In October 2026, a study published in *Transfusion Medicine* reported seemingly respectable numbers: a one-dimensional convolutional neural network achieved an overall accuracy of 0.7314 in predicting red blood cell (RBC) transfusion volume categories for hospitalized patients on a hold-out test set. Yet the authors themselves explicitly stated in the abstract that the model ‘missed high-volume transfusion events’, labeling this as a ‘critical safety limitation’: precisely those patients receiving the largest transfusion volumes were missed. This is not an isolated case. From September to November 2026, AI research in transfusion medicine appeared consecutively: prediction of donor reactions, blood-type inventory forecasting, trauma-related transfusion volume prediction, and a digital pilot for donor eligibility screening—alongside three review articles, one public letter exchange, and a thalassemia-focused series paper. Reading these ten publications side by side reveals a consistent weakness: the rarer the event, the poorer the model’s performance; and the safety-critical thresholds in transfusion medicine lie precisely at the rare end (this article’s summary).
## 0.7314: Propped Up by the Most Common Class
A team from the Department of Transfusion Medicine and the Medical Digital Intelligence Innovation Center at Beijing Friendship Hospital, Capital Medical University (seven authors) retrospectively collected data from 7,250 patients between April 2024 and June 2025. They trained a one-dimensional convolutional neural network using 19 standardized pre-transfusion hematologic and clinical variables to predict RBC transfusion volume categorized into five levels: 1, 2, 3, 4, or ≥5 units. Evaluation was performed on a 30% hold-out test set. The overall accuracy was 0.7314, and the authors claimed the model converged without overfitting (authors’ statement).
But when broken down, problems emerge: the 2-unit category accounted for 68.03% of the cohort, and its precision was exactly 0.7314—the same value as the overall accuracy (the identical figures are quoted verbatim from the abstract; the fact that overall accuracy matches the precision of a single class is itself an anomaly noted in the abstract, which this article reports faithfully without smoothing). In contrast, minority classes (≥5 units and 3 units) had low prevalence (both <5%) and overlapping hemoglobin values (60–80 g/L), limiting model performance. The authors termed this a ‘critical safety limitation’.
The clinical implication is straightforward (this article’s judgment): the ≥5-unit group comprises patients with the most severe bleeding—those for whom blood inventory preparation absolutely cannot fail. Systematic underestimation for this group means blood banks stock based on average volumes, leading to shortages during the most critical surgeries and trauma cases. Errors at the tail carry far greater consequences than errors in the middle. The authors listed three improvement directions: external multicenter validation, targeted data augmentation for minority classes, and development of indication-specific submodels. None of these has yet been implemented in this study.
## 0.587%: Under Rare Events, AUC Looks Good
Donor-side rare events exist too. A team from Zhejiang Provincial Blood Center (seven authors) analyzed 5,310,116 whole-blood donations recorded at the center between January 2018 and June 2026 to predict vasovagal reactions (VVR)—a common systemic adverse event during donation, occurring at only 0.587% in the final test cohort. This study’s design was the most rigorous among the ten publications (this article’s judgment): data from 2018–2023 were used for development; 2024 was reserved for internal temporal validation; and January 2025–June 2026 was held out for final temporal testing. Six candidate models were compared using 20 pre-donation available predictors, prioritizing average precision (AP) as the primary metric and ROC-AUC as secondary. The model, calibration function, and risk threshold were frozen once during development and remained unchanged across all subsequent cohorts.
The results yield two sets of numbers requiring separate interpretation. Under the AP-prioritized framework, XGBoost was selected. In the 2024 validation cohort, AP = 0.0336 and ROC-AUC = 0.7776. In the final test cohort—1,014,652 donations with 5,957 VVRs—AP = 0.0338 (95% CI: 0.0305–0.0374); ROC-AUC = 0.7633 (95% CI: 0.7572–0.7688), indicating declining discriminative ability over time. After applying the frozen calibration function, the observed-to-expected ratio was 1.169—i.e., actual reactions occurred ~17% more frequently than predicted (the ratio is taken verbatim from the abstract; the percentage is recalculated herein).
At such a rare baseline rate (0.587%), an AUC of 0.7633 sounds usable—but AP = 0.0338 is the honest number (this article’s judgment): it indicates that only a tiny fraction of individuals flagged by the model actually experienced VVR. The frozen alert threshold marked 1.55% of donors, among whom the VVR incidence was 3.45%—5.88 times the overall rate. The authors concluded the tool supports risk stratification and targeted observation at donation sites, while clearly stating prerequisites for routine use: multicenter external validation and prospective implementation studies. One additional figure merits noting: for donors with prior observable donation history, AUC = 0.7713; for those without prior history, AUC = 0.6822. The less data available per individual, the weaker the model’s discrimination (this article’s judgment).
