Computational methods for cancer survival classification using intermediate information

We study a potentially useful methodology based on machine learning (ML) involving integration of separate biomarker classes, to improve prediction and separation of ovarian cancer survival times. We also imported intermediate survival information for separating extreme two groups. For prediction of survival phenotypes, we use four classifiers, first two existing machine learning methods (support vector machine, SVM; random forest, RF), the second a new regression-based method (REG) feature selection together with Cox proportional hazards model (FSCR), FSCR_REG, the third SVMbased classifier using FSCR data sets (FSCR_SVM). We compared these four methods using three types of cancer tissue features: i) miRNA expression, ii) mRNA expression, and iii) integrated miRNA and mRNA expression information, the latter with features selected separately from miRNAs and mRNAs profiles. The accuracies of survival classification using the combined miRNA/mRNA profiles are higher than those using miRNA or mRNA alone . The latter differences indicate sometimes strong interactions between miRNA and mRNA features which are not visible in individual analyses.