ARTIFICIAL INTELLIGENCE FOR MORTALITY RISK PREDICTION IN PEDIATRIC INTENSIVE CARE UNITS: DEVELOPMENT, VALIDATION, AND CLINICAL IMPACT ANALYSIS

Abstract

Background: Accurate mortality risk prediction in pediatric intensive care units (PICUs) supports clinical decision-making, family communication, and resource allocation. Objective: To develop and externally validate an AI-based mortality prediction model for critically ill children and assess clinical utility compared to established pediatric mortality scores. Methods: A multicenter study across 12 PICUs included 18,437 admissions over three years. XGBoost ensemble models were trained on the first 24 hours of physiological, laboratory, and clinical data. Results: The AI model achieved AUC-ROC of 0.93 (95% CI: 0.91–0.95) in external validation, significantly outperforming PRISM-IV (AUC-ROC 0.84) and PIM-3 (AUC-ROC 0.83). Calibration was excellent (Brier Score 0.06). Clinician AI-assisted decisions showed 18.3% improvement in mortality prediction accuracy. Conclusion: AI-based mortality prediction substantially outperforms established pediatric critical care scoring systems and improves clinical prognostic accuracy when used as a decision support tool.