Increasing the sensitivity of cancer predictors using confidence based ensemble modeling

This paper discusses the use of symbolic regression based ensemble modeling for obtaining more sensitive cancer predictors. The ensemble models are generated on the basis of blood parameters acting as model inputs which have been coupled with diagnosis data in order to predict breast cancer. In addition to previous works this contribution focuses on the use of ensemble predictors in order to achieve more sensitive models. For achieving this goal the best models in terms of accuracy, sensitivity and in terms of a combined measure are selected based on training data in order to analyze to which extent the more sensitive model behavior is also reflected on test data. In addition to the a-posteriori selection of ensemble models with certain properties first results are shown that have been achieved with a new evaluation function which favors more sensitive predictors and guides the search towards more sensitive models already in the model generation phase.