LIDS Webinar: Dynamic Prediction Methods

Prediction models in clinical settings are routinely developed using traditional, prospective study designs that define a baseline (origin) at which predictors are measured and from which to predict future risk. However, the increased availability and use of electronic health records and data registers for research purposes provide a large wealth of dynamic information collected over time, information that is directly related to disease status, progression, cure, and relapse. The hope is that such information can be used to inform and individualize predictions based on a dynamic assessment of a patient's characteristics: for instance, biomarker values and their dynamics could be predictive of future risk. Therefore, accommodating these time-varying features within a prediction model can enable dynamic predictions for updating the prognosis of a patient whenever new data is available. Several estimators have been proposed for the task of dynamic prediction, mainly from two approaches: joint modeling and landmarking. These approaches differ in terms of what information is used and how, underlying modeling assumptions, and computational complexity. In this short workshop, we will introduce the joint modeling and landmarking approaches for dynamic prediction, including clear definitions of risk estimators, various modeling strategies, and performance metrics. The two approaches are illustrated in practice using openly available observational data on heart function after surgery. Finally, state-of-the-art developments in the field are introduced and discussed as well.