Hybrid Classifier via PNN-based Dimensionality Reduction Approach for Biomedical Engineering Task
Ivan Izonin, Roman Tkachenko, Michal Greguš, Liliia Ryvak, V. V. Kulyk, Valentyna Chopyak
Abstract
The problem of producing medical and biological products in the case when such a product must be biocompatible is a complex and resource-intensive task. The conventional apparatus for studying alloys, from which such a product will be created, involves lengthy and expensive tests. Today, some traditional methods can be replaced by artificial intelligence tools. This paper solves the problem of predicting the properties of alloys for the manufacture of biocompatible products using machine learning. The authors developed a hybrid approach that involves a combination of PNN, Ito decomposition, and Logistic regression classifier. The first step of our approach is reducing the dimension of the input data space by replacing the initial inputs with a set of probabilities obtained by using PNN. The next step is the use of Ito decomposition and the application of the classifier using the final dataset. The proposed approach is tested using a real task. It has been experimentally established that the replacement of the initial inputs with the outputs of PNN provides a significant increase in the generalization properties of the Logistic regression classifier. Additional application of the Ito decomposition provides an increase in the classification accuracy. Comparison with other methods showed that our approach provides the highest classification accuracy based on different indicators.