Mining Medical Databases with Modified Gini Index Classification

Decision tree classification is a common method used in data mining. It has been used for predicting medical diagnoses. Among data mining methods for classification, decision trees have several advantages such as they are simple to understand and interpret; they are able to handle both numerical and categorical attributes. However, it is well-known that when Gini index is used for classification, the method biases multivalued attributes. In addition to having difficulty when the number of classes is large, the method also tends to favor tests that result in equal-sized partitions and purity in all partitions. In this paper, we modify the Gini-based decision tree method. To overcome the known problems, we normalize the Gini indexes by taking into account information about the splitting status of all attributes. Instead of using the Gini index for attribute selection as usual, we use ratios of Gini indexes and their splitting values in order to reduce the biases. We report our experiments with several benchmark medical data bases. Comparisons between our method and other known classification algorithms are provided.