A preoperative risk prediction model for 30-day mortality following cardiac surgery in an Australian cohort.

Abstract

Background: Population-specific risk models are required to build consumer and provider confidence in clinical service delivery, particularly when the risks may be life-threatening. Cardiac surgery carries such risks. Currently, there is no model developed on the Australian cardiac surgery population and this article presents a novel risk prediction model for the Australian cohort with the aim to provide a guide for the surgeons and patients in assessing preoperative risk factors for cardiac surgery. Aim(s): This study aims to identify preoperative risk factors associated with 30-day mortality following cardiac surgery for an Australian population and to develop a preoperative model for risk prediction. Method(s): All patients (23 016) undergoing cardiac surgery between July 2001 and June 2008 recorded in the Australian Society of Cardiac and Thoracic Surgeons (ASCTS) database were included in this analysis. The data were divided randomly into model creation (13 810, 60%) and model validation (9206, 40%) sets. The model was developed on the creation set and then validated on the validation set. The bootstrap sampling and automated variable selection methods were used to develop several candidate models. The final model was selected from this group of candidate models by using prediction mean square error (MSE) and Bayesian Information Criteria (BIC). Using a multifold validation, the average receiver operating characteristic (ROC), p-value for Hosmer-Lemeshow chi-squared test and MSE were obtained. Risk thresholds for low-, moderate- and high-risk patients were defined. The expected and observed mortality for various risk groups were compared. The multicollinearity and first-order interaction effect between clinically meaningful risk factors were investigated. Result(s): A total of 23 016 patients underwent cardiac surgery and the 30-day mortality rate was 3.2% (728 patients). Independent predictors of mortality in the model were: age, sex, the New York Heart Association (NYHA) class, urgency of procedure, ejection fraction estimate, lipid-lowering treatment, preoperative dialysis, previous cardiac surgery, procedure type, inotropic medication, peripheral vascular disease and body mass index (BMI). The model had an average ROC 0.8223 (95% confidence interval (CI): 0.8118-0.8227), p-value 0.8883 (95% CI: 0.8765-0.90) and MSE 0.0251 (95% CI: 0.02515-0.02516). The validation set had observed mortality 3.0% (95% CI: 2.7-3.3%) and predicted mortality 2.9% (95% CI: 2.6-3.2%). The low-risk group (additive score 0-3) had 0.6% observed mortality (95% CI: 0.3-0.9%) and 0.5% predicted mortality (95% CI: 0.2-0.8%). The moderate-risk group (additive score 4-9) had 1.7% observed mortality (95% CI: 1.2-2.2%) and 1.4% predicted mortality (95% CI: 1.0-1.8%). The observed mortality for the high-risk group (additive score 9 plus) was 6.7% (95% CI: 5.8-7.6%) and the expected mortality was 6.7% (95% CI: 5.8-7.6%). Conclusion(s): A preoperative risk prediction model for 30-day mortality was developed for the Australian cardiac surgery population. © 2010 European Association for Cardio-Thoracic Surgery.