Ансамблі моделей неітеративного машинного навчання для аналізу біомедичних даних малих обсягів

Abstract

This dissertation addresses a significant theoretical problem of advancing the foundations of synthesizing neural-like ensemble structures based on non-iterative machine learning algorithm to improve the accuracy of solving regression and classification tasks on small-volume biomedical datasets. An analysis of existing methods for the intelligent processing of structured small-scale biomedical data is conducted, and the need for constructing neural-like ensemble structures using noniterative machine learning algorithm is substantiated. A novel approach is proposed for utilizing piecewise-linear approximation in the analysis of multidimensional response surfaces by means of isomorphic representation of each input feature as a set of local components. This allowed for nonlinear expansion of the input space and reduction of approximation error by 58% in RMSE using a linear non-iterative SGTM neural-like structure. A method for forming a combined response surface based on cluster analysis is improved, where the original feature space is extended solely by the cluster center’s target value. This minimized the increase in dimensionality while improving generalization and prediction accuracy. A method is developed for expanding the space of independent features in the dataset through the use of output signals from the summation layer of a Probabilistic Neural Network, which led to a 43% and 9% increase in classification accuracy (according to F1-score) compared to baseline artificial neural networks. A new methodology is developed for augmenting extremely small biomedical datasets using Cartesian squaring of the input vector components, which enabled the application of various nonlinear machine learning models, particularly SGTM neural-like structures with RBF-based input expansion. This approach improved regression accuracy (RMSE reduced to 59%). Methods for synthesizing ensembles of non-iterative linear and nonlinear neural-like models were proposed using piecewise-linear approximation, RBF-based input expansion, and symmetric random shifts, followed by aggregation of their outputs via a meta-model. A method for partial correction of systematic error in General Regression Neural Networks is developed using an ensemble of two such networks, achieving a 4–8% RMSE reduction. The method of global-local approximation was enhanced, yielding a 13% increase in accuracy. Additionally, direct and inverse error compensation methods for regression models were proposed based on rational function transformations, resulting in 9–11% reduction in MSE. All of the above-mentioned methods have been implemented in proprietary software tools protected by certificates of copyright registration. Their effectiveness has been experimentally validated using both synthetic and real biomedical data. The reliability of the obtained results is supported by comparative evaluation against existing machine learning methods and by successful approbation in a number of applied tasks in biomaterials science, rheumatology, traumatology, dentistry, and transplantology. The implementation was carried out in collaboration with medical institutions, academic organizations, and within the framework of international grant projects, which is confirmed by relevant implementation and utilization acts.