## The Hardest-to-Predict Blood Type Is the Rarest One
Now shifting to the blood bank side. A Czech team (Institute of Scientific Instruments of the Czech Academy of Sciences and Department of Transfusion and Tissue Medicine, University Hospital Brno; four authors) built an interpretable pipeline for a centralized Czech transfusion center: linear regression with automatic forward selection and backward elimination to forecast daily whole-blood inventory by ABO/Rh blood type. Daily data from 2021–2024 were used for training; 2025 was reserved for independent testing. Features included school holidays, proximity to non-working days, weekday and month indicator variables, national influenza surveillance metrics, and scheduled donation numbers from two hospitals.
Results were reported separately by blood type: normalized mean absolute error (nMAE) was 6.49% for A+, but 26.26% for AB−. The largest error fell on AB−—typically the least prevalent blood type in the population (prevalence ranking is common background knowledge, unmentioned in the abstract; this article’s judgment), representing a roughly 4.05-fold difference (recalculated herein). The model consistently outperformed a naive baseline using only scheduled donations; the authors claimed performance comparable to other machine learning methods (abstract assertion, no comparative details provided). The dominant predictor was scheduled donation numbers, with a coefficient near 1; calendar variables added incremental improvement; influenza indicators showed weak but statistically significant associations.
Two takeaways (this article’s judgment): First, safety margins for inventory must be set separately per blood type; preparing AB− inventory using the error magnitude of A+ is effectively no margin at all. Second, for tasks with clear structural signals—e.g., ‘more scheduled donations → higher inventory’—interpretable linear models need not lose to complex models; the authors explicitly emphasized interpretability in both methods and conclusions.
## The Most Accurately Predicted Outcome Is ‘No Transfusion Needed’
Pre-hospital trauma represents the heaviest component of transfusion volume prediction (this article’s judgment). A U.S. multi-institutional team (Department of Trauma, University of Pennsylvania; MIT Lincoln Laboratory; eleven authors) used the 2020 American College of Surgeons Trauma Quality Improvement Program (TQIP) registry to include 401,337 blunt, penetrating, and burn trauma patients aged 18–60: 314,790 blunt, 71,760 penetrating, and 5,726 burn; median age 37 years (IQR 27–50); median Injury Severity Score (ISS) 6, versus median ISS 26 among those receiving massive transfusion (>4 units within 4 hours). Their FAIT tool uses EMS and emergency department vital signs plus easily observable field injury patterns as inputs, modeling transfusion volume as a continuous variable via probabilistic distribution prediction covering RBCs, plasma, platelets, cryoprecipitate, and whole blood; methodology employed ten-fold cross-validated logistic regression with class weighting.
Results show a clear directional pattern: the AUC for predicting ‘no transfusion needed’ was 0.91 ± 0.00—the only AUC reported in the abstract with zero fluctuation notation (this article’s summary per abstract wording). For transfusion volume thresholds (>0, >4, >10, >20 units within 4 hours), AUC ranged from 0.91–0.95 using both EMS and ED data, and 0.87–0.91 using EMS vital signs alone. The authors also reported survival rates by transfusion volume: 71% for >5 units, 62% for >10 units, and 35% for >50 units. The heavier the tail, the lower the survival—and precisely there lies the greatest need for accurate prediction (this article’s summary).
Two points must be reported faithfully. First, the sum of the three injury types equals 392,276, differing from the abstract’s stated total cohort size of 401,337 by 9,061: this discrepancy originates entirely within the abstract itself; this article records both numbers without inferring causes (summation and discrepancy are recalculated herein). Second, ten-fold cross-validation constitutes internal validation; retrospective data from a single registry cannot demonstrate model performance at another hospital or in another year (this article’s judgment). Additionally, the phrase ‘AUC 0.94–0.95 ± 0.01’ appearing in parentheses at the end of the survival-rate sentence is not defined in the abstract regarding its referent; this article does not infer its meaning.
## 12 Questionnaires
Another gate exists before blood collection: donor eligibility screening. A team from the Department of Hematology, King Abdulaziz University, Jeddah, Saudi Arabia (two authors) converted the Saudi National Donor History Questionnaire (DHQ v1.0) and its 52-item eligibility flowchart into an AI-assisted, searchable digital reference and a digital questionnaire capable of generating eligibility determinations. A prospective before-and-after quality improvement pilot was conducted across two centers (April–August 2026), published on November 3 in the *Journal of Clinical Medicine*.
