Machine learning-enabled precision transfusion: research progress, challenges, and prospects
10.13303/j.cjbt.issn.1004-549x.2026.07.023
- VernacularTitle:机器学习赋能精准输血：研究进展、挑战与展望
- Author:
Jiang DENG
1
;
Chaojie WANG
1
;
Ning ZHAO
1
;
Liping LYU
1
;
Ping MA
1
;
Ke ZHANG
1
;
Yanyu ZHANG
1
Author Information
1. Academy of Military Medical Sciences, Beijing 100850, China
- Publication Type:Journal Article
- Keywords:
machine learning (ML);
transfusion prediction;
patient blood management;
clinical decision support;
precision transfusion
- From:
Chinese Journal of Blood Transfusion
2026;39(7):967-976
- CountryChina
- Language:Chinese
-
Abstract:
Blood transfusion is an important life-support measure in modern medicine. As clinical demand for transfusion continues to rise, blood supply remains under persistent strain, making the efficient utilization of this precious medical resource a critical component of blood management. Accurate prediction of transfusion needs is of great significance for optimizing blood resource allocation and conserving blood products. Traditional transfusion decision-making relies on clinical experience and simplified scoring systems, which are insufficient to meet the demands of precision medicine. Machine learning (ML) technology can integrate multidimensional data—including demographic characteristics, laboratory indicators, surgical information, and dynamic physiological waveforms—to construct high-performance predictive models, offering a new approach to transfusion prediction. This article systematically reviews the application progress of ML in transfusion prediction across clinical scenarios such as trauma, the perioperative period, and obstetrics. In trauma, ML has been applied to early massive transfusion prediction, prehospital transfusion decision-making, and precise prediction in pediatric trauma. In the perioperative field, applications span specialties including traumatic brain injury, orthopedic surgery, cardiovascular and major vascular surgery, and hepatic surgery. In obstetrics, models can effectively predict postpartum hemorrhage and transfusion requirements associated with cesarean section. Studies have shown that algorithms such as random forest, gradient boosting machines, and deep neural networks achieve area under the receiver operating characteristic curve (AUC) values of 0.83-0.98 in massive transfusion prediction, 0.75-0.97 in perioperative scenarios, and 0.80-0.89 in obstetric transfusion prediction—significantly outperforming traditional scoring tools. However, current studies are generally limited by single-center retrospective designs, insufficient external validation, and inadequate reporting standards; limited model interpretability and challenges in clinical integration further constrain practical translation. Future research should focus on constructing multicenter data cohorts, integrating multimodal data, conducting prospective implementation studies, and deeply embedding models within electronic health record systems, thereby advancing the application of ML to optimize clinical transfusion decision-making.
