7. Artificial Intelligence and Machine Learning in PBM

The next stage of PBM development is expected to benefit greatly from advances in the field of artificial intelligence (AI) and machine learning (ML).

AI and ML provide advanced capabilities to assimilate complex clinical data, enable personalized patient care, and support clinician decision making to improve outcomes. Over the past 15 years, there has been a marked increase in research focused on applying AI and ML across various domains of PBM. A 2022 scoping review identified 47 studies utilizing AI and ML in PBM, with most focusing on predicting transfusion requirements (30%) and bleeding risk (28%) (1).

This chapter presents a comprehensive overview of AI/ML applications across the core pillars of PBM: anemia management, minimization of blood loss, optimization of physiological tolerance to anemia, and decision making related to transfusion. It also highlights recent technological advancements, emerging clinical applications, and global perspectives, while also addressing existing challenges and outlining future directions for growth in the field.

Applications of AI and ML in PBM

1. Anemia Diagnosis and Management

Early identification and treatment of anemia is a cornerstone of PBM. Recent advances in AI and ML have shown promise in enhancing both diagnosis and therapeutic management of anemia.
ML algorithms have been developed to detect anemia using clinical photographs, particularly of the conjunctiva and gingiva, captured via smartphone cameras. These tools aim to identify pallor associated with anemia and offer a non-invasive, low-cost method of screening, particularly suited to resource-limited settings where conventional blood testing may be inaccessible or expensive (2). 
Differentiating iron deficiency anemia (IDA) from thalassemia trait is of particular clinical importance, especially in endemic regions. Traditionally, this requires specialized tests such as serum ferritin or hemoglobin electrophoresis. Recent studies have demonstrated that ML models trained on routine laboratory parameters, particularly complete blood counts (CBC), can achieve high accuracy in classifying anemia subtypes. For instance, a 2025 study from Thailand employed random forest and gradient boosting algorithms on a dataset of over 1,100 microcytic anemia patients. The models achieved an area under the curve (AUC) of approximately 0.95, effectively distinguishing IDA from thalassemia trait and outperforming traditional diagnostic indices (3).

AI based clinical decision support tools are also being used to guide anemia therapy, particularly the administration of erythropoiesis-stimulating agents (ESAs) in patients with chronic kidney disease. A recent randomized controlled trial by Lim et al evaluated AI based models for prescribing ESA in hemodialysis patients and found that the Random effects (REEM) trees model was non-inferior to physician guided dosing in maintaining target hemoglobin levels, without increasing adverse events. These findings support the potential role of AI in optimizing anemia management (4).

2. Predictive Analytics for Transfusion Decision Support

AI and ML are increasingly applied to support transfusion decision making through predictive analytics across different clinical settings. In trauma and critical care, where rapid identification of transfusion needs is critical, models such as XGBoost and random forest have demonstrated superior performance in forecasting transfusion requirements, compared to traditional scoring systems (5).

In surgical settings, AI algorithms have been applied to predict perioperative transfusion requirements using preoperative and intraoperative data, aiding in individualized blood ordering and resource allocation. Comparable accuracy has been observed between advanced ML models and well-designed logistic regression tools, emphasizing the need for context-specific implementation. In ICUs and general wards, predictive models that integrate hemoglobin trends with vital signs can improve anticipation of transfusion needs, potentially reducing inappropriate transfusions. 

3. Intraoperative Blood Loss Prediction and Conservation

Minimizing intraoperative blood loss forms another important pillar of PBM. Traditional bleeding surveillance techniques such as INR monitoring or point of care viscoelastic assays (TEG/ROTEM) offer only intermittent insights into a dynamic process. AI and ML tools are now being explored to provide real-time, data-driven forecasts of bleeding and guide hemostatic interventions more proactively.
ML models integrate diverse intraoperative data streams like vital signs, lab parameters and surgical details to predict significant blood loss before it becomes clinically apparent. Gigengack et al. (2023) developed a random forest algorithm that predicted blood loss > 250 mL in burn surgeries, enabling early implementation of cell salvage and antifibrinolytics (6). Similar models are under development for obstetric and orthopedic procedures, supporting forward planning and rapid intervention. In different surgical contexts, ML applications have demonstrated significant operational benefits. For example, in Thailand, ML based tool (naive Bayes) demonstrated acceptable accuracy for massive transfusions in neurosurgery patients (7).