All study data came from staff questionnaires—12 pre-intervention and 12 post-intervention—with an 86% response rate. Post-implementation, 33% of participants reported screening a donor in <5 minutes, versus 0% pre-implementation; uncertainty related to medications dropped from 75.0% to 50.0%; self-ratings for usability, retrieval speed, accuracy, process improvement, and satisfaction were all favorable; however, overall process ratings showed no change.
Two distinctions must be clarified (this article’s judgment): First, the ‘AI-generated’ label in the title refers to AI-assisted conversion of a paper manual into a digital tool; eligibility determinations were generated by the digital questionnaire, but the abstract does not describe its decision logic, and this article makes no inference. Nor does the abstract describe any autonomous AI judgment of eligibility. Second, the 12-by-12 pre-post self-assessments measured staff perception—not screening quality. The authors’ own framing is explicit: exploratory, self-reported outcomes requiring larger-scale studies with objective measurements for confirmation. One additional disclosure: the article’s PubMed XML contains a DOI field serialization error; this article verified and adopted the official Crossref-consistent DOI.
## 99% and 80%: At the Review Level
Numbers from primary studies come with caveats; review-level numbers suddenly swell. On October 12, *Transfusion* published a narrative review co-authored by the AABB Foundation (16 authors from institutions including Johns Hopkins University School of Medicine, Canadian Blood Services, Vitalant, OneBlood, and Carter BloodCare), mapping AI applications along the vein-to-vein continuum: personalized donor recruitment and individualized iron management at the donor end; fairness auditing; demand forecasting, waste minimization, and inventory optimization in the supply chain; federated learning enabling multi-institutional collaboration without sharing raw data; omics and genomics supporting high-throughput analysis, phenotype prediction, and precision transfusion support for rare blood types and alloimmunized patients; enhanced hemovigilance in hospitals, with large language models and retrieval-augmented generation supporting decision-making, patient blood management, and adaptive learning. The authors self-limited their scope: this is a narrative, non-exhaustive review intended as a ‘framework for ongoing monitoring’, not a fixed checklist. Realizing its potential requires governance, cross-population validation, privacy protection, bias mitigation, reproducibility, regulatory oversight, and human-in-the-loop involvement.
On September 1, *Immunohematology* published an expert review (three authors from AIIMS Nagpur Department of Transfusion Medicine and Hematology, India) offering different numbers: image- and deep-learning-based methods for automated red blood cell antigen phenotyping and genotyping are said to achieve ‘99% accuracy’; the Haem-Match Alliance’s bloodMatcher platform is said to ‘potentially reduce alloimmunization risk by up to 80%’. Both statements are review-level paraphrases, qualified by ‘said’, ‘potentially’, and ‘up to’, with no elaboration of primary validation designs in the abstract; this article reports them verbatim per the abstract’s phrasing, without verifying original studies. Placing these two numbers alongside each other (this article’s summary): primary studies report AUC = 0.7633, accuracy = 0.7314, AP = 0.0338, nMAE = 26.26%; review-level claims cite 99% and 80%. Neither set originates from the primary studies reviewed herein (original studies cited by the reviews are not specified in their abstracts), nor is the validation tier visible.
Also in *Transfusion*, a September review by five authors on biologics surveyed 12 AI applications across three thematic domains and presented a maturity spectrum: from early clinical adoption (HLA matching, production quality control) to conceptual stage (digital twins, individualized pre-treatment). On September 14, *Vox Sanguinis* published a reply by two authors to comments received on their review titled ‘Harnessing Big Data and Artificial Intelligence in Transfusion Medicine’ (Chinese translation of title is this article’s rendering). The reply lacks an abstract and is unavailable in full; this article cites only title-level information and refrains from inferring comment positions. However, the very existence of a public comment-and-reply exchange signals that the review’s assertions provoked open scholarly debate within the field (debate being the inherent purpose of the ‘comment’ literature type). A global perspective is completed by the thalassemia series in *The Lancet Haematology* (November; 15 authors): AI’s roles in screening, diagnosis, complication prediction, and remote care are reviewed, while a stronger caution is sounded—that most thalassemia patients reside in low- and middle-income countries, where access to specialist care and novel therapies remains severely limited; in rare-disease contexts, AI’s bottleneck is often not the algorithm, but whether people can access care (this article’s judgment).
## Three Judgments
Reading the ten publications side by side, this article distills three judgments (all synthesized herein—not conclusions drawn from any single paper).
First, error resides where rarity resides—and so does safety. Five primary studies reported four rare and one thin: rare adverse events (VVR incidence 0.587%), rare blood types (AB−), rare transfusion categories (≥5 units, prevalence <5%), rare critical illness (massive transfusion, >50 units associated with 35% survival), and the thinnest validation material (12 questionnaires). Model performance is weakest precisely at these points. Overall metrics systematically obscure this: AUC = 0.7633 and overall accuracy = 0.7314 appear acceptable, but disaggregated metrics—AP = 0.0338, per-class precision, blood-type-specific nMAE—expose the rarity-side weaknesses.
