Prediction of elective surgical red blood cell transfusion reveals surgeon-level variation

BACKGROUND Preoperative prediction of red blood cell (RBC) transfusion in elective surgery serves two complementary purposes: enabling patient blood management (PBM) interventions before surgery and supporting blood product supply planning for elective lists. Beyond patient risk factors, individual surgeon practice may also contribute to transfusion rates, but this effect has not been fully quantified. STUDY DESIGN AND METHODS In a retrospective cohort of 279,905 elective surgical procedures at Helsinki University Hospital (2021-2025), we trained four gradient-boosted tree (CatBoost) models crossing patient factors with procedure group and pseudonymised surgeon identity. All predictors were restricted to data available seven or more days before scheduled surgery. The primary outcome was any allogeneic RBC transfusion from operating-room entry through seven postoperative days. Discrimination was assessed by the area under the precision-recall curve (AUPRC) on a held-out 20% test set. Surgeon-level effects were characterised by Shapley additive explanation (SHAP) values. RESULTS The transfusion rate was 2.2%. The patient-only model achieved AUPRC 0.261 (no-skill baseline 0.022). Adding surgeon identity to the procedure-adjusted model increased AUPRC from 0.440 to 0.472 (+0.032; 95% CI 0.020-0.044), raising lift from 20.3 to 21.7. Surgeon identity was the third most important of 72 features by permutation importance. DISCUSSION The operating surgeon was associated with RBC transfusion beyond patient and procedure factors, and surgeon identity was the third most important predictor in a deployment-realistic model using only data available a week before surgery. This information is readily available from routine records and could support PBM audit and blood supply planning for elective surgery.