Rule extraction based on rough set theory and its application to medical data analysis

Abstract

Knowledge discovery from databases is an important theme not only in medical data analysis but also in many other practical fields. Recently, rough set theory has been attracting many researchers' attention as an effective method for knowledge discovery. The main idea of rough set theory is to obtain rules which are as simple as possible from the given database by reducing the database while holding the original degree of consistency. To this end, two kinds of approximation sets to the original rough set are introduced: the lower approximation provides an inevitable rule, while the upper approximation provides a possible rule. In addition, a method is suggested for reducing the number of attributes while keeping the degree of consistency of the database. The aim of the paper is to apply such techniques to medical data analysis. Traditional rough set theory can treat only categorical data. Unfortunately, however, many medical data have continuous numerical values. In order to convert continuous numerical data into categorical data, we apply an ID3-like technique on the basis of information quantity. In addition, an idea for utilizing inconsistent data is suggested by defining the quality of boundary. This provides us with more information on which attributes are important, and simpler rules from databases. Finally, those techniques are applied for finding rules which cause MA (macroangiopathy) to NIDDM (Non-Insulin Dependent Diabetes Mellitus) patients.