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

Area of Science:

• Medical Informatics
• Trauma Surgery
• Emergency Medicine

Background:

• Hemorrhagic shock is a leading cause of preventable trauma mortality.
• Early recognition and intervention are critical for managing hemorrhagic shock.
• Predictive tools can aid in timely transfusion decisions for trauma patients.

Purpose of the Study:

• To develop and optimize machine learning (ML) algorithms for predicting the need for packed red blood cell (PRBC) transfusion within 24 hours of injury in multiple trauma patients.
• To identify key clinical predictors influencing early PRBC transfusion requirements.

Main Methods:

• Retrospective analysis of 908 multiple trauma patients.
• Utilized SHAP analysis for feature selection, identifying Glasgow Coma Scale (GCS), hemoglobin (Hb), pulse rate (PR), systolic blood pressure (SBP), and pulse pressure as key predictors.
• Evaluated multiple ML algorithms (Random Forest, K-Nearest Neighbors, Logistic Regression) using AUC, F1 score, sensitivity, and specificity.

Main Results:

• The Random Forest model demonstrated superior performance with an AUC of 0.997, sensitivity of 0.938, and specificity of 0.994.
• PRBC transfusions were associated with lower GCS, higher PR, lower SBP, lower pulse pressure, and lower Hb levels.
• K-Nearest Neighbors and Logistic Regression also showed high specificity but lower sensitivity.

Conclusions:

• Machine learning, particularly the Random Forest algorithm, effectively predicts the need for early PRBC transfusion in trauma patients.
• GCS, Hb, PR, SBP, and pulse pressure are significant predictors for early transfusion.
• Further multicenter validation is recommended to confirm the clinical applicability of these predictive models.