Predicting the complexity and mortality of polytrauma patients with machine learning models

Area of Science:

• Medical Informatics
• Computational Biology
• Trauma Surgery

Background:

• Polytrauma patients present complex prognoses requiring accurate prediction of mortality and complexity.
• Existing trauma evaluation scores have limitations in precisely assessing polytrauma outcomes.
• Machine learning offers a promising avenue for enhancing prognostic accuracy in critical care settings.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting polytrauma patient mortality and complexity.
• To compare the performance of machine learning models against established trauma scoring systems.
• To identify key clinical predictors contributing to polytrauma outcomes.

Main Methods:

• Retrospective analysis of 756 polytrauma patients admitted to the ICU.
• Development of machine learning models (SVM, RF, ANN, XGBoost) using clinical features.
• Comparison of optimal ML model performance against Injury Severity Score (ISS), Trauma Index (TI), and Glasgow Coma Scale (GCS).

Main Results:

• The XGBoost model achieved 90% accuracy and 88% F-score for mortality prediction, outperforming other ML models and conventional scores.
• External validation confirmed the XGBoost model's stability and generalization (91% accuracy, 82% AUC).
• XGBoost also demonstrated superior performance in predicting polytrauma complexity, with notable predictors like Intracranial Hematoma (ICH).

Conclusions:

• Machine learning algorithms, particularly XGBoost, can significantly enhance the prognostic estimation of polytrauma patients.
• The developed ML models offer improved accuracy over traditional scoring systems for predicting mortality and complexity.
• Leveraging ML in polytrauma care holds potential value for optimizing patient management and outcomes.