Consensus outlier detection in survival analysis using the rank product test

Abstract

Survival analysis is a well known technique in the medical field. The identification of individuals whose survival time is too short or to long given their profile, assumes great importance for the detection of new prognostic factors. The study of these outlying observations have gained increasing relevancy with the availability of high-throughput molecular and clinical data for large cohorts of patients. Several methods for outlier detection in survival data have been proposed, which include the analysis of the residuals, the measurement of the concordance c-index, and methods based on quantile regression for censored data. However, different results are obtained depending on the type of method used. In order to solve the disparity of results we proposed to apply the Rank Product test. A simulated dataset, and two clinical datasets were used to illustrate our proposed consensus outlier detection method, one from myeloma disease and the other from The Cancer Genome Atlas (TCGA) ovarian cancer. Finally, the Rank Product with multiple testing corrections was performed in order to identify which observations have the highest rank amongst the methods considered. Our results illustrate the potential of this consensus approach for the automated retrieval of outliers and also the identification of biomarkers associated with survival in large datasets.