Ensemble-Based Selective-Voting Algorithm for Cancer Classification

• Activity Number: 359
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Chuanlei Zhang*+ and Radhakrishnan Nagarajan and Eric Siegel and Ralph Kodell
• Companies: University of Arkansas for Medical Sciences and University of Arkansas for Medical Sciences and University of Arkansas for Medical Sciences and University of Arkansas for Medical Sciences
• Keywords: Classification ; Genomics ; Cancer ; Selective voting ; Ensemble ; Convex hull

Abstract

There is much interest in using gene-expression data to classify cancer patients. However, sufficient predictive ability for clinical application has not been shown using traditional classification algorithms. Recently, a model-free ensemble algorithm has been proposed for classifying patients using high-dimensional genomic data. In this paper, the convex-hull structure of that algorithm is used to develop a new ensemble-based selective-voting algorithm. This new algorithm allows members of the ensemble to vote when test points fall inside reduced two-dimensional non-overlapping convex hulls defined by pairs of predictor variables. We study two different pruning methods to trim overlapping convex hulls to achieve separation of classes, and we investigate various numbers of bivariate regression models to select gene pairs as predictor variables. Only gene pairs for which either member does not appear in a higher-ranked pair based on the regression R-squared are kept as unique sets of potential voters. The algorithm is tested on a publicly available colon cancer dataset. Classification accuracy is shown to be improved using the new algorithm.