Risk Prediction In Cardiovascular Epidemiology: Considerations for Modeling Composite Endpoints

Abstract

This thesis reviews construction and evaluation of risk prediction scores in the context of cardiovascular disease. We give an overview of the clinical guidelines that make use of cardiovascular disease risk scores as well as current practices in recalibration of risk scores outside of the development population. Because cardiovascular disease is a composite of heart attack and stroke, we summarize approaches to modeling and recalibrating composite endpoints and give our major reservations about the clinical utility of such a composite endpoint and the ability to recalibrate such an endpoint. A method of predicting individual outcomes is proposed that is roughly statistically equivalent to the current practice in terms of calibration and discrimination but increases useful information for clinicians in that, for a given composite risk, risks for the component outcomes can vary greatly. Knowledge of these component risks should influence clinical recommendations. Limitations are discussed and next steps are suggested, largely in terms of simulating competing risks survival data and comparing the performance of recalibration methods for composite endpoint whose component risk models differ substantially.