AI is increasingly being explored to enhance the interpretation of viscoelastic assays. For instance, ML models have been developed to rapidly predict thromboelastography (TEG) curve outputs, such as clot firmness from easily measured blood protein concentrations, offering a practical tool for future real-time coagulation management in emergency settings (8). While still experimental, these systems could support real-time transfusion decisions during complex surgeries.

4. Blood Demand Forecasting

National and cross-border initiatives are increasingly integrating AI in clinical practises. In the United Kingdom, the National Health Service Blood and Transplant (NHSBT) piloted an AI as a service platform (“Kortical”) for platelet demand forecasting, reporting reductions in shortage and wastage (9). A modelling study carried out in Germany demonstrated that Deep Neural Networks can substantially improve platelet demand prediction accuracy, leading to reduced platelet wastage and fewer shortage events (10). Emerging trends include integration of donor databases, real-time usage feeds, and hospital inventory data into unified predictive platforms. Federated learning and privacy preserving, data sharing frameworks are being explored to enable robust models while complying with data protection regulations. Such approaches may guide both local operational decisions and national transfusion policies.

Methodological Advances in AI/ML for PBM

1. Natural Language Processing (NLP): PBM requires interpretation of structured laboratory values and vitals, free‑text clinical notes, immunohematology assays and time‑series data such as vital‑sign trends. Modern AI pipelines use data fusion techniques to harmonise these heterogeneous inputs. NLP can extract adverse transfusion reaction descriptions from unstructured clinical notes. A recent scoping review reported that NLP driven information retrieval improved detection of transfusion reactions compared with manual review, demonstrating how unstructured data can enrich PBM models (11).
2. Retrieval Augmented Generation (RAG):  RAG is an approach that combines LLMs with domain specific knowledge bases to generate context aware, evidence based outputs. In the setting of PBM, RAG can operationalize clinical guidelines and real-world data to provide precise, context based recommendations. For example, RAG systems could retrieve transfusion thresholds from consensus guidelines, align them with dynamic patient parameters (hemoglobin levels, comorbidities), and generate actionable transfusion triggers at the bedside. The recent Almanac framework (Retrieval-Augmented Language Models for Clinical Medicine) demonstrated that coupling LLMs with curated biomedical sources substantially improved factual accuracy and adherence to clinical guidelines (12). Applying similar architectures to PBM can enhance both the safety and efficiency of transfusion practice, making AI systems more trustworthy for clinical integration.
3. Deep learning for Morphology and Immunohematology:  Deep learning is emerging as a powerful tool to support PBM, offering applications that span red cell morphology, anemia monitoring, and immunohematology. Convolutional neural networks have been used to classify red blood cell shapes with a very high accuracy (13). RBC Match, a semi-supervised model applied to peripheral blood smears, has shown promise in monitoring anemia recovery, allowing clinicians to track morphological changes over time and align transfusion timing more closely with patient-specific needs (14). Another study demonstrated that both ABO and Rhesus (Rh) blood groups can be rapidly and accurately phenotyped by integrating deep learning algorithms for image interpretation, with high accuracy and reduced discrepancies in antigen typing (15).These advances will make PBM more precise, standardized, and patient centred, ultimately improving both safety and efficiency in transfusion practice.
4. Federated Learning:  To improve model generalisability without sharing patient specific data, federated learning architectures train models at separate centres and exchange only model weights. A federated deep‑learning framework for red‑cell abnormality detection achieved 94–95 % accuracy, similar to the accuracy of centrally trained models (16). Such approaches facilitate multi centre collaborations in PBM while maintaining patient privacy.
5. Random Forest and Other Predictive Models:  Random Forest (RF) is an approach that combines multiple decision trees, each trained on different subsets of the data, to enhance predictive accuracy. Evidence from studies in Australia (17), China (18) and Brazil has consistently shown that RF and other ML based algorithms outperform conventional risk scoring systems in predicting perioperative transfusion needs, especially in cardiac surgery cohorts. Classical time‑series models remain strong baselines for predicting blood demand. Accurate forecasting of blood demand is integral for implementing a successful PBM program. A study compared Auto Regressive Integrated Moving Average (ARIMA), Artificial Neural Networks (ANN), and hybrid models using multiyear hospital demand data in Iran and concluded ARIMA provided the more accurate predictions for most blood groups (19).
6. AI Agents: AI agents—or instances of AI models capable of autonomously receiving data and executing actions on behalf of clinicians or patients within defined contexts—can translate and execute AI predictions into real-world recommendations. Agents are increasingly recognized as having a critical role in complex medical environments such as perioperative care settings. Longstanding challenges—such as inefficient scheduling, fragmented communication, unpredictable case durations, administrative burdens, resource mismanagement, and postoperative complications—continue to undermine both efficiency and patient outcomes. By harnessing predictive analytics, automation, and real-time data, AI agents can bolster streamline clinical workflows across the surgical continuum: preoperatively, they can refine risk stratification, scheduling, and patient communication; intraoperatively, they can support real-time decision-making, and transfusion resource coordination; postoperatively, agents enable continuous monitoring and tailored care orders. Multiple agents specialized for individual tasks can also work together to orchestrate clinical care with the physician as the ‘human-in-the-loop’ for safety and critical decisions. Given the complexity of surgical care, perioperative care has many opportunities where specialized AI agents have the potential to deliver rapid, high-quality, evidence-based care coordination. 

