Early prediction of transfusion requirements in trauma patients using explainable machine learning

Hemorrhagic shock is the most common preventable cause of death in trauma patients. Early transfusion significantly improves survivability in patients suffering from hemorrhagic shock. We hypothesized that machine learning models could reduce time to initial transfusion by more rapidly identifying patients likely to require blood products.

We conducted a retrospective model-development and internal-validation study using the Trauma Quality Improvement Program (TQIP) dataset from 2019 to 2022. Adult patients aged 18 years or older were included; the only exclusion criterion was missing transfusion outcome data. Models evaluated included logistic regression, ridge regression, Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest, and extreme gradient boosting (XGBoost) assessed using an 80/20 train-test split. Five-fold cross-validation was used for hyperparameter tuning. Area under the receiver operating characteristic curve (AUROC) was used for model optimization.

A total of 1,232,956 patients were included, of whom 109,819 (8.9%) received transfusions. XGBoost achieved the highest AUROC (0.8452). With a decision threshold of 0.0816 selected by maximizing balanced accuracy on the test set, accuracy was 0.7298, sensitivity was 0.7927, and specificity was 0.7236. SHapley Additive exPlanations (SHAP) values showed that Injury Severity Score, systolic blood pressure, activation level, pulse rate, and age were the most influential predictors.

XGBoost demonstrated strong predictive performance using only arrival-time variables, and SHAP explanations offered clinically intuitive reasoning. These characteristics highlight the model's potential as a rapid decision-support tool, although external validation remains an important next step.