Building cancer prognosis systems with survival function clusters. (15th March 2018)

• Record Type: Journal Article
• Authors: Muñoz, Johanna; Murua, Alejandro
• Is Part Of: Statistical analysis and data mining. Volume 11:Number 3(2018)
• Page Start: 98
• Page End: 110
• Publication Date: 2018-03-15
• Subjects: classification -- Gap statistics -- lung cancer -- Potts model clustering -- Runt Pruning -- survival analysis; Data mining -- Statistical methods -- Periodicals; 006.312
• DOI: 10.1002/sam.11373
• Languages: English
• ISSNs: 1932-1864

Abstract

Abstract : In oncology, risk groups are usually constructed by dividing the population in blocks of patients with similar health conditions and demographics levels. Even for a handful of factors, the number of risk groups may be large, which complicates the analyses. There is a need to cluster together homogeneous blocks of patients into larger entities with similar survival characteristics. We develop and compare several techniques to detect these patient meta‐blocks. Our prognosis systems are based on the integrated absolute distance between the survival functions associated with patient blocks. We propose the use of vectorization of survival curves and of principled ensemble algorithms for clustering. We test these methods on different complexity scenarios. The best performing methods are then used to create prognosis systems for the NCIC lung cancer database, a longitudinal study of the U.S. lung cancer patients who were followed from 1988 to 2009.