Challenges and limitations

Despite rapid progress, several issues need be resolved and addressed before AI can be integrated into routine PBM. Data quality and standardisation are major concerns: hospital datasets often contain missing information or use different definitions for the same variables, which can reduce the accuracy of AI predictions. Another limitation is generalisation. Models trained in a single hospital or setting may not be equally efficient in other settings, since transfusion thresholds, patient characteristics, and local clinical practices vary. Federated learning, a method that allows hospitals to share knowledge without directly sharing patient data, is being explored to overcome this challenge.

Workflow integration is another major obstacle in implementing AI for routine applications. AI tools must deliver timely, role-appropriate information within existing hospital systems, otherwise excessive or poorly designed alerts may cause “alert fatigue,” where users start ignoring them.

Regulatory and ethical issues add further complexity. High quality, prospective, multi-centre trials are still  needed to demonstrate clinical effectiveness, patient benefit, and cost efficiency of the AI and ML models before they can be integrated in clinical Practice of PBM. In this context, the SPIRIT- AI (Standard Protocol Items Recommendations for Interventional Trials) and CONSORT-AI (Consolidated Standards of Reporting Trials) guidelines provide essential frameworks for the design and reporting of future studies (20).

Future directions

AI and ML in PBM is likely to evolve from passive prediction towards actionable decision support. Future systems should deliver contextualised recommendations, such as personalised anaemia management or transfusion triggers, directly within electronic health records.
Precision PBM, guided by patient-specific profiles, could enable tailored blood conservation strategies, while expansion into outpatient and preoperative settings may support earlier detection of anaemia and coagulopathy. 

International federated learning initiatives will be critical for building generalizable models without compromising data privacy, and continuous learning systems that recalibrate with new data and clinician feedback will help sustain accuracy over time. The rise of agentic AI, Model Context Protocol, and ambient AI holds promise for embedding real-time outputs into clinical workflows, thereby enhancing PBM decision making at the point of care. To ensure safe and equitable adoption, future platforms must address bias, promote transparency, and must be coupled with clinician training and organisational readiness as regulatory frameworks mature.
 

