"Dynamic Ensemble, Then Knowledge Distillation": A SHAP-Driven Two-Stage Framework for Sepsis Mortality Prediction.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Sepsis Research

Background:

• Dynamic ensemble learning (DEL) models aid sepsis progression monitoring.
• Existing DEL methods struggle with patient heterogeneity and lack explainability.

Purpose of the Study:

• To propose a novel SHapley Additive exPlanations (SHAP)-driven Dynamic Ensemble then Knowledge Distillation (DEKD) framework.
• To enhance sepsis mortality prediction accuracy and model explainability.

Main Methods:

• DEKD employs a two-stage approach: patient clustering using SHAP values and knowledge distillation.
• Base models are built on patient clusters, with predictions aggregated via a weighted distance-based ensemble.
• A student model is trained using knowledge distillation for overall explanation and performance improvement.

Main Results:

• DEKD achieved an AUROC of 0.955 for 48-hour mortality and 0.924 for in-hospital mortality on the MIMIC-III dataset.
• The framework demonstrated significant improvement in diversity compared to benchmark methods.
• DEKD enhances both predictive performance and model explainability.

Conclusions:

• The proposed DEKD framework effectively addresses limitations of existing DEL methods in sepsis mortality prediction.
• DEKD offers a promising approach for accurate and interpretable sepsis outcome prediction.
• This method advances the application of machine learning in critical care settings.