Modeling risk stratification in human cancer

Motivation: Despite huge prognostic promises, gene expression-based survival assessment is rarely used in clinical routine. Main reasons include difficulties in performing and reporting analyses and restriction in most methods to one high-risk group with the vast majority of patients being unassessed. The present study aims at limiting these difficulties by (i) mathematically defining the number of risk groups without any a priori assumption; (ii) computing the risk of an independent cohort by considering each patient as a new patient incorporated to the validation cohort and (iii) providing an open-access Web site to freely compute risk for every new patient. Results: Using the gene expression profiles of 551 patients with multiple myeloma, 602 with breast-cancer and 460 with glioma, we developed a model combining running log-rank tests under controlled chi-square conditions and multiple testing corrections to build a risk score and a classification algorithm using simultaneous global and between-group log-rank chi-square maximization. For each cancer entity, we provide a statistically significant three-group risk prediction model, which is corroborated with publicly available validation cohorts. Conclusion: In constraining between-group significances, the risk score compares favorably with previous risk classifications.