References

1. Meier, J. M., & Tschoellitsch, T. ( Artificial Intelligence and Machine Learning in Patient Blood Management: A Scoping Review. Anesthesia and analgesia, 2022, 135(3), 524–531.
2. Chatterjee, S., Malaiappan, S., & Yadalam, P. K. . Artificial intelligence (AI)-based detection of anemia using the clinical appearance of the gingiva. Cureus, 2024, 16(6), e62792.
3. Tepakhan, W., Srisintorn, W., Penglong, T., & Saelue, P. Machine learning approach for differentiating iron deficiency anemia and thalassemia using random forest and gradient boosting algorithms. Scientific Reports, 2025,15(1), Article 16917.
4. Lim, L. M., Lin, M. Y., Hsu, C., Ku, C., Chen, Y. P et al. Computer-assisted prescription of erythropoiesis-stimulating agents in patients undergoing maintenance hemodialysis: a randomized control trial for artificial intelligence model selection. 2025, JAMIA open
5. Feng, Y. N., Xu, Z. H., Liu, J. T., Sun, X. L., Wang, D. Q., & Yu, Y. Intelligent prediction of RBC demand in trauma patients using decision tree methods. Military Medical Research, 2021
6. Gigengack, R. K., Taha, D., Martijn Kuijper, T., Roukema, G. R., Dokter, J., Koopman, S. S. H. A., & Van der Vlies, C. H. Predicting blood loss in burn excisional surgery. Burns : journal of the International Society for Burn Injuries, 2023
7. Taweesomboonyat, C., Kaewborisutsakul, A., & Sungkaro, K. Prediction of massive transfusions in neurosurgical operations using machine learning. Asian Journal of Transfusion Science. 2022
8. Ghetmiri, D.E., Venturi, A.J., Cohen, M.J. et al. Quick model-based viscoelastic clot strength predictions from blood protein concentrations for cybermedical coagulation control. Nat Commun. 2024
9. Kortical Ltd. (2021). AI for platelet forecasting in UK blood services. Retrieved from https://kortical.com
10. Schilling, M., Rickmann, L., Hutschenreuter, G., & Spreckelsen, C. Reduction of Platelet Outdating and Shortage by Forecasting Demand With Statistical Learning and Deep Neural Networks: Modeling Study. JMIR medical informatics, 2022. 
11. Maynard, S., Farrington, J., Alimam, S., Evans, H., Li, K., Wong, W. K., & Stanworth, S. J. Machine learning in transfusion medicine: A scoping review. Transfusion, 2024.
12. Zakka, C., Shad, R., Chaurasia, A., Dalal, A. R., Kim, J. L. et al. Almanac - Retrieval-Augmented Language Models for Clinical Medicine. NEJM AI, 2024.
13. Routt, A. H., Yang, N., Piety, N. Z., Lu, M., & Shevkoplyas, S. S. Deep ensemble learning enables highly accurate classification of stored red blood cell morphology. Scientific reports,2023.
14. Yan, Q., Zhang, Y., Wei, L., Liu, X., & Wang, X. Assessment of anemia recovery using peripheral blood smears by deep semi-supervised learning. Annals of hematology, 2025.
15. Larpant, N., et al. Simultaneous phenotyping of five Rh red blood cell antigens on a paper-based analytical device combined with deep learning for rapid and accurate interpretation. Analytica Chimica Acta, 2022.
16. Dipto, S. M., Reza, M. T., Mim, N. T., Ksibi, A., Alsenan, S., Uddin, J., & Samad, M. A. An analysis of decipherable red blood cell abnormality detection under federated environment leveraging XAI incorporated deep learning. Scientific reports, 2024.
17. Hui V, Litton E, Edibam C, Geldenhuys A, Hahn R, Larbalestier R, et al. Using machine learning to predict bleeding after cardiac surgery. European Journal of Cardio-thoracic Surgery. 2023
18. Lee, S. M., Lee, G., Kim, T. K., Le, T., Hao,  et al. Development and Validation of a Prediction Model for Need for Massive Transfusion During Surgery Using Intraoperative Hemodynamic Monitoring Data. JAMA network open, 2022.
19. Sarvestani, S. E., Hatam, N., Seif, M., Kasraian, L., Lari, F. S., & Bayati, M. Forecasting blood demand for different blood groups in Shiraz using auto regressive integrated moving average (ARIMA) and artificial neural network (ANN) and a hybrid approaches. Scientific reports, 2022.
20. Ibrahim, H., Liu, X., Rivera, S. C., Moher, D., Chan, A. W., Sydes, M. R., Calvert, M. J., & Denniston, A. K. Reporting guidelines for clinical trials of artificial intelligence interventions: the SPIRIT-AI and CONSORT-AI guidelines. Trials, 2021.

The authors

The first version was written by Mark Yazer, the updated version was created by Rounak Dubey, Sheharyar Raza and Ruchika Goel.

 Click the below link to go back to the Patient Blood Management Resources Homepage.