AI and Blood Banking: Predicting Transfusion Demand – A Systematic Review of Forecasting Approaches

Transfusion Medicine and Hemotherapy

Systematic Review

Merlin Engelke¹  Osman Alperen Çinar-Koraş¹  Jens Kleesiek^1,2,3  Peter A. Horn⁴ 

[1] Institute for Transfusion Medicine, University Medicine Essen, Essen, Germany

[2] Institute for Artificial Intelligence in Medicine, University Medicine Essen, Essen, Germany

[3] Faculty of Computer Science and Medical Faculty, University of Duisburg-Essen, Essen, Germany

[4] Department of Physics, TU Dortmund University, Dortmund, Germany

关键词: Machine learning;    Blood transfusion;    Demand prediction;    Forecasting;    Platelets;   

DOI  :  10.1159/000552969

 received in 2026-02-17, accepted in 2026-05-12,  发布年份 2026

来源: Karger

【 摘 要 】

AbstractIntroduction: Accurate prediction of blood product demand is essential for maintaining adequate supply while minimizing wastage. Artificial intelligence (AI) and machine learning (ML) approaches have emerged as promising tools for transfusion demand forecasting across multiple scales, ranging from individual patients to regional supply chains. This review provides a structured overview of AI applications in blood banking, organized by prediction scope and clinical objective. Methods: We conducted a systematic review following PRISMA 2020 guidelines, searching PubMed and Scopus for studies published between January 2015 and December 2025. Studies applying AI or ML to transfusion demand prediction were included. We organized findings into a four-tier framework based on prediction scope and clinical goal and systematically compared technical approaches. Results: Out of 535 identified records, supplemented by a targeted search during peer review, 24 studies met inclusion criteria. The four-tier framework revealed distinct prediction challenges: (1) patient-individual prediction (goal: improve individual care quality) achieved areas under the curve (AUCs) of 0.45–0.97 across five studies for 24-h to 72-h prediction, (2) procedure-specific models (goal: optimize surgical blood ordering) achieved AUCs of 0.70–0.91 across six studies, (3) facility-level optimization (goal: reduce wastage and shortages) approximately halved platelet outdating rates and achieved mean absolute percentage errors as low as 4.18% across eight studies for daily to weekly forecasting, and (4) regional forecasting (goal: secure long-term supply) achieved R2 values up to 0.85 across five studies for monthly forecasting. Deep learning excelled for temporal patterns in individual patients, while traditional statistical methods (ARIMA, SARIMA) proved effective for aggregated regional forecasting, where deep learning approaches showed no consistent advantage over simpler models. Prediction horizons varied by blood product: 3–7 days for platelets versus weekly to monthly for red blood cells. Conclusion: AI-driven demand prediction offers substantial potential to improve blood supply management across all tiers of the blood supply chain. The choice of modeling approach should be guided by the prediction scope, available data infrastructure, and specific clinical objectives. Standardized external validation remains critical for clinical deployment.

【 授权许可】

Others   
© 2026 The Author(s). Published by S. Karger AG, Basel