Second, primary studies are more honest than reviews. Authors of the five primary studies openly disclosed limitations in their abstracts (‘critical safety limitation’, ‘missed high-volume transfusion events’, poorer performance in donors without prior history, largest error for AB−, need for objective confirmation of self-reported data) and explicitly listed external validation, prospective studies, and objective measurement as prerequisites. Review-level optimistic numbers are mostly qualified paraphrases, with primary validation designs invisible at the abstract level. The public comment-and-reply exchange confirms the field itself is rigorously scrutinizing the boundaries of such claims.
Third, the validation tier is the common prerequisite for all numbers. Temporal validation (Zhejiang), independent-year testing (Czech Republic), single-center hold-out testing (Beijing), internal registry ten-fold cross-validation (FAIT), and 12-person self-reported pre-post comparison (Saudi Arabia)—five tiers occupy distinct positions, yet none answers the question ‘What changed after deployment?’: Did donor injuries decrease? Did blood wastage decline? Did transfusion decisions shift? None of the ten publications provides an answer (this article’s judgment). When reading any transfusion AI number, the first question should be: At which validation tier was it validated?
## What Different Readers Can Take Away
Blood center and blood bank managers: The Zhejiang study provides a directly applicable template—freezing models, calibrating functions and thresholds, performing temporal validation using future cohorts, reporting Average Precision (AP) and observed-to-expected ratios rather than AUC alone; a 1.55% labeling rate yields 5.88-fold enrichment, and the operational model of targeted observation for labeled populations is directly evaluable. However, the authors explicitly state prerequisites: multicenter external validation and prospective implementation studies.
Hospital transfusion departments and clinicians: An overall accuracy of 0.7314 must not be misread as deployable—it is largely sustained by two major unit classes, while units ≥5 are systematically missed; none of the three prerequisites listed by the authors (external validation, data augmentation, indication-specific submodels) has been implemented; FAIT-style prehospital prediction holds potential for triggering massive transfusion protocols, but currently rests only on internal validation from a single registry.
Blood supply chain managers: Scheduled donor attendance is already the dominant signal; calendar and influenza variables offer only modest improvement; for rare blood types such as AB−, safety margins must be prepared at a daily error magnitude of 26.26%; interpretable linear regression suffices for this task—complex models need not be default choices.
Model developers: For rare-event prediction, AP and calibration drift must be reported; minority classes must be reported separately—not buried within overall accuracy; temporal validation must be standard practice, not an optional bonus.
Donors and patients: These ten documents do not alter any individual’s eligibility determination or transfusion plan; if risk stratification ever becomes operational, its form will more likely involve staff conducting heightened observation of high-risk donors (targeted observation being the authors’ conclusion), yet this remains pending prospective validation (as summarized herein).
Final boundary statement: This article reviews summary-level materials from ten sources; all numbers are transcribed verbatim from respective abstracts; dates reflect PubMed-recorded publication dates; comment replies contain no abstracts and full texts are unavailable—only title-level information is cited, with no content paraphrasing; the eligibility determination logic in the Saudi study is not described in its abstract, and this article makes no inference about it; the DOI field in its XML contains a serial error, which has been disclosed; this article adopts the official DOI value verified against Crossref; the figures “99%” and “80%” are summary-level paraphrases with qualifiers—their original primary sources fall outside the scope of verification herein; the AUC parenthetical in the FAIT survival-rate sentence is ambiguous, and this article makes no inference; the discrepancy of 9,061 between itemized sums and totals reflects inconsistency within the abstract itself—this is reproduced verbatim without inferring causes; the claim that performance is “comparable to other machine learning methods” is the authors’ assertion, with no comparative details provided; the statement that the model is “converged and not overfitted” is the authors’ own description; conflict-of-interest and funding information are reported solely as stated in each abstract—this article makes no assertion of “no conflict of interest.”
Sources: Transfusion Clinique et Biologique, October 29, 2026 (PMID 42957083, https://doi.org/10.1016/j.tracli.2026.10.006); Journal of Clinical Medicine, November 3, 2026 (PMID 43030691); Vox Sanguinis, October 29, 2026 (PMID 42957183); Transfusion Medicine, October 7, 2026 (PMID 42839773); Injury, September 14, 2026 (PMID 42760195); Transfusion, October 12, 2026 (PMID 42864904); Immunohematology, September 1, 2026 (PMID 42859974); Transfusion, September 2026 (PMID 42798768); Vox Sanguinis, September 14, 2026 (PMID 42736258); The Lancet Haematology, November 2026 (PMID 42838072)