Improving the accuracy of cancer prediction by ensemble confidence evaluation

This paper discusses a novel approach for the prediction of breast cancer, melanoma and cancer in the respiratory system using ensemble modeling techniques. For each type of cancer, a set of unequally complex predictors are learned by symbolic classification based on genetic programming. In addition to standard ensemble modeling, where the prediction is based on a majority voting of the prediction models, two confidence parameters are used which aim to quantify the trustworthiness of each single prediction based on the clearness of the majority voting. Based on the calculated confidence of each ensemble prediction, predictions might be considered uncertain. The experimental part of this paper discusses the increase of accuracy that can be obtained for those samples which are considered trustable depending on the ratio of predictions that are considered trustable.