Derivation and validation of a simplified predictive index for renal replacement therapy after cardiac surgery

Context A predictive index for renal replacement therapy (RRT; hemodialysis or continuous venovenous hemodiafiltration) after cardiac surgery may improve clinical decision making and research design. Objectives To develop a predictive index for RRT using preoperative information. Design, Setting, and Participants Retrospective cohort of 20 131 cardiac surgery patients at 2 hospitals in Ontario, Canada. The derivation cohort consisted of 10 751 patients at Toronto General Hospital (1999-2004). The validation cohorts consisted of 2566 patients at Toronto General Hospital (2004-2005) and 6814 patients at Ottawa Heart Institute (1999-2003). Main Outcome Measure Postoperative RRT. Results RRT rates in the derivation, Toronto validation, and Ottawa validation cohorts were 1.3%, 1.8%, and 2.2%, respectively. Multivariable predictors of RRT were preoperative estimated glomerular filtration rate, diabetes mellitus requiring medication, left ventricular ejection fraction, previous cardiac surgery, procedure, urgency of surgery, and preoperative intra-aortic balloon pump. The predictive index was scored from 0 to 8 points. An estimated glomerular filtration rate less than or equal to 30 mL/min was assigned 2 points; other components were assigned 1 point each: estimated glomerular filtration rate 31 to 60 mL/min, diabetes mellitus, ejection fraction less than or equal to 40%, previous cardiac surgery, procedure other than coronary artery bypass grafting, intra-aortic balloon pump, and nonelective case. Among the 53% of patients with low risk scores (<= 1), the risk of RRT was 0.4%; by comparison, this risk was 10% among the 6% of patients with high-risk scores (>= 4). The predictive index had areas under the receiver operating characteristic curve in the derivation, Toronto validation, and Ottawa validation cohorts of 0.81, 0.78, and 0.78, respectively. When these cohorts were stratified based on index scores, likelihood ratios for RRT were more concordant than observed RRT rates. Conclusions RRT after cardiac surgery is predicted by readily available preoperative information. A simple predictive index based on this information discriminated well between low- and high-risk patients in derivation and validation cohorts. The index had improved generalizability when used to predict likelihood ratios for RRT.