Background: The aim of this study was to develop two novel risk prediction scores for transfusion and bleeding that would be used to inform treatment decisions, quality assurance, and clinical trial design in cardiac surgery. Methods: Clinical data prospectively collected from 26 UK cardiac surgical centres and a single European centre were used to develop two risk prediction models: one for any red blood cell (RBC) transfusion, and the other for large volume blood transfusion (≥4 RBC units; LVBT), an index of severe blood loss. 'Complete case' data were available for 24 749 patients. Multiple imputation was used for missing covariate data (typically <5% per variable), with the imputed data set containing 39 970 patients. Risk models were developed in the complete case data set, with internal validation using leave-one-centre-out cross-validation. The final selected models were fitted to the imputed data set. Final risk scores were then compared with the performance of three existing scores: the Transfusion Risk and Clinical Knowledge score (TRACK), the Transfusion Risk Understanding Scoring Tool (TRUST), and the Papworth Bleeding Risk Score (BRiSc). Results: The area under the receiver operating characteristic curve (AUC) was 0.77 (95% confidence interval 0.77-0.77) for the any RBC transfusion score and AUC 0.80 (0.79-0.80) for the LVBT score in the imputed data set. The LVBT model also showed excellent discrimination (Hosmer-Lemeshow P=0.32). In the imputed data set, the AUCs for the TRACK and TRUST scores for any RBC transfusion were 0.71 and 0.71, respectively, and for LVBT the AUC for the BRiSc score was 0.69. Conclusions: Two new risk scores for any RBC transfusion or LVBT among cardiac surgery patients have excellent discrimination, and could inform clinical decision making.
- blood transfusion
- cardiac anaesthesia