Machine Learning Models for Dynamic Assessment of Extubation Readiness in Pediatric Critical Care

Determining the optimal timing for extubation in critically ill children remains challenging, with premature extubation leading to increased morbidity and mortality, while prolonged ventilation exposes patients to ventilator-associated complications. To develop and validate machine learning models for dynamic assessment of extubation failure risk (nowcasting) and extubation readiness (forecasting) in mechanically ventilated children. Retrospective cohort study using electronic health records from two pediatric intensive care units in London, UK (2013-2022), including 3,815 ventilation episodes in children aged 0-17 years. Mechanical ventilation via endotracheal tube using pressure control-BIPAP (PC) or spontaneous pressure support CPAP (PS) modes. Primary outcomes were extubation failure (requiring reintubation within 48 hours) and extubation readiness (successful extubation within 12 hours). Both models incorporated demographic, physiological, ventilation, and medication data, with varying historical context lengths to optimize prediction accuracy. The median age of children in the study cohort (n=3815 ventilation episodes) was 8.0 months (56.2% male); extubation failure occurred in 315/3815 (8.3%). The nowcasting model achieved an area-under-the-receiver-operating-characteristic curve (AUROC) of 0.77. The forecasting model reached an AUROC of 0.85. Ventilation parameters dominated the nowcasting model, while medication response and patient characteristics drove the forecasting model. Our dual-model approach offers a structured framework for extubation decision-making in critically ill children, combining continuous monitoring of readiness with snapshot assessment of extubation failure risk. Prospective validation is needed; however, this strategy may help clinicians optimize the timing of ventilation liberation within pediatric intensive care.