Optimising hospital pharmacy inventory for high-value, intermittent-demand drugs using a machine-learning approach with LightGBM-Tweedie regression: a single-centre, retrospective study.

OBJECTIVE: Hospital pharmacies operate under a persistent dilemma: ensuring uninterrupted access to essential medicines while minimising capital tied up in inventory. This challenge is most acute for drugs with lumpy or erratic demand, where traditional manual ordering often fails because of limited forecasting capability. Therefore, this study aimed to develop and evaluate a LightGBM-Tweedie regression model for forecasting weekly demand of high-value intermittent-demand drugs.

METHODS: We conducted a retrospective machine learning-based study using weekly pharmacy usage data from a large tertiary hospital in Beijing, China, from January 2021 to December 2024. Data were split into a training set (2021–2023) and a testing set (2024). Stock keeping units were classified by the Syntetos–Boylan average demand interval/coefficient of variation squared framework into smooth, erratic, lumpy and intermittent demand. The LightGBM-Tweedie model was compared with naïve, autoregressive integrated moving average, random forest and long short-term memory models. Performance was assessed using weighted mean absolute percentage error (wMAPE) on the 2024 test set and stratified by demand pattern. Financial impact was evaluated by comparing model forecasts with historical manual procurement orders using overstock and shortage cost proxies.

RESULTS: In the targeted analysis of complex demand patterns, our model achieved a wMAPE of 21.4%, with approximately 3% improvement on the baseline naïve model. Financially, the analysis revealed a critical operational insight: while manual procurement led to minimal overstock but shortage costs significantly higher than overstock costs, the model substantially reduced estimated shortage-related financial losses while moderately increasing holding costs. By implementing a strategic inventory buffer, the model reduced the estimated financial impact of stockouts by approximately 70% for erratic-demand drugs compared with manual practices.

CONCLUSION: The developed framework may support more balanced inventory decision-making by mitigating systematic under-ordering behaviours. Rather than replacing professional judgement, it functions as a risk-stratification tool that provides a quantitative basis for balancing holding costs and availability. Integrating such decision-support systems into hospital policy potentially offers a scalable pathway to reduce clinical shortages and enhance the financial resilience of the pharmaceutical supply chain.