A model to predict risk of blood transfusion after gynecologic surgery

Abstract

A model that predicts a patient's risk of receiving a blood transfusion may facilitate selective preoperative testing and more efficient perioperative blood management utilization. We sought to construct and validate a model that predicts a patient's risk of receiving a blood transfusion after gynecologic surgery. In all, 18,319 women who underwent gynecologic surgery at 10 institutions in a single health system by 116 surgeons from January 2010 through June 2014 were analyzed. The data set was split into a model training cohort of 12,219 surgeries performed from January 2010 through December 2012 and a separate validation cohort of 6100 surgeries performed from January 2013 through June 2014. In all, 47 candidate risk factors for transfusion were collected. Multiple logistic models were fit onto the training cohort to predict transfusion within 30 days of surgery. Variables were removed using stepwise backward reduction to find the best parsimonious model. Model discrimination was measured using the concordance index. The model was internally validated using 1000 bootstrapped samples and temporally validated by testing the model's performance in the validation cohort. Calibration and decision curves were plotted to inform clinicians about the accuracy of predicted probabilities and whether the model adds clinical benefit when making decisions. The transfusion rate in the training cohort was 2% (95%confidence interval, 1.72-2.22). The model had excellent discrimination and calibration during internal validation (bias-corrected concordance index, 0.906; 95% confidence interval, 0.890-0.928) and maintained accuracy during temporal validation using the separate validation cohort (concordance index, 0.915; 95% confidence interval, 0.872-0.954). Calibration curves demonstrated the model was accurate up to 40% then it began to overpredict risk. The model provides superior net benefit when clinical decision thresholds are between 0-50% predicted risk. This model accurately predicts a patient's risk of transfusion after gynecologic surgery facilitating selective preoperative testing and more efficient perioperative blood management utilization.