An explainable knowledge distillation method with XGBoost for ICU mortality prediction.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Decision Support

Background:

• Mortality prediction in intensive care units (ICUs) is crucial for assessing patient severity.
• Current scoring systems have limitations in performance due to model non-specificity and linearity.
• Electronic health records (EHRs) enable advanced deep learning models but lack clinical explainability.

Purpose of the Study:

• To develop an explainable Knowledge Distillation method with XGBoost (XGB-KD) for enhanced mortality prediction.
• To improve the predictive performance of XGBoost models in clinical settings.
• To provide better explainability for complex patient data.

Main Methods:

• Utilized deep learning models as teachers to learn complex patterns from high-dimensional time series data.
• Distilled knowledge from teacher models' soft labels to train an XGBoost student model.
• Employed feature engineering, model calibration, and SHapley Additive exPlanations (SHAP) for insights.

Main Results:

• XGB-KD demonstrated superior predictive performance compared to vanilla XGBoost, deep learning models, and other state-of-the-art baselines.
• The method achieved better results on the MIMIC-III dataset.
• The approach successfully provided intuitive explanations for predictions.

Conclusions:

• The proposed XGB-KD method effectively enhances XGBoost predictive performance through knowledge distillation.
• The method offers meaningful explanations, addressing the explainability gap in clinical deep learning.
• XGB-KD is a valuable tool for improving ICU mortality prediction and clinical decision-making.