Machine Learning Models for Predicting the Need for Early Packed Red Blood Cell Transfusion in Multiple Trauma...

One of the preventable contributors to trauma mortality is hemorrhagic shock, which requires early recognition and immediate intervention. In this retrospective analysis, we aimed to develop and optimize machine learning (ML) algorithms to predict the need for packed red blood cell (PRBC) transfusion within 24 hours of injury in multiple trauma patients.

This retrospective longitudinal study analyzed consecutive multiple trauma patients admitted to the emergency department. The outcome was transfusion of at least one unit of PRBC within the first 24 hours of traumatic injury. SHAP analysis was employed for feature selection, and the five key predictors were identified and entered in the models: Glasgow Coma Scale (GCS), hemoglobin (Hb), pulse rate (PR), systolic blood pressure (SBP), and pulse pressure. The dataset was split 80:20 for training/testing, and multiple machine learning algorithms were evaluated based on area under the receiver operating characteristic curve (AUC), F1 score, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).

The study cohort consisted of 908 patients, with a median age of 34 years. PRBC transfusions were more common in older adults with lower GCS scores, higher PR, lower SBP, lower pulse pressure, and lower Hb levels on admission. Among the machine learning models, Random Forest performed best (AUC: 0.997, sensitivity: 0.938, specificity: 0.994), followed by K-Nearest Neighbors and Logistic Regression, both of which showed perfect specificity but lower sensitivity.

Random Forest outperformed other ML algorithms, achieving high discriminative ability, sensitivity, and specificity. PR, GCS, Hb, SBP, and pulse pressure were the most influential predictors of the need for early transfusion. Despite promising results, further multicenter validation studies are needed to confirm the real-world applicability of these models.