Size (and frequency) matter: Evaluating the best management practices for blood products with spotty demand

Blood banks often use simple ordering rules that replenish stock to a target level when inventory falls below a trigger point. However, for products such as Low-Titer group O Whole Blood (LTOWB), where utilization is intermittent, simple rules may not always be effective. This paper evaluates ordering policies for blood products with intermittent ("spotty") demand.

LTOWB inventory was evaluated under zero-inflated Poisson (ZIP) (i.e., spotty) demand. Four strategies were compared against an exact benchmark: (i) a classical reorder-point/order-up-to policy, (ii) a scenario-based rolling horizon heuristic (Sc-RHA), (iii) model-based approximate dynamic programming (ADP), and (iv) model-free reinforcement learning using Proximal Policy Optimization (PPO). Performance was assessed using total expected cost relative to the exact solution.

Simple reorder-point/order-up-to policies work well when demand is consistent or regular. When expected daily demand is approximately 1.0 unit/day, this simple policy is close to optimal. Sc-RHA reduces the threshold for reliable performance to approximately 0.4 units/day by incorporating short-term age and demand information. ADP remains stable and near-optimal across the full demand range, including highly sparse settings. However, PPO is effective only when an extremely large amount of simulated training data is available, which is operationally impractical in typical LTOWB settings.

Centers with robust demand of at least 1.0 unit/day can safely rely on simple reorder-point/order-up-to policies. For locations with demand between 0.4 and 1.0 unit/day, Sc-RHA provides an age-aware alternative with modest added complexity. For extreme sparsity, ADP becomes the most appropriate approach.