Intra-institutional prediction of outcome after cardiac surgery: comparison between a locally derived model and the EuroSCORE

OBJECTIVE: To construct models for predicting mortality, morbidity and length of intensive care unit (ICU) stay after cardiac surgery and to compare the performance of these models with that of the EuroSCORE in two independent validation databases. METHODS: Clinical data on 4592 cardiac surgery patients operated between 1992 and 1996 were retrospectively collected. In order to derive predictive models and to validate them, the patient population was randomly divided into a derivation database (n=3061) and a validation database (n=1531). Variables that were significant in univariate analyses were entered into a backward stepwise logistic regression model. The outcome was defined as mortality within 30 days after surgery, predefined morbidity, and the length of ICU stay lasting >2 days. In addition to the retrospective database, the models were validated also in a prospectively collected database of cardiac surgical patients operated in 1998-1999 (n=821). The EuroSCORE was tested in two validation databases, i.e. the retrospective and prospective one. Hosmer-Lemeshow goodness-of-fit was used to study the calibration of the predictive models. Area under the receiver operating characteristic (ROC) curve was used to study the discrimination ability of the models. RESULTS: The overall mortality in the retrospective and the prospective data was 2 and 1%, and morbidity 22 and 18%, respectively. The created predictive models fitted well in the validation databases. Our models and the EuroSCORE were equally good in discriminating patients. Thus, in the prospective validation database, the mean areas under the ROC curve for our models and for the EuroSCORE were similar, i.e. 0.84 and 0.77 for mortality, 0.74 and 0.74 for morbidity, and 0.81 and 0.79 for the length of intensive care unit stay lasting for 2 days or more, respectively. CONCLUSIONS: Our models and the EuroSCORE were equally good in discriminating the patients in respect to outcome. However, our model provided also well calibrated estimation of the probability of prolonged ICU stay for each patient. As was originally suggested, the EuroSCORE may be an appropriate tool in categorizing cardiac surgical patients into various subgroups in interinstitutional comparisons. Our models may have additive value especially in resource allocation and quality assurance purposes for local use.