Explainable machine learning model for predicting acute pancreatitis mortality in the intensive care unit.

Area of Science:

• Intensive Care Medicine
• Machine Learning in Healthcare
• Clinical Prediction Models

Background:

• Current models for predicting mortality risk in acute pancreatitis (AP) are suboptimal.
• Machine learning (ML) offers a potential improvement for risk stratification in AP patients admitted to the intensive care unit (ICU).

Purpose of the Study:

• To develop and validate an explainable ML model for predicting mortality risk in AP patients.
• To compare the performance of the ML model against established clinical scoring systems.

Main Methods:

• Developed a gradient-boosting ML (XGBoost) model using the MIMIC database (training) and validated it on the eICU-CRD (validation).
• Compared XGBoost model performance against APACHE IV, SOFA, and BISAP scores using Area Under the Receiver Operating Characteristic Curve (AUC) analysis.
• Utilized the SHAP (SHapley Additive exPlanations) method for model interpretability.

Main Results:

• The XGBoost model achieved a superior AUC of 0.89 (95% CI: 0.84-0.94) in predicting mortality risk for AP patients.
• At 100% sensitivity for death prediction, the XGBoost model demonstrated a specificity of 38%, significantly outperforming APACHE IV (1%), SOFA (16%), and BISAP (1%).

Conclusions:

• The developed explainable ML model accurately predicts mortality risk in AP patients.
• This model may enable the safe monitoring of very low-risk patients in general wards, improving resource allocation.
• Explainability of the model enhances physician understanding of prediction reasoning.