An Approximation Cascade Scheme via Rational Fractions for Biomedical Data Analysis

• Main Authors: Izonin, Ivan, Tkachenko, Roman, Shcherbii, Ostap, Berezsky, Oleh, Krak, Iurii, Oliinyk, Maksym
• Format: Conference Proceeding
• Language: English
• Subjects: approximation; biomedical engineering; cascades; Computer science; Data models; direct approach; ensemble; fast training algorithms; Iterative algorithms; Machine learning; Machine learning algorithms; Mathematical models; rational fracture; Training
• ISSN: 2766-3639
• DOI: 10.1109/CSIT61576.2023.10324122

The accuracy of solving the approximation tasks in the case of analysis of large volumes of tabular biomedical datasets by machine learning methods is not always high. This is explained by the complex, non-linear dependencies inside a multi-parametric dataset of a large volume, which combines medical-biological and engineering-technical features. The task of increasing the accuracy of the machine learning algorithm, ANN, or a certain mathematical model, in particular for the case when they are a "black box" during the intellectual analysis of such data, is urgent. The paper proposed a new two-step method for the refinement of regression model parameters (a direct approach). It is based on the principles of approximation by rational fractions, which have many advantages compared to other known methods. Simulation of the developed method is implemented as a cascade of two machine learning algorithms. Experimental studies were carried out based on SGD and extended-input SGTM neural-like structure using a real-world large biomedical dataset. We show a significant increase in the approximation accuracy of the designed method in comparison with several known methods, including cascades.