Ensemble fusion methods for medical data classification

Medical data classification is acknowledged as an area of increasing importance, yet also poses many difficulties. One of these is that medical datasets are often imbalanced; that is that there are (potentially many) more samples of some classes compared to others. In this paper, a dedicated algorithm - Undersampling Balanced Ensemble (USBE) - is used to deal with this problem. We then conduct an experimental study to investigate the quality of different fusion methods for combining classifiers in an ensemble. Several fusion techniques based on discrete and continuous responses from (neural network) base classifiers are evaluated and it is shown that a careful choice of fusion method can boost the recognition rate of the minority class. In particular, a neural network trained fuser is shown to provide the best classification performance on two separate breast cancer datasets.