Mortality prediction in ICU Using a Stacked Ensemble Model.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Prediction Modeling

Background:

• Electronic clinical data facilitates machine learning for predicting patient outcomes.
• Artificial intelligence (AI) offers significant potential for developing survival prediction models in intensive care units (ICUs).
• Innovative algorithms are key to enhancing the performance of predictive models.

Purpose of the Study:

• To develop and evaluate a stacked ensemble model (SEM) for predicting mortality in critically ill ICU patients.
• To compare the performance of the SEM with and without the integration of clinical severity scoring results.
• To assess the impact of different feature selection techniques on model prediction accuracy.

Main Methods:

• Utilized a stacked ensemble model incorporating multiple machine learning algorithms.
• Integrated clinical severity scoring results into the prediction model.
• Employed feature selection techniques to obtain optimized feature subsets (SetS and SetT).
• Applied SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• The stacked ensemble model integrating severity scoring achieved a higher Area Under the Curve (AUC) of 0.879 compared to models without it (AUC 0.862).
• The SEM utilizing the SetS feature subset demonstrated superior performance with AUC values of 0.879 and 0.860.
• SHAP analysis provided insights into the relationship between risk features and patient outcomes.

Conclusions:

• Stacked ensemble models incorporating clinical severity scores enhance the prediction of ICU patient mortality.
• Feature selection is crucial for optimizing AI model performance in critical care.
• AI-driven predictive models, interpreted via SHAP, can improve understanding of patient risk factors.