Forecast-Driven Dynamic Physician Staffing in a Pediatric Emergency Department: A Prospective Quasi-Experimental Pilot Study.

Abstract

Emergency department crowding is a persistent threat to acute care quality, yet predictive models for ED demand have rarely been translated into prospective operational staffing interventions. Here we report a prospective single-center quasi-experimental pilot study evaluating forecast-driven dynamic physician scheduling in the pediatric ED of Hacettepe University Ihsan Dogramaci Children's Hospital (December 2024 to May 2025). Using a deep learning demand forecasting model (TiDE-RIN) combined with linear programming, we determined daily physician counts (range 3 to 6) for evening shifts (16:00 to 24:00) during days 1 to 15 of each month; days 16 to end of month maintained the institution's standard fixed four-physician schedule. Among after-hours visits with valid disposition timestamps (n = 9,626), mean post-evaluation length of stay (PE-LOS) was 175.9 min in the intervention arm and 184.0 min in the control arm (unadjusted difference 8.2 min; propensity-score-matched 7.9 min, median 10.0; stabilised inverse-probability weighting 8.5 min). Because staffing was allocated by calendar day, inference was clustered on the 182 study days: the two-way fixed-effects arm contrast was a reduction of 8.9 min (95% CI - 22.5 to + 4.8; p = 0.20) and the fully covariate-adjusted contrast 11.0 min (95% CI - 24.4 to + 2.5; p = 0.11). All design-consistent specifications placed the reduction between 4 and 11 min, none excluded no effect under day-level clustering, and the pilot was not powered for a difference of this size. Diagnostic testing was modestly lower (any-test rate 0.51 versus 0.54; p = 0.009). Effects were concentrated among lower-acuity patients and during the early-evening demand peak, and the optimised schedule allocated physicians demand-responsively (mean 4.31 versus 4.00 per shift) with no signal of compromised short-term patient safety, although per-physician workload effects did not reach significance in this pilot. Spillover analysis confirmed no progressive improvement in the concurrent control period. In this prospective pilot, coupling deep learning demand forecasting with optimisation-based physician scheduling was associated with a consistent but modest reduction in PE-LOS whose confidence interval included no effect, supporting feasibility and motivating multi-centre evaluation, and directly addressing the translational gap between predictive model development and real-world clinical implementation.