MIRFCS: An Ensemble of Filter Methods for Classification of Disease Data

Abstract

A facility for early disease diagnosis is very important in the medical field. Since most disease datasets are huge in dimension, an automatic diagnosis process through computing devices becomes complex and time-consuming. Feature selection methods can be used to eliminate unnecessary information from a dataset. Among the existing filter-based feature selection methods, one filter may end up eliminating important features. This is where our ensemble-based filter methods of Mutual Information, ReliefF, and Chi-Square come into play, where even if one filter eliminates an important feature, another filter may compensate for it. This has been done by forming a union of the reduced set produced by each filter. The three datasets on which the evaluation has been done are PID, DLBCL and Prostate cancer. From the union, accuracies are calculated using different classifiers and the classification accuracies of 97.18%, 97.32% and 91.90% are achieved in the three datasets, respectively.