Data and Information Integration for Risk Prediction in the Era of Big Data

点击翻译按钮获取中文摘要

Abstract Toward precision medicine and precision disease prevention, the overarching goal of this proposal is to develop innovative statistical methods for accurate risk prediction. We address three challenges that plague studies on the value of candidate risk predictors that adds to established predictors for improved predictive accuracy: there is often a lack of independent validation data, the source population for the study sample and the target population of prediction are often different, no statistical methods are currently available for developing risk prediction models using individually-matched case-control data, and there is a lack of statistical methods for helping assess study feasibility beyond standard power calculation for testing predictor-outcome association. On the other hand, data and information that are external to the study may well exist and can be exploited to alleviate these challenges. For example, a model with only standard predictors often exists and has been validated, and the distribution of standard risk predictors in the target population of prediction is often available. We propose that external data and information can be exploited to address the above-mentioned challenges for candidate predictor evaluation, and develop innovative statistical methods to bring this idea to fruition. Considering prediction of a binary outcome, we propose a novel method to building logistic prediction models that are guaranteed to calibrate well in the target population, an innovative method for risk prediction with individually matched case-control data, and a method to project the added value of candidate predictors to help assess study feasibility. Our methods, accompanied by user-friendly software, will facilitate cost effective and timely predictor evaluation for predicting binary outcomes. Our methods were motivated by and will be applied to several PI Chen's collaborative studies.