Hospital trust network features inform machine learning for medicines demand forecasting and resilient supply chain decisions

Hospital pharmacies require short-horizon demand forecasts that remain useful across thousands of medicines with widely different issue volumes. This study evaluates an artificial intelligence (AI)-enabled feature-engineering framework for adaptive decision support using 72 corrected finalised Secondary Care Medicines Data (SCMD) files published by the NHS Business Services Authority (NHSBSA) for April 2019 to March 2025. The analysis processes 22.67 million records into 522,813 product-month observations and 488,924 exact next-calendar-month positive-issue forecasting rows. Demand volatility is treated as an operational signal for disruption forecasting because unstable medicine flows can increase the risk of review delays, emergency purchasing or stock imbalance even when shortage labels are unavailable. Target construction, lag variables and expanding product summaries are calendar-indexed and use only information available at the forecast origin. Descriptive statistics, unit-root and trend diagnostics, temporal holdout validation, seven full rolling-origin refits, trend-aware baselines and paired Wilcoxon tests are used. The raw holdout target is extremely skewed (skewness 38.05), whereas log(1 + quantity) is approximately symmetric (skewness 0.04). Trust-network-informed Light Gradient Boosting Machine (LightGBM) achieves the lowest holdout root mean squared logarithmic error (RMSLE = 0.4832) and mean rolling-origin RMSLE (0.4910), while rolling means achieve the lowest weighted absolute percentage error (WAPE). A transparent monitoring score identifies 10.0% of product-months that account for 69.0% of actual quantity and 65.1% of predicted indicative-cost exposure. By combining real-world operational data, rigorous temporal validation, and interpretable monitoring outputs, this work contributes to AI-enabled decision support for resilient supply chains under uncertainty.