Dynamic Prediction of Intensive Care Unit Transfer and Critical Interventions After Pediatric Allogeneic...

National and international hematopoietic stem cell transplantation (HSCT) societies recommend the use of early warning systems to detect clinical deterioration after pediatric HSCT. Commonly utilized tools, such as vital-sign-based early warning scores, do not incorporate pretransplantation risk factors, rely on categorization of vital sign ranges, and generate static scores applicable to single points in time. To predict the future need for (1) ICU transfer and (2) ICU transfer with critical intervention following pediatric allogeneic HSCT by integrating baseline demographic and transplant characteristics with dynamic changes in post-transplantation physiologic variables. Retrospective observational cohort study of patients admitted for first allogeneic HSCT at Children's National Hospital from January 1, 2012, through December 31, 2022. Patients were observed for ICU outcomes from transplantation to day 100 or hospital discharge. Statistical analyses included Kaplan-Meier survival analysis and time-dependent univariate and multivariate Cox proportional hazards regression. We also report 100-day and 1-year mortality. Among 307 pediatric HSCT recipients, 73 (24%) required ICU transfer and 37 (12%) received ≥1 critical intervention (invasive mechanical ventilation, vasoactive medication administration, and/or renal replacement therapy). The cumulative incidence of both outcomes plateaued by day 45 post-HSCT, with 65% of at-risk patients remaining free of ICU transfer and 81% free of ICU critical interventions. The final multivariate Cox proportional hazards models incorporated pretransplantation variables and dynamic post-transplantation physiologic changes, demonstrating excellent performance with daily C-statistics ranging from 0.83 to 0.86 for ICU transfer and 0.89 to 0.93 for ICU transfer with critical intervention. Overall mortality was 6% in 100 days and 15% in 1 year, but was higher among patients requiring ICU transfer (22% in 100 days, 37% in 1 year) and highest among those requiring critical interventions (38% in 100 days, 65% in 1 year). Dynamic modeling of pretransplantation factors and evolving post-transplantation physiologic trajectories can quantify the changing risk of clinical deterioration over time after pediatric HSCT. The performance of our models is appropriate for implementation, meeting the recommendations of national and international societies for the use of early warning systems. Implementation of dynamic early warning systems such as this may enable earlier recognition, enhanced surveillance, and potentially improved outcomes for high-risk patients.