Ensemble for Medulloblastoma outcome prediction

A clinical challenge concerning medulloblastoma is the highly variable response of patients to therapy. Whereas some patients are cured by chemotherapy and radiation, others have progressive diseases. In this paper we show that a random subspace based ensemble of Fisher's linear classifiers permits to obtain a more reliable system for medulloblastoma outcome prediction than those proposed in literature. The dimension of the feature space is very high, hence we select a pool of features using a feature selection algorithm, we project this reduced space onto a lower dimensional space by Karhunen-Loeve feature transform. The random subspace based ensemble is built starting from the Karhunen-Loeve Space and not from the original space. The area under the ROC curve obtained by the best method proposed in the literature was 0.83, the system proposed in this paper obtains an area under the ROC curve of 0.91.