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

Abstract

Objective To evaluate machine learning–based models for predicting all-cause mortality in pediatric ICU patients using comprehensive biochemical panels, with a focus on addressing missing data and class imbalance. Materials and methods A retrospective analysis was performed on a publicly available PICU dataset comprising 8,629 patients aged 28 days to 18 years. Twenty-two biochemical variables measured on the first PICU day were analyzed. Missing values were addressed using Multiple Imputation. Lasso regression was applied for feature selection. Models were trained using 5-fold cross-validation with the Synthetic Minority Oversampling Technique (SMOTE). Two ML-based imbalance-handling strategies, stacking ensemble and data subdivision were evaluated. Pairwise DeLong tests were used to compare AUC performance across models. Results Among the 8,629 included patients, there were 476 non-survivors (5.5 percent). Multiple imputation followed by SMOTE improved model performance across all algorithms. The single-model classifiers achieved AUC-ROC values of 0.82 (Random Forest), 0.79 (CatBoost), 0.83 (Extra Trees), and 0.79 (Logistic Regression). The stacking ensemble demonstrated the best overall performance, with an AUC-ROC of 0.88 and an AUC-PRC of 0.45. The data subdivision approaches also produced strong discriminative performance, achieving AUC-ROC value up to 0.83 for three-subdivision, and 0.82 for five-subdivision strategies. Calibration analysis showed that the stacking model achieved the lowest Brier score (0.04), indicating superior probabilistic accuracy compared with individual classifiers. Feature importance analyses across all MI-based models consistently highlighted coagulation markers (D-dimer, reference TT, PTT, INR), electrolytes (chloride, potassium, sodium), and metabolic and organ-dysfunction indicators (AST, ALT, creatinine) as key predictors of mortality. Conclusions This study demonstrates that ensemble stacking is a more effective strategy than data subdivision for addressing class imbalance in PICU mortality prediction.