Introduction: Pediatric mortality prediction remains a critical task in intensive care. Traditional statistical approaches struggle to fully utilize rich patient data such as vitals and lab results. This study proposes a novel framework, PEDICTOR, that integrates expert domain knowledge with machine learning (ML) and deep learning (DL) techniques to improve mortality prediction in pediatric intensive care unit (PICU) patients. Methods: Pediatric mortality prediction remains a critical task in intensive care. Traditional statistical approaches struggle to fully utilize rich patient data such as vitals and lab results. This study proposes a novel framework, PEDICTOR, that integrates expert domain knowledge with machine learning (ML) and deep learning (DL) techniques to improve mortality prediction in pediatric intensive care unit (PICU) patients. Results: The PEDICTOR stacker achieved an average per-minute F1-score of 0.441 (95% CI: 0.439–0.442) and an AUCROC of 0.932 (95% CI: 0.9314–0.9325) for predicting expired cases. The framework also attained an average 30-minute prediction F1-score of 0.591 (95% CI: 0.589–0.592) and an average hourly prediction F1-score of 0.649 (95% CI: 0.648–0.651). Results were consistent across all five pediatric age groups. Conclusions: PEDICTOR offers a robust, age-aware mortality prediction system leveraging ensemble ML-DL stacking and cluster-based oversampling. Its ability to accurately predict per-minute, 30 minute, and hourly outcomes can support timely clinical decision-making. This framework can help clinicians intervene early, allocate resources efficiently, and potentially save lives in pediatric critical care settings.
Siddiqui et al. (Sun,) studied this question.