Risk Factors for Common Complications Following Adult Heart Surgery

Abstract

Background and aims: Cardiac surgery carries a substantial risk of mortality and other postoperative complications. Morbidity following cardiac surgery has become more important in recent years as mortality has decreased. Pre-operative risk prediction may be useful for patient counseling, perioperative care planning and quality control. However, mortality scores like the EuroSCORE have often not showed sufficient discrimination or calibration for prediction of other complications. Our aims were therefore to develop separate risk prediction models for postoperative cardiac dysfunction, prolonged mechanical ventilation, and prolonged stay in the intensive care unit (ICU), and to compare our models with various previously published risk scores. The endpoints were chosen because they are common conditions in clinical practice.

Our models were developed in a cohort of approximately 5000 Norwegian patients operated at St. Olavs University Hospital from 2000 through 2007. For the cardiac dysfunction and prolonged ventilation models we employed logistic regression. The model for a prolonged ICU stay was developed using Cox regression. Internal validation was performed by bootstrapping. Discrimination was assessed with areas under the receiver operating characteristics curve, and calibration was assessed with the Hosmer-Lemeshow test and graphical comparison of predicted and observed probability.

Results: Using preoperative variables that are usually easily obtainable in clinical routine, we developed models showing excellent discrimination and good calibration for prediction of postoperative cardiac dysfunction and prolonged mechanical ventilation. Addition of a few intraoperative variables improved discrimination and calibration. Internal validation indicated that the models should behave with little error in future datasets. Our preoperative models showed better performance than alternative previously published scores. The model for prediction of a prolonged ICU stay showed good calibration and excellent discrimination for a stay of more than 2, 5 or 7 days. Discrimination for ICU stay prediction by the EuroSCORE II and other published models was good, but calibration was poor. None of the models were useful for prediction of ICU stay in individual patients because most patients in all risk categories of all models had short ICU stays (75th percentiles: 1 day).

Comments: Our findings support that locally developed models with easily obtainable preoperative variables may accurately predict various endpoints. Even if addition of intraoperative variables marginally improved the accuracy, preoperative models are more useful for clinical practice. A locally developed model may be more suitable than a model that was originally developed from a different population. Our data showed that a good discriminative ability does not guarantee good calibration. The use of ill-calibrated models may lead to serious under- or overestimation of the true risk and thereby to incorrect information to patients and suboptimal treatment decisions. Our findings demonstrated that length of stay in the ICU does not solely depend on clinical reasons, but also on the policy of the institution. For this reason, it is difficult to make universal a model for prediction of ICU stay.