Predicting Subject-Specific Outcome via an Optimal Stratification Procedure with Baseline Covariates

• Activity Number: 564
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Florence Yong*+ and Lu Tian and Sheng Yu and Lee Jen Wei
• Companies: Harvard and Stanford University and Harvard and Harvard
• Keywords: discretization algorithm ; predictability assessment ; 3-in-1 dataset modeling strategy ; clinically meaningful difference ; risk scoring system ; stratified medicine

Abstract

To predict a future subject's outcome with multiple baseline variables, one generally fits the study data relating the outcomes and covariates via a working model. Typically a continuous scoring system comprising of individual scores for outcome prediction can be obtained. However, these individual level prediction scores can be quite imprecise and difficult to translate into action. A common practice is to categorize subjects with similar scores by pre-determined grouping such as quartiles for stratified medicine. In light of a lack of a systematic approach to obtain a clinically meaningful stratification strategy, we propose an objective and interpretable procedure to optimally separate subjects into different groups with a desirable meaningful risk difference between them. We illustrate how to build and evaluate various risk scoring systems for more generalizable outcome predictions. Validation and statistical inference for the resulting actionable distinctive strata were illustrated using data from two clinical studies involving binary outcome and censored survival time. Vast applications for more personalized therapeutic strategy are envisioned.