Development and validation of web-based, interpretable predictive models for sepsis and mortality in extensive burns

Area of Science:

• Medical research
• Computational biology
• Trauma surgery

Background:

• Extensive burn injuries (≥ 50% total body surface area) are linked to high sepsis and mortality rates.
• Identifying risk factors and developing predictive models are crucial for managing these severe cases.

Purpose of the Study:

• To identify risk factors for sepsis and mortality in extensively burned patients.
• To develop accurate and interpretable machine learning models for predicting sepsis and mortality.

Main Methods:

• Retrospective cohort study of 237 extensively burned patients (2012-2023).
• Applied ten machine learning algorithms (e.g., Random Forest, Gradient Boosting Tree) to predict sepsis and mortality.
• Evaluated models using AUC, precision, recall, accuracy, F1 score; compared with SOFA score; utilized SHAP for interpretability.

Main Results:

• Key sepsis predictors: SOFA score, new shock, albumin, BUN, third-degree burn area, TBSA burned, WBC count, inhalation injury.
• Key mortality predictors: ALT, SOFA score, burn type, new shock, third-degree burn area, TBSA burned, sepsis.
• Random Forest model achieved high sepsis prediction (AUC=0.977); Gradient Boosting Tree model excelled in mortality prediction (AUC=0.981).

Conclusions:

• Machine learning models (RF, GBT) effectively predict sepsis and mortality in extensive burn patients.
• SHAP analysis enhances model transparency for clinical interpretation and early intervention.
• Developed web-based calculators to aid clinical decision-making and improve patient outcomes.