Rule generation and model selection used for medical diagnosis

Abstract

In medical data analysis classification combined with rule generation is a common technique to obtain diagnosis results together with a rule based explanation. In this contribution we apply a neural network based rule generator in the domain of septic shock research. The septic shock is of special interest in intensive care medicine due to its high lethality rate. We describe the functionality of the neuro-fuzzy algorithm and present classification and rule generation results of our analysis. Because we repeated our analysis with randomly selected test data to calculate statistically valid mean results, we generated one neural network with different architecture for each repetition. To decide the important question which of the different models should be used in the application phase, we propose a useful method based on similarity measures for rules resp. rule sets to select one representative network out of the set of trained networks.