Early prediction of unplanned critical care transfers in children using EHR-based ensemble machine learning.

Background: Early recognition of unplanned ICU transfer is critical for improving outcomes, yet existing pediatric warning systems lack accuracy and timeliness. We developed machine learning models predicting unplanned ICU transfer up to 24 hours in advance. Methods: We conducted a retrospective cohort study using electronic health record data at a quaternary-care children's hospital with 175 PICU/NICU beds. Patients aged 0-24 years admitted to medical or surgical wards for ≥30 hours between 2019 and 2024 were included. The cohort was divided into development (2019-2022) and external validation (2023-2024). The F-WIN model is an ensemble of extreme gradient-boosting models trained at varying prediction horizons (1-24 hours). The primary outcome was unplanned transfer to PICU/NICU, defined as transfer followed by clinical deterioration within 12 hours, including death, administration of vasopressors or inotropes, intubation, or new non-invasive ventilation. We compared F-WIN's performance with the bedside Pediatric Early Warning System and the Watcher and Clinician Concern programs. Metrics included areas under the receiver operating characteristic (AUROC) and precision-recall curve (AUPRC). Results: Among 72 419 floor stays (49 283 patients) in the full dataset, 816 stays (1.1%) involving 769 patients (1.6%) ended in unplanned ICU transfer. Using 647 variables, F-WIN achieved strong discrimination across horizons (8-hour AUROC: 0.93; [95%CI, 0.91-0.95], AUPRC: 0.36 [95%CI, 0.29-0.43]); 24-hour AUROC: 0.91 ([95%CI, 0.89-0.93], AUPRC: 0.22 [95%CI, 0.17-0.29]) and outperformed all comparators. A parsimonious model preserved performance (24-hour Conclusions: F-WIN accurately predicted unplanned transfer up to 24 hours in advance, outperforming existing tools. A parsimonious model maintained similar performance.