Selecting Marker Interaction Using Maximum Entropy Conditional Probability Models for Partially Censored Survival Outcomes

• Activity Number: 604
• Type: Contributed
• Date/Time: Thursday, August 7, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Aotian Yang*+ and Qing Pan
• Companies: and George Washington University
• Keywords: maximum entropy ; lagrange multiplier ; nonparametric constraint ; censoring ; EM algorithm ; Buckley-James estimator

Abstract

We propose novel bioinformatics methods targeting at identifying biomarker interactions underlying longitudinal clinical symptoms. Specifically, we are interested in SNP combinations associated with proliferative diabetic retinopathy risks among type 1 diabetes patients. GWAS data from patients of the Epidemiology of Diabetes Intervention and Complication trials are employed. The proposed method combines the computation intensive maximum entropy conditional probability models (MECPM) with advanced survival techniques. MECPM assumes constraints in the form of combinations of several biomarkers and one phenotype, fits nonparametric conditional models under the maximum entropy dogma and selects a harmonious yet simple model using the minimum description length criterion. Furthermore, we expand the current MECPM algorithm, which is restricted to binary case-control outcomes, to longitudinal follow-up data. To impute the censored event times, Buckley-James method is employed in the screening, and EM algorithm is employed in the final selection. The proposed method considers the synergy between biomarkers as well as the longitudinal property of disease symptoms.