Advancing polytrauma care: developing and validating machine learning models for early mortality prediction.

Area of Science:

• Medical Informatics
• Trauma Surgery
• Machine Learning in Healthcare

Background:

• Early identification of high-risk polytrauma patients is critical for effective intervention and improved survival rates.
• Machine learning (ML) offers a promising approach for developing predictive models using clinical data.

Purpose of the Study:

• To develop and validate ML models for predicting 72-hour mortality in adult polytrauma patients.
• To identify key clinical parameters influencing mortality prediction.

Main Methods:

• Retrospective analysis of polytrauma patients from Dryad and institutional databases.
• Development and validation of Random Forest (RF), neural network, and XGBoost models.
• Utilized SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) for model interpretability.

Main Results:

• Key predictors of 72-hour mortality included age, BMI, GCS, ISS, pH, base excess, and lactate.
• The RF model achieved high performance: AUROC 0.87 (internal) and 0.98 (external), AUPRC 0.67 (internal) and 0.88 (external), and accuracy 0.83 (internal) and 0.97 (external).
• RF model demonstrated superior predictive efficacy and clinical utility compared to other models.

Conclusions:

• The Random Forest model is highly effective in predicting 72-hour mortality in adult polytrauma patients.
• This ML model can assist clinicians in identifying at-risk patients, thereby guiding clinical decision-making and improving patient outcomes.