Evidence combination in medical data mining

In this work we apply a formal evidence combination technique for mining medical data. The mining task is classification of skin lesions and breast cancer cases. The goal is to combine beliefs from different sources of evidence, namely different classifiers, to arrive at a final classification. Dempster-Shafer theory takes into consideration evidences in the form of mathematically evaluated beliefs. It also considers uncertainty associated with beliefs. The evidences considered here are the beliefs obtained from three classifiers: k-nearest neighbor, Bayesian and Decision Tree. The classifier uncertainty based on its discriminative power is computed dynamically. Dempster's rule of combination combines the beliefs to arrive at a final decision. Our evaluation shows more accuracy than classifications based on individual beliefs. We study the circumstances under which the evidence combination approach improves classification.