An interpretable machine learning algorithm enables dynamic 48-hour mortality prediction during an ICU stay.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Critical Care Medicine

Background:

• Accurate short-term mortality prediction is crucial for intensive care unit (ICU) management.
• Existing models often use static data, failing to capture the dynamic nature of critical illness.
• This study addresses the need for dynamic prediction models throughout the ICU stay.

Purpose of the Study:

• To develop and validate an interpretable machine learning algorithm for dynamic 48-hour mortality prediction in the ICU.
• To enable continuous risk assessment and adaptation to patient status changes.
• To improve clinical decision-making through real-time predictive insights.

Main Methods:

• Retrospective cohort study of 9,786 ICU patients (2018-2022) at a German university hospital.
• Development of a Light Gradient-Boosting Machine (LGBM-48h) model for 48-hour mortality prediction, updated daily.
• Nested cross-validation for training/evaluation; external validation on the MIMIC-IV database; SHAP values for feature importance analysis.

Main Results:

• The LGBM-48h model achieved high performance: AUROC of 0.909 (training) and 0.886 (testing).
• External validation on MIMIC-IV yielded an AUROC of 0.859, demonstrating generalizability.
• The model effectively stratified risk dynamically and highlighted key features influencing mortality predictions via time-varying SHAP values.

Conclusions:

• LGBM-48h offers a dynamic and interpretable approach to predicting short-term ICU mortality.
• The model has the potential to support clinical decision-making and care prioritization.
• Further prospective validation is recommended for real-world clinical implementation.