distances in the algorithm of different implementations of a probabilistic neural network for detecting Parkinson's disease based on biomedical voice indicators in the case of small high-dimensional data was investigated. Experimental modeling of three different variants of PNN implementation was carried out using the following distances: Chebyshov, Manhattan, Minkowski, cosine, and Canberra. The results of the study showed different values of the F₁-measure when applying different distances. It was found that the use of the Euclidean metric in the structure of a probabilistic neural network is not always the best option. In particular, the application of non-Euclidean metrics provided a significant increase in accuracy for the analyzed dataset. This indicates the need for correctly selecting this parameter of the probabilistic neural network to obtain the highest accuracy when solving medical diagnosis problems.