Prospective Validation of Multi-week Inpatient Census Forecasting for Pediatric HSCT and Cellular Therapy.

BACKGROUND: Accurate inpatient census forecasting is important for cell therapy and blood and marrow transplantation (BMT) programs because bed capacity, staffing, and coordination of transplant care depend on anticipating occupancy several weeks in advance. Forecasting BMT census is challenging because the service combines planned admissions, including scheduled transplants, with less predictable admissions related to treatment complications. Existing literature has focused on shorter-term prediction, leaving limited evidence for forecasting over the 15- to 30-day horizons most relevant to BMT operational planning. We therefore developed a forecasting approach that reflects the operational structure of BMT care by modeling the distinct processes that determine census rather than treating census as a single time series. OBJECTIVE: To develop and prospectively validate a component-based model for forecasting pediatric hematopoietic cell transplantation and cellular therapy inpatient census 15 to 30 days in advance. STUDY DESIGN: We performed retrospective model development followed by prospective validation of a deployed forecasting workflow. The model decomposed future census into expected admissions and expected discharges. Forecasts of upcoming admissions were obtained by integrating known, planned transplant schedules with predicted, unplanned admissions estimated from historical admission patterns. For patients already hospitalized, discharge probabilities were estimated at the patient level using gradient-boosted decision trees. These components were combined to generate daily census forecasts over 15-, 30-, and longer-range horizons. Performance was assessed using forecast accuracy metrics including mean absolute percentage error (MAPE) and root mean squared error (RMSE). RESULTS: In retrospective evaluation, the component-based model achieved strong performance across clinically relevant forecast horizons. Accuracy was strongest at the most operationally useful intervals, with MAPE of 13% at 15 days and 16% at 30 days. These results suggest that integrating known transplant schedules with modeled unplanned admissions and patient-level discharge probabilities can provide accurate medium-range census forecasts for our overall BMT service. In prospective validation after deployment, performance remained stable in real-world use, confirming feasibility of routine operational forecasting and reducing concern for retrospective information leakage. CONCLUSIONS: Accurate 15- to 30-day inpatient census forecasting for pediatric hematopoietic stem cell transplantation (HSCT) and cellular therapy programs is achievable when models are integrated with transplant operational data and aligned with the mechanisms that drive occupancy. A component-based framework combining scheduled transplant data, modeled unplanned admissions, and patient-level discharge prediction performed well in both retrospective and prospective evaluation. This approach can support more reliable capacity planning for HSCT and cellular therapy programs and may be adaptable to other complex inpatient services with both planned and unplanned utilization.