Approaches to diabetes data mining

This thesis examines the power of inductive decision trees, neural networks and support vector machines in a medical decision support system. We use diabetes data collected from the Diabetes Care Program of Nova Scotia and a data set from UCI Machine Learning repository. The influence of imbalanced data sets on classification is addressed. It is proved that support vector machines outperform the others in handling the skewed data set. We use improved association mining methods to discover knowledge about diabetes data. A filter is designed to prune the rule set generated by Apriori algorithm. We suggest to use association mining to optimize feature selection. We also suggest combining the usage of decision tree and association mining may take advantage of the goal-oriented character of decision trees and the more general results of association mining.