Development and validation of a multivariable prediction model for mediastinitis poststernotomy

Methods Data of 4526 patients admitted after MHS to our Cardiovascular ICU (January 2005,-June 2011) were prospectively collected. After exclusion of 655 patients, 3970 were analyzed to identify risk factors for mediastinitis. A model was generated by logistic regression analysis in a subgroup of 2618 randomly selected patients and validated in a second cohort of 1352. The discriminatory power was evaluated by area under the receiver operating characteristic curves (AUC-ROC). Calibration was done by the Hosmer-Lemeshow test. The mediastinitisscore (“MED-Score”) was created using the points estimate for each variable, with beta coefficients of the final model. For stratification three risk levels were generated. We assessed the ability of “MED-Score, Euroscore (ES)1, and the Society of Thoracic Surgeons score (STS)2 to predict mediastinitis using ROC curves comparison (MannWhitney). Concordance of risk stratification by scoring systems was evaluated using weighted Kappa index. We used SPSS vs15.0. and STATA 11.0. The outcome “mediastinitis” was defined according to CDC criteria. Results 94 (2.4%) patients developed mediastinitis. Four preoperative (age >70, COPD, BMI>30 and antiplatelet therapy) and 3 perioperative predictors of mediastinitis (ischemia time, emergency reoperation and prolonged intubation) were identified and included in the logistic model, which accurately predicted outcome (AUC.ROC 0.80). Hosmer-Lemeshow test p = 0.46. The AUC for preoperative Society of Thoracic Surgeons (STS), and combined STS, logEuroScore had a lower discrimination power than MED-Score. This difference was statistically significant. Concordance of risk stratification by the different scoring systems was poor (Kappa< 0.20).