Abstract
<title>Abstract</title> <p>Background Procedure-based blood-ordering schedules may not capture patient-level variation in perioperative transfusion risk. We aimed to develop and temporally evaluate a preoperative red-cell transfusion model and quantify the consequences of development-derived crossmatching thresholds. Methods This single-centre retrospective study included adults undergoing elective abdominal or pelvic cancer surgery in 2024–2025. Patients treated in 2024 formed the development cohort and those treated in 2025 formed the temporal evaluation cohort. The outcome was any red-cell transfusion from the intraoperative period through 72 h after surgery. Ridge-penalised logistic regression used routinely available preoperative patient and surgical variables. Performance was assessed using discrimination, Brier score, and calibration. Sensitivity-based thresholds were derived only in the development cohort and applied unchanged to the temporal cohort. Results Among 1,068 patients, 541 were included in the development cohort and 527 in the temporal cohort; 115 (21.3%) and 86 (16.3%), respectively, received red-cell transfusion within 72 h. The primary continuous-haemoglobin model had an optimism-corrected area under the curve of 0.812 in development and a temporal area under the curve of 0.790 (95% confidence interval 0.738–0.844), compared with 0.713 for the binary-anaemia model and 0.622 for the procedure-only model. The temporal Brier score was 0.114, with a calibration intercept of − 0.480 and slope of 0.781, indicating overall risk overestimation and overly dispersed predictions. The spline model had a temporal area under the curve of 0.808. At the primary 95% sensitivity target, the locked 3.83% threshold recommended crossmatching for 421 patients (79.9%), with sensitivity 97.7% and negative predictive value 98.1%. In simulation, this represented 95 fewer recommended crossmatches than observed practice, while two of 86 transfusion events occurred below the threshold. Conclusions A parsimonious model based on routinely available preoperative variables retained discrimination in a subsequent-year cohort, but temporal miscalibration limits direct implementation. Risk-based crossmatching warrants further evaluation as a decision-support strategy, with local recalibration, external validation, and prospective impact assessment required before clinical adoption.</p>