Background: We compared the abilities of surgeons and of an established risk model to predict operative mortality after aortic valve replacement (AVR), and we investigated scenarios that give rise to discrepancies between these predictions.
Materials and methods: We reviewed all AVR procedures performed at a Veterans Affairs institution between 1993 and 2008 (n = 317). The abilities of the Continuous Improvement in Cardiac Surgery Program (CICSP) risk model and of the surgeons to predict operative mortality were assessed by computing the area under the receiver operating characteristic curve (AUC). We investigated cases in which there was a significant discrepancy (2-fold or greater) between the surgeons' and the CICSP model's predictions.
Results: The predictive abilities of both the surgeons and the CICSP risk model were good-AUC values were 0.73 and 0.75, respectively (P = 0.84)-but the surgeons' mean estimate of mortality risk (8.3% +/- 8.3%) exceeded both the CICSP model's estimate (6.6% +/- 8.3%) (P < 0.0001) and the actual mortality rate (5.4%). There was significant discrepancy between the two sources of prediction in 38% (122/317) of cases. In this subset of cases, the CICSP did not adjust for factors that influenced risk stratification by the surgeon in 33% (40/122) of cases; the most common of these factors were anticipation of a more extensive procedure, severe pulmonary disease other than chronic obstructive pulmonary disease, hepatic disease, and pulmonary hypertension.
Conclusions: Both surgeons and the CICSP model performed well in risk-stratifying AVR patients, but the surgeons tended to overestimate the risk. The CICSP model did not capture some disease entities considered relevant in estimating mortality by surgeons.
Published by Elsevier Inc.