From appointments to inventory: Predicting amount of daily stocked blood by blood type

Maintaining adequate blood inventories requires reliable short-term estimates of available blood units, yet daily availability varies due to calendar effects, donor behavior, and operational constraints. This study presents an interpretable computational pipeline for predicting daily stocked amounts of whole-blood units across major ABO/Rh blood types in a centralized transfusion center in Czechia.

We employed linear regression with automated forward selection and backward elimination to identify statistically significant predictors. Models were trained on retrospective daily data from 2021 to 2024, with 2025 reserved as an independent test set. Features included school holidays, proximity to non-working days, weekday and month indicators, national influenza surveillance variables, and counts of scheduled donors from two hospitals. Performance was evaluated using normalized mean absolute error (nMAE) and compared with a naïve baseline based solely on scheduled donations.

Across blood types, nMAE ranged from 6.49% (A+) to 26.26% (AB-). The proposed models consistently outperformed the naïve baseline and achieved performance comparable to other machine learning approaches. Scheduled donations were the dominant predictors, with coefficients close to one, while weekday and holiday variables provided additional improvements. Influenza indicators showed weak but statistically significant associations.

The proposed interpretable modeling approach provides accurate short-term predictions of blood inventory levels. Incorporating calendar and epidemiological variables offers modest improvements beyond scheduled donations, supporting more effective blood supply management.