DYNAMIC PREDICTION OF SURVIVAL DATA USING SINGLE OR MULTIPLE LONGITUDINAL MARKERS

Recurrent events and time-to-event data occur frequently in longitudinal studies. In large clinical trials with survival endpoints, researchers collect a multitude of longitudinal markers. There is a growing need to utilize these rich longitudinal information to build prediction models and assess their prognostic performance. In this dissertation research, I propose a novel approach of integrating longitudinal markers in modeling the recurrent event or terminal event data, and conduct dynamic prediction of event risks. Under joint a model framework, I jointly model a longitudinal outcome and a recurrent event process with the two process correlated via shared latent function. The probability of having a new occurrence of recurrent event in a given time interval is predicted based on subject-specific longitudinal profile and disease history. When multivariate longitudinal outcomes are considered, traditional joint model method has limitation on specifying ap propriate longitudinal structures and computation problem occur when using Bayesian approach. To avoid these potential issues, I employ multivariate functional principal component analysis approach which is more flexible, robust and time efficient. For terminal event data, I specify a prognostic model incorporating multivariate longitudinal information, the prediction can be updated with accumulated data over time. I also propose a recurrent event model integrating multiple longitudinal markers and conduct personalized dynamic prediction of new recurrent event risk, which helps physicians to identify patients at risk and give personalized health care