Enhancing hospital drug and medical supply request processes through a clinically prioritized and demand-variability-aware adaptive decision support framework.

Abstract

Rule-based replenishment thresholds in hospital information systems (HIS) are feasible and auditable but may respond slowly to demand variability and clinical criticality. To evaluate a decision-support framework that preserves a three-level request structure (minimum/critical/maximum) while adapting thresholds using short-horizon demand forecasts and uncertainty, stratified by VEN clinical priority (Vital-Essential-Non-essential), and to benchmark its forecasting component against a naïve rolling 180-day mean predictor (μ180). A retrospective unit-item-day panel was constructed from routine hospital records and classified by VEN. XGBoost forecasted 14- and 30-day cumulative demand; uncertainty was modeled using a rolling 30-day consumption standard deviation scaled by VEN-specific coefficients. Adaptive targets were mapped to existing HIS request levels and backtested against the baseline rule (7×/15×/30 × the rolling 180-day mean daily consumption) on a 30-day ML-available panel (35,801 unit-item-day records; 1269 series; 2025-09-30-2026-01-22). Forecast accuracy was compared with μ180 using RMSE and WAPE, supported by calibration and Diebold-Mariano testing. Policies were evaluated using paired Wilcoxon tests and lead-time sensitivity analyses. The proposed policy reduced stock-out days (54 to 10; -81.5%), days below critical (11,612 to 6396; -44.9%), ordering days (1936 to 550; -71.6%), and ordered quantity (853,119 to 579,795; p < 0.001). Stock-outs were eliminated for Vital items. XGBoost reduced WAPE by 62%-71% versus μ180 across horizons, with absolute-error superiority confirmed by Diebold-Mariano testing. Inventory exposure increased (366,406 to 802,049), indicating a measurable safety-inventory trade-off. VEN-prioritized, forecast- and volatility-aware threshold adaptation improved service continuity and ordering efficiency within an auditable HIS workflow while quantifying trade-offs under alternative lead-time assumptions.