Prototype-based classification in unbalanced biomedical problems

Medical diagnosis can be easily assimilated to a classification problem devoted at identifying the presence or not of a disease. Since a pathology is often much rarer than the healthy condition, medical diagnosis may require a classifier to cope with the problem of under-represented classes. Class imbalance, which has revealed rather common in many other application domains, contravenes the traditional assumption of machine learning methods about the similar prior probabilities of target classes. In this respect, due to their unrestricted generalization ability, classifiers such as decision trees and Naïve Bayesian are not the proper classification methods. On the contrary, the basic feature of case-based classifiers to reason on representative samples of each class makes them appear a more suitable method for such a task. In this chapter, the behavior of a case-based classifier, ProtoClass, on unbalanced biomedical classification problems is evaluated in different settings of the case-base configuration. Comparison with other classification methods showed the effectiveness of such an approach to unbalanced classification problems and, hence, to medical diagnostic classification.