Mulesa P. Medical data mining based on hybrid neural networks

The goal of thesis is synthesis of hybrid neuro-fuzzy systems for solving tasks of effective information analysis and processing based on dynamical medical data mining which presented by multivariate tables and non-stationary nonlinear signals with local properties under a-priory and current uncertainty. The learning method for neuro-fuzzy compressor is proposed. In this architecture we used activation function with linear derivatives, that allows increasing speed of data processing and reduces computational methods realization. The learning-self-learning method of single layer classification-clustering neural network is proposed. Such method can process information both in supervised and unsupervised learning mode and allows solving the classification-clustering tasks with fuzzy clusters in on-line mode. Multilayered diagnosis neural-network system based on Takagi-Sugeno-Kang approach with additional non-linear diagnosis layer is proposed. The learning method based on pattern recognition criterion is modified. Such system is characterized by increasing learning rate and simplicity of computational realization. Medical information preprocessing method for structuring of factors space in diagnostic medical tasks based on fuzzy decision trees, model of multi-criterion choice and fuzzy logic is improved. This method allows to provide the most important factors ranging.