Predicting Mortality Using Machine Learning Algorithms in Patients Who Require Renal Replacement Therapy in the...

Area of Science:

• Nephrology
• Critical Care Medicine
• Data Science in Healthcare

Background:

• Existing severity of illness scores lack precision in predicting mortality for acute kidney injury (AKI) patients requiring renal replacement therapy (RRT).
• There is a need for improved predictive models in this high-risk patient population.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting mortality in AKI patients undergoing RRT.
• To compare the performance of ML models against established scoring systems such as SOFA and HELENICC.

Main Methods:

• Routinely collected clinical data from AKI patients requiring RRT were extracted from the MIMIC and eICU databases.
• Four ML models (MLP, logistic regression, XGBoost, RF) were developed and validated.
• Model performance was assessed using AUC, accuracy, calibration, sensitivity, specificity, PPV, and NPV, and compared to SOFA, nonrenal SOFA, and HELENICC scores.

Main Results:

• The XGBoost model demonstrated the highest mortality prediction performance with an AUC of 0.823 (95% CI, 0.791−0.854) in the testing dataset.
• XGBoost achieved the highest accuracy (0.758) and showed no significant lack of fit (p > 0.05) via the Hosmer−Lemeshow test.
• Machine learning models generally outperformed SOFA and HELENICC scores in predicting mortality.

Conclusions:

• The XGBoost model offers superior performance for mortality prediction in AKI patients requiring RRT compared to existing scoring systems.
• Machine learning approaches hold significant promise for enhancing clinical decision-making in critical care settings.
• Further validation and implementation of ML models can improve patient outcomes.