Artificial Intelligence-based predictive models for adverse blood donor reactions: a systematic review of immediate...

A significant challenge in blood donation is the occurrence of adverse donor reactions (ADRs) and their subsequent negative impact on the blood supply and public health. A promising strategy to mitigate these events is the deployment of non-invasive, cost-effective artificial intelligence (AI) models for donor screening and monitoring. This study aims to systematically review the AI models utilized in the identification and prediction of ADRs.

This study used a systematic review approach in line with the PRISMA 2020 guidelines. A search was performed across various databases, including Web of Science, PubMed, Embase, Google Scholar, and Scopus. The results were combined narratively and presented using descriptive statistics. The quality of eligible studies was assessed using the Newcastle-Ottawa Scale (NOS) tool.

Among the 13 studies, nine were classified as immediate reactions and four as delayed reactions. The commonly used models included regression models, classical statistical models, and machine learning algorithms such as Random Forests, Gradient Boosting Machines (GBMs), XGBoost, and Artificial Neural Networks (ANNs). The main standard evaluation metrics for the models included Odds Ratio, Precision-Recall Area Under the Curve (PR-AUC), F1 Score, Precision, and Recall.

Adverse reactions among blood donors negatively impact donor retention and, by extension, the stability of the blood supply for patient care. In this context, AI models may offer a promising tool for supporting the prediction and monitoring of ADRs. However, the available studies are not sufficient to support the widespread adoption of these models in clinical or operational decision-making. Heterogeneity in study design, outcomes, and evaluation metrics together with limitations such as limited implementation, risk of bias, unclear reference standards, and a lack of external validation has constrained the interpretability and generalizability of the findings. Therefore, future research with more rigorous study designs, standardized reporting, harmonized evaluations, and external validation is essential to establish the effectiveness and reliability of these models.