Explainable AI for mortality prediction: a comparative study using the MIMIC-III dataset.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems

Background:

• Predicting patient mortality in intensive care units (ICUs) is crucial for treatment optimization and resource management.
• Machine learning (ML) models demonstrate potential in outperforming traditional scoring systems for ICU mortality prediction.
• The 'black-box' nature of ML hinders clinical adoption, necessitating explainable AI (XAI) methods.

Purpose of the Study:

• To evaluate the accuracy of various ML algorithms in predicting ICU mortality using the MIMIC-III dataset.
• To apply XAI techniques, specifically SHapley Additive exPlanations (SHAP), to identify key predictors of mortality.
• To assess the potential of interpretable ML models in supporting clinical decision-making.

Main Methods:

• A retrospective analysis of 600 patient records from the MIMIC-III database.
• Implementation and comparison of eight ML algorithms: SVM, KNN, DT, GB, RF, NB, LR, and ET.
• Model performance evaluation using threefold cross-validation, F1 Score, sensitivity, specificity, and accuracy.
• Application of SHAP for identifying significant mortality predictors.

Main Results:

• Extra Trees (ET) and Gradient Boosting (GB) achieved the highest accuracy (98.33% and 98.23%) with F1-scores over 96%.
• Support Vector Machine (SVM) also showed strong performance (97.50% accuracy).
• SHAP analysis identified hypertension, tumors, and endocrine/digestive diseases as leading mortality predictors.

Conclusions:

• ML algorithms, particularly ET and GB, are highly effective for predicting ICU mortality.
• Explainable AI (XAI) is essential for building trust and facilitating the adoption of ML in clinical settings.
• Interpretable ML models can safely support informed ICU decisions and improve patient outcomes.