TY - JOUR
T1 - Prescriptive analytics for hospital drug inventory
T2 - A context-aware multi-channel neural network for integrated forecasting and optimization
AU - Chen, Jie
AU - Xu, Zeshui
AU - Zhu, Bin
N1 - Publisher Copyright:
© 2026 Elsevier Ltd
PY - 2026/9
Y1 - 2026/9
N2 - Inventory management in hospital pharmacy settings focuses on two interdependent tasks: forecasting drug demand and determining ordering quantities under uncertainty. However, effectively integrating these tasks remains challenging due to the underutilization of contextual data and the disconnect between prediction accuracy and decision quality. This study proposes a prescriptive analytics framework for hospital drug inventory that integrates forecasting and decision-making within a context-aware multi-channel neural network. Specifically, for demand forecasting, we develop a multi-channel neural network that captures heterogeneous temporal patterns from historical data including ordering records, epidemiological indicators, behavioral signals, environmental variables, and demographic factors. For inventory decisions, we train the model to directly generate ordering quantities by minimizing a decision-oriented loss function that accounts for the asymmetric costs of overstocking and understocking. We prove that incorporating multi-source contextual data and integrating prediction with decision-making leads to lower expected costs compared to conventional two-stage methods. Empirical results based on real data from hospitals in Beijing support this finding, showing consistent improvements in cost performance and inventory stability.
AB - Inventory management in hospital pharmacy settings focuses on two interdependent tasks: forecasting drug demand and determining ordering quantities under uncertainty. However, effectively integrating these tasks remains challenging due to the underutilization of contextual data and the disconnect between prediction accuracy and decision quality. This study proposes a prescriptive analytics framework for hospital drug inventory that integrates forecasting and decision-making within a context-aware multi-channel neural network. Specifically, for demand forecasting, we develop a multi-channel neural network that captures heterogeneous temporal patterns from historical data including ordering records, epidemiological indicators, behavioral signals, environmental variables, and demographic factors. For inventory decisions, we train the model to directly generate ordering quantities by minimizing a decision-oriented loss function that accounts for the asymmetric costs of overstocking and understocking. We prove that incorporating multi-source contextual data and integrating prediction with decision-making leads to lower expected costs compared to conventional two-stage methods. Empirical results based on real data from hospitals in Beijing support this finding, showing consistent improvements in cost performance and inventory stability.
KW - Demand forecasting
KW - Inventory decision-making
KW - Multi-channel neural network
KW - Multi-source data
KW - Prescriptive analytics
UR - https://www.scopus.com/pages/publications/105033405185
U2 - 10.1016/j.omega.2026.103557
DO - 10.1016/j.omega.2026.103557
M3 - Article
AN - SCOPUS:105033405185
SN - 0305-0483
VL - 143
JO - Omega (United Kingdom)
JF - Omega (United Kingdom)
M1 - 103557
ER -