Dynamic and personalized prediction of complex cardiovascular events.

点击翻译按钮获取中文摘要

PROJECT SUMMARY Statistical learning methods such as random forests have proven useful in medical research. With the availability of massive biomedical and event history data collected during the course of diseases, dynamic and personalized risk prediction of future clinical events can provide valuable information to identify high-risk individuals and initiate timely treatments or interventions. Our application is motivated by the NHLBI Pooled Cohorts Study, where risk factors were measured intermittently at follow-up visits, and multiple cardiovascular disease (CVD) events could occur during the follow-up period. Existing statistical learning methods usually focus on time to the first event with baseline predictors; methods that can handle the second and subsequent clinical events or repeatedly measured time-dependent risk factors are lacking. We develop flexible random forest methods for multiple event data, where the complex event history information is fully utilized without pre-specifying the dependence structure of different events. The proposed methods can deal with the case where events are of different degrees of clinical importance and competing risks exist. The methodology will be applied to the pooled cohorts to build accurate risk prediction tools and to identify important risk factors for both CVD incidence and recurrence. We will conduct validation analysis to test whether novel statistical learning methods can outperform existing methods such as Cox-type models; we will also use forest models to provide guidance in building Cox-type models for CVD recurrence. The proposed research has the potential to advance dynamic and personalized risk prediction and to facilitate more effective prevention and treatment strategies for CVD recurrence.