Interpretable machine learning for predicting sepsis risk in emergency triage patients.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Emergency Medicine

Background:

• Early sepsis detection is critical for patient outcomes.
• Traditional sepsis screening often relies solely on vital signs.
• Integrating broader patient data can enhance predictive accuracy.

Purpose of the Study:

• To develop and validate a sepsis prediction model using structured electronic medical records (sEMR) and machine learning (ML) in emergency triage.
• To compare the effectiveness of a comprehensive model against one using only vital signs.
• To improve early sepsis screening by incorporating diverse triage information.

Main Methods:

• Retrospective cohort study using the MIMIC-IV database (189,617 patients).
• Developed two models: Model 1 (vital signs only) and Model 2 (vital signs, demographics, medical history, chief complaints).
• Employed eight ML algorithms and evaluated performance using AUC, F1 Score, and calibration curves; utilized SHAP and LIME for interpretability.

Main Results:

• Model 2, incorporating comprehensive data, consistently outperformed Model 1 across most algorithms.
• Gradient Boosting in Model 2 achieved the highest AUC of 0.83; Extra Tree, Random Forest, and SVM achieved 0.82.
• SHAP analysis provided interpretable insights into the Gradient Boosting model's predictions.

Conclusions:

• Comprehensive triage information integrated via sEMR and ML is more effective for sepsis prediction than vital signs alone.
• Interpretable ML enhances model transparency and provides sepsis prediction probabilities.
• This approach offers a feasible strategy for early sepsis screening, supporting clinical decision-making in triage.