Selection of Important Attributes for Medical Diagnosis Systems

Success of machine learning algorithms is usually dependent on a quality of a dataset they operate on. For datasets containing noisy, inadequate or irrelevant information these algorithms may produce less accurate results. Therefore a common pre-processing step in data mining domain is a selection of highly predictive attributes. In this case study we select subsets of attributes from medical data using filter feature selection algorithms. To validate the algorithms we induce decision rules from the selected subsets of attributes and compare classification accuracy on both training and test datasets. Additionally medical relevance of the selected attributes is checked with help of domain experts.