Data based prediction of cancer diagnoses using heterogeneous model ensembles

Abstract

In this paper we discuss heterogeneous estimation model ensembles for cancer diagnoses produced using various machine learning algorithms. Based on patients' data records including standard blood parameters, tumor markers, and information about the diagnosis of tumors, the goal is to identify mathematical models for estimating cancer diagnoses. Several machine learning approaches implemented in HeuristicLab and WEKA have been applied for identifying estimators for selected cancer diagnoses: k-nearest neighbor learning, decision trees, artificial neural networks, support vector machines, random forests, and genetic programming. The models produced using these methods have been combined to heterogeneous model ensembles. All models trained during the learning phase are applied during the test phase; the final classification is annotated with a confidence value that specifies how reliable the models are regarding the presented decision: We calculate the final estimation for each sample via majority voting, and the relative ratio of a sample's majority vote is used for calculating the confidence in the final estimation. We use a confidence threshold that specifies the minimum confidence level that has to be reached; if this threshold is not reached for a sample, then there is no prediction for that specific sample. As we show in the results section, the accuracies of diagnoses of breast cancer, melanoma, and respiratory system cancer can so be increased significantly. We see that increasing the confidence threshold leads to higher classification accuracies, bearing in mind that the ratio of samples, for which there is a classification statement, is significantly decreased.