Comparison of Conventional and Novel Risk-Prediction Methods Under High-Dimensional Setting

• Activity Number: 363
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Cathy Jenkins* and Dandan Liu and Sean Collins and Alan Storrow and Frank Harrell
• Companies: Vanderbilt University and Vanderbilt University and Vanderbilt University and Vanderbilt University and Vanderbilt University
• Keywords: risk prediction ; high dimensional

Abstract

Accurate risk stratification using prediction modeling is important in acute care settings as efficient resource allocation becomes critical. Conventional methods for developing prediction models include pre-specification, principal component analysis, and step-wise regression. When the scope of predictors greatly exceeds the sample size these methods do not always perform well. The 'pre-conditioning' method proposed by Paul et. al (Annals of Statistics, 2008) offers a potential alternative to classic methods for high dimensional settings. It separates risk prediction and model selection, utilizing the correlation structure of the predictors. Its strength is identifying a set of predictors that is consistent for increasing numbers of observations and predictors. We will compare these methods when developing a 30-day adverse event risk prediction model for acute heart failure patients in the emergency department (ED). The goal of this prediction model is to aid the physician in identifying patients safe for ED discharge. We will assess model performance based on Somer's Dxy, area under the curve (AUC) of Receiver Operating Characteristic (ROC) curves, and calibration curves.