Performance Evaluation for Fresh Blood Component Demand Forecasting and Allocation

Abstract

Efficient inventory management for fresh blood components is critical for patient care but is challenged by perishability and demand uncertainty. Although demand forecasting is increasingly integrated into replenishment decisions, forecasting models are often evaluated using statistical accuracy metrics, while inventory performance is assessed using cost measures. This separation limits understanding of how forecast errors affect downstream outcomes. This study examined when improved forecast accuracy translates into operational benefit under alternative replenishment policies, using platelet inventory as a case study. A two-stage framework was developed. Stage 1 generated one-day-ahead platelet demand forecasts using Long Short-Term Memory (LSTM) models trained with mean squared error (MSE), mean absolute error (MAE), Huber loss, and mean absolute percentage error (MAPE) loss functions under baseline demand and four uncertainty scenarios. Stage 2 embedded forecasts into perishable inventory base-stock policies: forecast-only (FO) and data-driven (DD), while classical (CC), which does not use forecasts, served as the baseline policy. Forecasting and operational performance were evaluated using twelve accuracy metrics and annual total cost, respectively. As demand uncertainty increased from Scenario 1 to Scenario 3, forecast error and annual total cost increased across policies. Models trained with MSE, MAE, and Huber loss produced comparable forecast accuracy and annual total costs under DD and FO, whereas models trained with MAPE produced substantially higher forecast error. Under FO, this poor accuracy translated into much higher annual total cost, reaching 46,988 CU in Scenario 3 compared with 14,592 CU for the model trained with MSE. Under DD, however, the same poor forecasts produced annual total costs comparable to those of the model trained with MSE (11,204 CU versus 11,243 CU) because DD adaptively tuned forecast reliance and safety stock weekly. These findings show that the operational value of improved forecast accuracy is policy-dependent. They support decision-aware evaluation, where forecasting models and inventory policies are assessed jointly. For hospital managers, an adaptive DD policy could offer a robust no-regret choice, performing comparably to the CC policy when forecast accuracy is poor while delivering measurable gains when forecasts are reliable.