Optimizing ICU Care: Machine Learning and PCA for Early Prediction of Renal Replacement Therapy Requirement

Area of Science:

• Critical Care Medicine
• Nephrology
• Data Science in Healthcare

Background:

• Early identification of Renal Replacement Therapy (RRT) need in intensive care units (ICUs) is vital for patient outcomes and resource management.
• Machine learning offers potential for developing predictive models for RRT initiation.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting the requirement of RRT within 24 hours of ICU admission.
• To compare the performance of different machine learning algorithms and feature selection methods.

Main Methods:

• Utilized a dataset of 34,000 ICU admissions.
• Assessed Random Forest (RF), Neural Network (NN), and XGBoost models.
• Employed SHapley Additive exPlanations (SHAP) for model interpretability.
• Evaluated model performance using AUPRC and AUC-ROC metrics.
• Compared full datasets, Principal Component Analysis (PCA) reduced data, and top 10 feature models.

Main Results:

• XGBoost demonstrated superior performance in Area Under the Precision-Recall Curve (AUPRC).
• Random Forest (RF) achieved better performance in Area Under the Receiver Operating Characteristic Curve (AUC-ROC).
• Models utilizing the top 10 features outperformed the PCA-reduced model with fewer input variables.
• Model performance remained consistent across different data preprocessing techniques.

Conclusions:

• Machine learning models can accurately predict the need for RRT within 24 hours of ICU admission.
• Top feature models offer an efficient approach to RRT prediction, balancing accuracy and complexity.
• These predictive tools can support clinical decision-making and resource allocation in critical care settings.