Data subdivision approach enhances machine learning-based mortality prediction in pediatric ICU patients

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Pediatric Critical Care

Background:

• Predicting mortality in pediatric intensive care units (PICU) is crucial for patient management.
• Machine learning (ML) offers potential for improving mortality prediction accuracy.
• Addressing missing data and class imbalance are key challenges in developing robust ML models.

Purpose of the Study:

• To evaluate ML models for predicting all-cause mortality in PICU patients.
• To assess the impact of comprehensive biochemical panels on prediction accuracy.
• To compare ML strategies for handling missing data and class imbalance.

Main Methods:

• Retrospective analysis of 8,629 PICU patients using 22 biochemical variables.
• Multiple Imputation for missing data and Synthetic Minority Oversampling Technique (SMOTE) for class imbalance.
• Evaluation of ML models including Random Forest, CatBoost, Extra Trees, Logistic Regression, and ensemble stacking.

Main Results:

• Ensemble stacking achieved the highest AUC-ROC (0.88) and AUC-PRC (0.45).
• Coagulation markers, electrolytes, and organ-dysfunction indicators were key predictors of mortality.
• Stacking models demonstrated superior probabilistic accuracy with the lowest Brier score (0.04).

Conclusions:

• Ensemble stacking is a more effective strategy than data subdivision for addressing class imbalance in PICU mortality prediction.
• ML models utilizing comprehensive biochemical panels can significantly improve mortality prediction in pediatric ICU settings.