Prediction models are increasingly used in routine clinical practice. This is at least in part because of rapid increases in the range of prognostic information available to the clinician, and the consequent need to integrate that information to aid clinical decision-making. The conventional biostatistical approach to medical prediction is essentially static: a closed data set, typically from a single cohort, is analyzed using data at a specific time point in a patient’s disease trajectory, and the resulting model published in a journal. This model - supposedly true for all patients at all times – may then be subject to validation studies in other cohorts, with the results again published in a journal in the form of the model being “valid” or “not valid”.
The prediction approach best known to the general public is that of Netflix, an online movie rental company, that uses a member’s ratings of previously viewed movies to make predictions about how the member will enjoy other movies. A key aspect of the Netflix approach is that it is dynamic: predictions changes as the member changes (e.g. he or she starts to rate movies in a different way) and as more data become available (e.g. other members provide more ratings).
We have started to apply the concept of dynamic prediction in urology. For example, clinicians can access predictions about whether patients will eventually recover urinary and erectile function after prostate cancer surgery based on data available at various time points after surgery. Challenges and solutions to dynamic prediction modeling will be discussed including data management, informatics and evaluation.