Dual-phase machine learning framework for paediatric ventilator liberation: forecasting extubation readiness and nowcasting extubation outcomes

Objective To develop and validate a machine learning framework for predicting optimal extubation timing in paediatric intensive care units (PICUs), addressing the clinical need for continuous monitoring of weaning readiness and extubation safety. Methods and analysis We developed a dual-phase machine learning framework using data from 3815 ventilation episodes across two UK PICUs. The framework includes (1) a forecasting model that predicts extubation readiness up to 12 hours in advance, and (2) a nowcasting model that assesses immediate extubation failure risk. Both models used hourly time-series data and were trained using Gated Recurrent Units (GRUs) to capture temporal patterns. Area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve were used for evaluation on a held-out test set. Results The forecasting model achieved an AUROC of 0.85 (95% CI 0.845 to 0.855) and the nowcasting model reached 0.77 (95% CI 0.760 to 0.780). Forecasting performance benefited from longer observation windows (up to 24 hours) and incorporated broader clinical factors (medications, patient characteristics), while nowcasting relied primarily on immediate ventilation parameters. The temporal evolution of risk scores differed markedly between the two models, with the nowcasting model erroneously signalling early readiness. Conclusion Our findings demonstrate that distinct modelling strategies are needed to support different stages of ventilator liberation. The proposed dual-phase framework reflects the sequential nature of clinical decision-making and enables structured monitoring throughout the ventilator liberation process. This study represents the largest machine learning investigation of PICU ventilator liberation to date and demonstrates the feasibility of artificial intelligence-assisted extubation timing in paediatric critical care.