Early Prediction of Massive Transfusion for Patients With Traumatic Hemorrhage: Development of a Multivariable
Andrew J Benjamin1,2, Andrew J Young3, John B Holcomb4,5
1From the Division of Traumatology, Surgical Critical Care & Emergency Surgery, Department of Surgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA.
Objective:
Develop a novel machine learning (ML) model to rapidly identify trauma patients with severe hemorrhage at risk of early mortality.
Background:
The critical administration threshold (CAT, 3 or more units of red blood cells in a 60-minute period) indicates severe hemorrhage and predicts mortality, whereas early identification of such patients improves survival.
Methods:
Patients from the PRospective, Observational, Multicenter, Major Trauma Transfusion and Pragmatic, Randomized Optimal Platelet, and Plasma Ratio studies were identified as either CAT+ or CAT-. Candidate variables were separated into 4 tiers based on the anticipated time of availability during the patient's assessment. ML models were created with the stepwise addition of variables and compared with the baseline performance of the assessment of blood consumption (ABC) score for CAT+ prediction using a cross-validated training set and a hold-out validation test set.
Results:
Of 1245 PRospective, Observational, Multicenter, Major Trauma Transfusion and 680 Pragmatic, Randomized Optimal Platelet and Plasma Ratio study patients, 1312 were included in this analysis, including 862 CAT+ and 450 CAT-. A CatBoost gradient-boosted decision tree model performed best. Using only variables available prehospital or on initial assessment (Tier 1), the ML model performed superior to the ABC score in predicting CAT+ patients [area under the receiver-operator curve (AUC = 0.71 vs 0.62)]. Model discrimination increased with the addition of Tier 2 (AUC = 0.75), Tier 3 (AUC = 0.77), and Tier 4 (AUC = 0.81) variables.
Conclusions:
A dynamic ML model reliably identified CAT+ trauma patients with data available within minutes of trauma center arrival, and the quality of the prediction improved as more patient-level data became available. Such an approach can optimize the accuracy and timeliness of massive transfusion protocol activation.
