Development and Validation of Machine Learning Models for Real-Time Mortality Prediction in Critically Ill Patients...

Area of Science:

• Critical Care Medicine
• Nephrology
• Machine Learning in Healthcare

Background:

• Sepsis-associated acute kidney injury (SA-AKI) is a frequent complication in critically ill patients, significantly increasing mortality rates.
• Current mortality prediction tools lack sufficient accuracy and fail to capture the dynamic clinical status of SA-AKI patients.

Purpose of the Study:

• To develop and validate machine learning-based models for real-time mortality prediction in critically ill patients with SA-AKI.
• To improve the accuracy and timeliness of mortality risk assessment for SA-AKI patients.

Main Methods:

• A multi-center retrospective study utilizing data from MIMIC-IV and eICU-CRD databases.
• Development and validation of eXtreme Gradient Boosting (XGBoost) models using routine clinical variables updated every 12 hours.
• Evaluation of model performance using Area Under the Receiver Operating Characteristic Curves (AUCs) for predicting mortality at various time points.

Main Results:

• XGBoost models demonstrated superior performance in mortality prediction compared to SOFA and SAPS-II scores.
• Internal test set AUCs ranged from 0.848 to 0.804, and external test set AUCs ranged from 0.818 to 0.748.
• Shapley Additive Explanation (SHAP) method provided model interpretability, clarifying the relationship between predictors and mortality.

Conclusions:

• Interpretable machine learning XGBoost models show significant promise for real-time mortality prediction in SA-AKI patients.
• These models can serve as valuable tools for early identification of high-risk individuals.
• Timely clinical interventions can be facilitated by the accurate and dynamic risk stratification provided by these models.