Model updating methods, such as logistic recalibration, offer an efficient alternative to de novo model development for maintaining the predictive performance of cardiac surgery risk models over time.
Risk prediction models in cardiac surgery tend to lose their predictive performance over time. This statistical primer aims to provide an overview of updating methods with their strengths and weaknesses. This is important, as model updating may be an efficient and good alternative to the de novo development of risk models. The discussed methods are intercept recalibration, logistic recalibration, model revision, closed test procedure and Bayesian modelling. It is recommended to report an updated model according to the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) statement and to include calibration and discrimination plots of the original and updated models to assess the model performance. An example is provided for updating the EuroSCORE II model in a national cohort from the Netherlands. Logistic recalibration results in a significant improvement of model performance, without the risk of overfitting. The example illustrates that more data allow for more extensive updating methods.
Siregar et al. (2018) studied this question.