We consider the setting where one or more longitudinal biomarkers have been measured in patients, and the objective is to obtain estimates of the probability of an event of interest within a given time window. This prediction probability is to be provided during follow-up of the patient, at a given prediction time point, taking into account the information about the biomarker(s) until that time (dynamic prediction). The event of interest may be subject to one or more competing risks.
Our aim in this presentation is to investigate the use of three methods for obtaining dynamic prediction probabilities with longitudinal biomarkers in competing risks, namely joint models, landmarking and revival modeling. These models differ with respect to underlying assumptions and computational complexity.
We apply these methods to data on patients with end-stage liver disease, on the waiting list for liver transplantation. The biomarker of interest is the MELD score, and the competing events are death, transplantation and removal from the waiting list. Methods will be assessed with respect to predictive accuracy, as measured by cross-validated Brier and Kullback-Leibler scores.