Visualization of Biomarkers for Multiple-Class Disease When Data Is of High Dimensions and Subject to Severe Collinearity: A Latent Structure Approach

• Activity Number: 338
• Type: Contributed
• Date/Time: Tuesday, July 31, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Su Qian*+ and Chris Andrews
• Companies: SUNY at Buffalo and SUNY at Buffalo
• Keywords: biomarker ; latent structure ; visulization ; risk minimization ; diagnostic function

Abstract

We present a visualization tool for identification of biomarkers of disease status in high dimensional data. It performs well even in the presence of multiple classes of disease and collinearity. Our strategy is to assume both the development of the disease and the changes in the predictors are driven by a latent structure. The latent structure is approximated by projection to latent structures (PLS). The connection between disease status and the latent structure his established through a classification function designed to minimize risk according to a general loss matrix. The connection between the predictors and the latent structure is established through the PLS loading matrix. We visualize the biomarkers by studying the simultaneous movements of disease status and predictors in response to changes in the latent structure. Biomarkers are identified from background noise using Otsu's threshold detection. A series of simulation studies demonstrate our model system works as designed.