Abstract P346: Modeling Novel Biomarkers with Individual Risk Factors and an “Improved” Bayes Approach are Preferred Strategies for Assessing Predictive Utility of Biomarkers

Abstract

Given the rapid growth of new prognostic biomarkers, it is critical to assess their incremental utility for risk prediction while considering standard risk factors. This assessment may be influenced by the approach used to model new biomarkers. We hypothesized that the performance of a putative biomarker is best assessed by adding it to a model that includes standard risk factors as individual variables, as compared to adding it to a composite risk score (based on standard risk factors) estimated from the current study or to a composite risk score from a published study. We also compared 3 approaches of adjusting the prior absolute risk of an event using the information from a new biomarker, when data regarding prior risk are limited, hypothesizing that conditioning the biomarker residuals on prior risk (Improved Bayes approach) or adjusting the intercept of a model that includes the prior risk estimate are superior to the Naïve Bayes approach. Incremental performance was evaluated by comparing measures of improvement in discrimination. Using 1000 simulated datasets, similar incremental performance was observed when a putative biomarker was added to a model with the individual risk factors as compared to adding it to a model with a risk score estimated from the current study. Including a biomarker in a model with a published risk score resulted in an overestimation of its contribution ( Table ).These findings were supported by Framingham Heart Study data predicting incident atrial fibrillation using CRP and BNP.The Improved Bayes approach was a better strategy for updating the prior risk estimate as compared to the Naïve Bayes approach, using information from a new biomarker (Table). Our theoretical and empirical results identified that adding a new biomarker into a multivariable prediction model that includes the individual risk factors is the preferred strategy for assessing the incremental yield of a novel biomarker, and using the Naive Bayes approach (when information on the prior absolute risk of an event is scarce) is suboptimal.