Development and evaluation of a blood transfusion prediction algorithm tailored for precision and utility in civilian and military trauma patients

Early decisions regarding blood transfusion in trauma are time-critical and made under substantial uncertainty. Existing prediction models typically classify patients using fixed thresholds and do not estimate the volume or timing of transfusion. This limits their utility for real-time clinical decision-making and resource planning. We developed and validated a Bayesian network model, informed by domain knowledge and clinical data, to predict transfusion requirements in trauma using injury and physiological variables available early after injury. The model estimates transfusion volume across multiple timeframes (2, 6, 12, and 24 h). It was developed using data from a prospective civilian cohort and military trauma registries ( n = 2361) and evaluated using internal cross-validation and external validation in an independent dataset ( n = 13,347). The model achieved AUROC values of 0·82–0·94 for binary outcomes across timeframes. For ordinal volume predictions, Ranked Probability Scores were 0·07 for 0–12 h predictions and 0·04 for 12–24 h predictions. In scenario-based evaluation, agreement with model predictions was observed in 213 of 225 (95%) expert assessments. Prediction models based on binary thresholds may have limited impact on clinical decision-making despite good performance. A decision-support approach that estimates transfusion requirements by volume and timing, using early-available clinical data, may provide more clinically useful decision support in trauma care.