Blood Procurement With Prescriptive Analytics and Deep-Learning Method
本研究以血小板为例，利用规范分析和深度学习方法，在供应不确定和需求分优先级（紧急需求优先）下直接预测最优订购量，通过真实数据实验证明两种方法均优于传统方法，且规范分析在小数据下稳健，深度学习在大数据下成本更优。
Blood is a scarce and valuable resource in healthcare, making the effective management of blood inventory a persistent challenge for blood centers and hospitals due to irregular supply and heterogeneous demand. To address severe blood shortages, this study focuses on platelets as a representative case. It employs data-driven prescriptive analytics and deep learning to manage inventory under uncertain supply and prioritized demand, where emergency demand takes precedence over regular demand. First, the prescriptive analytics approach combines a local learning predictive method with optimization to directly predict order quantities based on historical demand and observed features, utilizing a distribution-free, one-step machine learning algorithm. Second, the deep neural network approach utilizes historical demand and feature data to build a neural network that produces optimal ordering decisions. Both approaches are evaluated against conventional methods through numerical experiments using real data from a regional blood center. Extensive multi-year out-of-sample evaluations demonstrate that prescriptive analytics and deep learning consistently outperform traditional data-driven methods across multiple performance metrics. Their relative performance evolves with dataset size: prescriptive analytics approaches exhibit strong robustness under limited data, whereas deep learning achieves superior cost efficiency as historical data accumulates. The results remain stable under varying cost parameters and feature configurations, confirming the robustness and scalability of the proposed framework. Feature selection is also crucial for improving algorithm performance. The findings provide both theoretical and practical contributions by offering insights into demand behavior and supporting cost-efficient, patient-centered blood inventory management strategies.