Sequential Application of Feature Selection and Extraction for Predicting Breast Cancer Aggressiveness

Breast cancer is a heterogenous disease with a large variance in prognosis of patients. It is hard to identify patients who would need adjuvant chemotherapy to survive. Using microarray based technology and various feature selection techniques, a number of prognostic gene expression signatures have been proposed recently. It has been shown that these signatures outperform traditional clinical guidelines for estimating prognosis. This paper studies the applicability of state-of-the-art feature extraction methods together with feature selection methods to develop more powerful prognosis estimators. Feature selection is used to remove features not related with the clinical issue investigated. If the resulted dataset is still described by a high number of probes, feature extraction methods can be applied to further reduce the dimension of the data set. In addition we derived six new signatures using three independent data sets, containing in total 610 samples.