Context: Cardiac dysfunction following open heart surgery is a major determinant of subsequent morbidity and mortality.
Objectives: To develop a specific risk prediction model for postoperative cardiac dysfunction based on preoperative variables, to investigate whether prediction could be improved by inclusion of selected intraoperative variables and to compare our model with five previously published risk scores.
Design: Single-centre prospectively collected data.
Setting: Tertiary care centre, Middle Norway.
Patients: Four thousand nine hundred and eighty-nine patients (all eligible) undergoing open cardiac surgery from 2000 to 2007.
Main outcome measures: Logistic regression models for postoperative cardiac dysfunction: predictive accuracy/calibration, discrimination as shown by area under the receiver operating characteristics curve, internal validity as indicated by bootstrapping, comparison of goodness-of-fit with predictions based on alternative risk scores.
Results: The preoperative model included chronic cardiac insufficiency, previous myocardial infarction, previous cardiac operation, pulmonary hypertension, renal dysfunction, low haemoglobin concentration, urgent operation and operation type other than isolated coronary artery bypass surgery. The area under the curve was 0.838 [95% confidence interval (CI) 0.812-0.864]. Risk prediction was accurate apart from a slight overestimation in the 2% of highest risk patients. Inclusion of a few intraoperative variables (inotropic or vasoconstrictor drugs, plasma or red cell transfusion) improved the model slightly, increasing the area under the curve to 0.875 (95% CI 0.854-0.896) or 0.890 (95% CI 0.863-0.902) for two equivalent models. On the basis of estimated shrinkage factors of 0.94, 0.97 and 0.98, respectively, the models should behave with 6% or less error in future datasets. Our preoperative model was significantly better than the previously published risk scores (P < 0.0002 for comparison of area under the curves).
Conclusion: The preoperative model including variables obtained easily in routine clinical work performed well and was improved only slightly by inclusion of intraoperative variables. Performance was better than those of the five previously published risk scores.