Risk-prediction models for mortality after coronary artery bypass surgery: application to individual patients.

INTRODUCTION: We derived a risk-assessment model for predicting mortality after coronary artery bypass surgery from patient data from the 1990s and examined the model's accuracy in predicting early mortality in more contemporary patients. We also examined the accuracy of a completely new model and an older model recalibrated with newer data.
MATERIALS AND METHODS: Three mortality-prediction models were derived: an "old" model from 8959 patients treated during 1993-1999, a "new" model from 5281 patients treated during 2000-2004, and a version of the old model "recalibrated" with the 2000-2004 data. Each model's discriminatory ability was assessed by computing area under receiver-operator characteristic (ROC) curves, and precision was estimated by comparing observed and predicted mortality rates. To test the temporal applicability of the old model, we applied it to the 2000-2004 data and to data from 1879 patients operated on during 2005-2007. To compare the recalibration and remodeling strategies, the new and recalibrated models were applied to the 2005-2007 data.
RESULTS: The old model showed good discrimination (ROC, 0.80) and concordance between observed and predicted mortality for the 2000-2004 patients but overpredicted mortality for the 2005-2007 patients. The new and recalibrated models had good discriminatory ability (ROC, 0.81 and 0.79) and showed similarly good concordance between observed and predicted mortality when applied to the 2005-2007 data.
CONCLUSIONS: Predictive models for mortality after cardiac surgery lose precision within a few years after development. Recalibrating old models and creating new models (i.e., remodeling) are equally good strategies for predicting outcomes in contemporary patient cohorts.