BMC Med Inform Decis Mak. 2025 Oct 10;25(1):371. doi: 10.1186/s12911-025-03211-4.
ABSTRACT
OBJECTIVE: To develop and evaluate machine learning models combined with survival analysis for predicting 7-, 14-, and 28-day mortality in critically ill children with acute kidney injury (AKI), identifying key predictors to guide risk stratification and early intervention.
METHODS: Using the Pediatric Intensive Care (PIC) database, we analyzed data from 3,624 children with AKI admitted between 2010 and 2018. Nine machine learning algorithms, including CatBoost, were trained to predict mortality, with feature importance assessed via SHapley Additive exPlanations (SHAP). Time-to-event analyses, including Kaplan-Meier and restricted cubic spline methods, examined the temporal impact of predictors on 28-day mortality, stratified by age and AKI stage.
RESULTS: CatBoost achieved the highest area under the curve (AUC) values: 0.871 (95% CI: 0.824-0.918) for 7-day, 0.871 (95% CI: 0.829-0.913) for 14-day, and 0.867 (95% CI: 0.829-0.905) for 28-day mortality. Lactate was the top predictor across all models. Time-to-event analyses revealed a linear association between elevated lactate (cut-off: 1.5 mmol/L) and 28-day mortality (p-overall < 0.001), with stronger effects in infants (0-3 years) and AKI stage 1 patients (HR > 1).
CONCLUSIONS: Machine learning, particularly CatBoost, combined with survival analysis, accurately predicts AKI-related mortality in critically ill children, with lactate as a pivotal marker. These findings support precision risk stratification and early lactate-targeted interventions, though multicenter validation is needed for clinical adoption.
PMID:41074021 | DOI:10.1186/s12911-025-03211-4
