Translating Time-Series Foundation Models into AI-Enhanced Medication Inventory Control in a Large Urban Health System

Hospital medication inventory management must maintain medication availability under sparse, volatile cabinet-level demand while limiting excess stock and pharmacy refill workload. Motivated by the empirical success of pretrained time-series foundation models (TSFMs) in retail demand forecasting, we develop a framework that converts probabilistic demand forecasts from TSFMs into interpretable replenishment policies. We use demand scenarios derived from these forecasts to optimize an interpretable replenishment policy that minimizes expected holding cost subject to stockout-risk and refill-workload constraints. We solve the model using a differentiable primal--dual procedure and evaluate our approach in a large urban health system in New York City to inform future implementation. Compared with the historical policy, our policy achieves a better trade-off across various outcomes: across nine medication–station pairs, the learned policy reduces refill workload in seven and holding cost in three, while achieving fewer or equal stockout days in six. Aggregated across both medications, it simultaneously reduces average holding cost, refill events, and total stockout days by 3\%--28\%. These results demonstrate the promise of forecast-driven cabinet policies for various operational objectives.