A Wrapper Based Feature Subset Evaluation Using Fuzzy Rough K-NN

Abstract — Application of data mining techniques on medical databases is a challenging task considering the high volume, complexity, and poor quality of the medical databases. Data mining in medical domain could greatly contribute in the discovery of disease associations and provide the physicians with valuable and previously unavailable knowledge. Among the hundreds or thousands of features in the medical databases only very few features predominantly contribute for medical decision making. The small subset of informative features, selected from a whole set of features, may carry enough information to construct reasonably accurate prognostic or diagnostic models. The objective is to find the optimal feature subset of the medical databases that could enhance the Accuracy, the Sensitivity and the Specificity of the classification algorithms. In this paper, a wrapper based feature subset selection approach with Fuzzy Rough K-Nearest Neighbor (Fuzzy Rough KNN) classification algorithm is used to select the discriminatory features of the Indian Liver Patient Dataset. The empirical results show that the feature selection approach could achieve a feature reduction of 70 % and enhance the performance of the classifier Fuzzy Rough KNN by 7%