Validation and calibration of the ERASL-pre and post risk scores Results of a Dutch and Japanese investigation

In most medical centers the use of duration models is now part of everyday operation. Often only the relative risk information from a duration models is used to order patients and assign them to distinct risk groups. In contrast, applications from marketing research predominantly use the richer absolute risk predictions in the form of hazard rates and survival probabilities. Many of these applications can transfer to medicine, provided the predictions are reliability and accurate. Due to the high consequences of medical decisions external validation of proposed models is paramount before implementation. This paper elaborates and applies established external validation methods. Increased attention is given to the lesser known calibration to determine to what extend the absolute risk predictions match with observed data, and calibrate them if need be. The methods are applied on two new validation data sets with the aim to validate the recently published ERASL risk scores. It was found that the model overall orders patients moderate to well for the Okayama data set, and poorly in the Rotterdam data set. Furthermore, was found that the original model systematically over estimates recurrence free survival. This was corrected successfully by embedding the ERASL risk scores in a Weibull calibration model.