Explainable artificial intelligence model for mortality risk prediction in the intensive care unit: a derivation and...

Area of Science:

• Critical care medicine
• Machine learning in healthcare
• Predictive analytics

Background:

• Machine learning (ML) models for mortality risk prediction often lack transparency.
• Improving the explainability of ML algorithms is crucial for clinical adoption.

Purpose of the Study:

• To enhance transparency in ML-based mortality risk prediction for critically ill patients.
• To develop and validate an interpretable ML model using SHapley Additive exPlanation (SHAP).

Main Methods:

• Data from the Medical Information Mart for Intensive Care IV database was used.
• Nine ML models were developed, with the optimal model selected based on accuracy and AUC.
• SHAP methodology was employed for model interpretability.

Main Results:

• A cohort of 21,395 critically ill patients was analyzed.
• The Random Forest model demonstrated the highest accuracy (87.62%) and AUC (0.89).
• Key predictors identified by SHAP included Glasgow Coma Scale, urine output, and blood urea nitrogen.

Conclusions:

• A transparent ML model for predicting outcomes in critically ill patients is feasible and effective.
• SHAP significantly enhances the explainability of ML models in critical care settings.