Nearest neighbor method using non-nested generalized exemplars in breast cancer diagnosis

Every year there are several million people in the world who die from cancer while breast cancer is belonging to the most prevalent cancers diagnosed in women. In this paper, a nearest neighbor method which uses non-nested generalized exemplars is analyzed for diagnosis of breast cancer. The aim is to improve its accuracy so that the severity of a mammographic mass lesion is predicted more accurately from BI-RADS attributes and the age of the patient. The improvement consists in a change of distance computation between attributes with missing values and the use of several exemplars in diagnosis for a patient. Experiments on mammographic mass data make use of 10-fold cross-validation where sensitivity, specificity and overall accuracy are computed. Achieved results show increases in the sum of sensitivity and specificity as a combined measure for minimization of life-threatening situations and costs. Overall, the amount of unnecessary biopsies is decreased in the analyzed method.