Improving Dynamic Predictions from Joint Models of Longitudinal and Survival Data Using Time-Varying Effects

• Activity Number: 136 - Toward a Learning-Health System: Methods and Strategies for Data-Driven Health Care
• Type: Invited
• Date/Time: Monday, July 31, 2017 : 10:30 AM to 12:20 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Dimitris Rizopoulos*
• Companies: Erasmus University Medical Center
• Keywords: dynamic predictions ; joint models ; predictions error ; AUC ; personalized medicine

Abstract

Joint models for longitudinal and time-to-event data have become a popular framework for calculating individualized predictions. These predictions have a dynamic nature and update each time a new longitudinal measurements is recorded or we simply know that the patient was event-free. This feature makes them a valuable tool in the context of precision medicine. Considerable work has been done in the literature illustrating how these predictions can be calculated from joint models, and how to evaluate their accuracy. In this work we show that we can improve the accuracy of dynamic predictions by postulating a time-varying function to link the longitudinal outcome with the hazard. Flexible estimation of this function is achieved using P-splines that efficiently control the degree of smoothness. We base estimation on a Bayesian approach that allows for an automatic selection of the penalty parameter by a suitable prior specification of the parameters of the time-varying function. Simulation studies and a real data example from cardio-thoracic surgery illustrate that this extension improves the discriminative capability of joint models and reduces the prediction error.