Multiple Biomarkers Algorithms for Optimal Disease Prediction

• Activity Number: 420
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Ping Xu* and Jeffrey Krischer
• Companies: University of South Florida and University of South Florida
• Keywords: biomaker evaluation ; Borda Count ; AUCROC

Abstract

Biomarker evaluation plays an important role in medical research. For many diseases,single biomarkers do not have good prediction for the associated disease outcome. Statistical methods for combining multiple biomarkers are needed. We propose a novel approach to combining multiple continuous biomarkers to optimize the overall prediction accuracy without assuming the distribution of the predictors. For all predictors, we replace the original values of biomarker with their rank positions in the sample. Using Borda count technique, a composite score is assigned to each subject based on their individual biomarkers' ranks. The performance of this score is assessed by the area under the receiver-operator characteristics curve(AUCROC), a summary measure of the discriminatory ability of a single continuous diagnostic biomarker for binary disease outcome. We obtain an AUCROC for each possible combination of biomarkers and rank the prediction performance based on these AUCROCs. The combination of biomarkers with largest accuracy gain over the single biomarkers thus can be utilized for overall prediction. It is illustrated with an application to data from a type 1 diabetes biomarker study.