Cooler by cooler - Forecasting transfusion demand and volumes in trauma

Predictive algorithms may optimize the delivery of care by facilitating early identification of resource requirements. We hypothesize that injury patterns and vitals obtained early after trauma can predict blood volume requirements for blunt and penetrating trauma across a wide array of environments.

Patients 18-60 years with blunt, penetrating and burn were included from the 2020 American College of Surgeons-Trauma Quality Improvement Program (TQIP) registry. Emergency Medical Services (EMS) and emergency department (ED) vital signs were analyzed. External injuries that could be easily noted in the field were curated from ICD codes. Blood volumes were treated as a continuous variable with probabilistic distributions predicted. Blood products characterized include packed red blood cells, plasma, platelets, cryoprecipitate and whole blood. The Field Artificial Intelligence Triage (FAIT) tool utilizes 10-fold cross validation using logistic regression and class-based weighting for predictions.

The population was 401,337 patients; (314,790 blunt, 71,760 penetrating, and 5726 burn) patients. The median age was 37 (interquartile range (IQR), 27-50). Median injury severity score (ISS) of 6 (IQR, 4-12) with ISS of 26 (IQR, 17-35) for those receiving massive transfusion (>4 units/4 h). FAIT accurately predicted when patients did not require blood; 0.91 ± 0.00 (AUC). Overall predictive capabilities for transfusion volume across ranges corresponding to transfusion (>0 units/4 h), massive transfusion (>4 units/4 h), and ultra-massive transfusion (>10 units/4 h and >20 units/4 h) had an AUC of 0.91-0.95 for the EMS + ED and 0.87-0.91 for the EMS vitals alone. Overall survival for different transfusion thresholds was 71% for > 5 units, 62% for > 10 units and 35% for > 50 units (AUC 0.94-0.95 ± 0.01).

The FAIT tool capably distinguishes patients who require transfusion from those who do not across a wide array of injury patterns, severity, and environments. With only limited pre-hospital inputs, the model predicts transfusion volumes within clinically relevant margins.