Governing AI-Assisted Red Blood Cell Transfusion Decisions in Viet Nam: Clinical Indication Review, Clinician Approval and Hemovigilance

Governing AI-Assisted Red Blood Cell Transfusion Decisions in Viet Nam: Clinical Indication Review, Clinician Approval and Hemovigilance

A physician is reviewing a request for two units of red blood cells for a patient recovering from major abdominal surgery. The latest hemoglobin concentration is 7.6 g/dL. The patient is hemodynamically stable, with no evidence of ongoing bleeding. An artificial intelligence (AI) system identifies the order as potentially inconsistent with a restrictive transfusion strategy and prompts the physician to review the indication.

Elsewhere in the hospital, another patient has a similar hemoglobin concentration but has recently experienced an acute myocardial infarction. Applying the same transfusion recommendation to both patients could have very different clinical consequences.

The difference is not simply the laboratory result. It is the patient's underlying condition, symptoms, physiological status, treatment goals and the evidence relevant to that particular situation.

Even when transfusion is appropriate, the prescribing decision is only one part of safe care. The correct blood component must be selected, compatibility established, patient and blood-unit identities verified, and the transfusion administered and monitored according to established procedures.

AI may help healthcare organizations identify orders requiring review, bring relevant clinical information together and support more consistent application of transfusion guidance. However, an AI-generated recommendation cannot independently establish whether a patient should receive blood, nor can it replace the professional and laboratory safeguards that protect patients throughout the transfusion process.

For healthcare organizations in Viet Nam, the challenge is to integrate AI into transfusion decision-making without delaying necessary treatment, encouraging inappropriate blood use or weakening existing safety controls.

The objective is not simply to transfuse fewer units. It is to ensure that patients receive blood when clinically needed, avoid unnecessary exposure when transfusion is unlikely to benefit them, and remain protected by reliable clinical assessment, safe administration and organizational learning.

Red Blood Cell Transfusion as a Clinical Governance Responsibility

Red blood cell transfusion is an essential treatment for selected patients with clinically significant anemia or blood loss. By increasing oxygen-carrying capacity, it can be lifesaving in appropriate circumstances.

Transfusion also carries potential harms. Patients may experience acute or delayed hemolytic reactions, allergic or febrile nonhemolytic reactions, transfusion-associated circulatory overload, transfusion-related acute lung injury and other complications. Although blood screening and safety measures have substantially reduced infectious risks, transfusion-transmitted infections remain a consideration within the broader blood safety system.

Avoiding unnecessary transfusion can reduce avoidable exposure. However, withholding or delaying necessary transfusion may also cause serious harm, particularly in patients with major hemorrhage, impaired tissue oxygen delivery or certain cardiovascular conditions. Safe practice must balance these competing risks.

The World Health Organization (WHO) describes Patient Blood Management (PBM) as a patient-centered, systematic and evidence-based approach to improving patient outcomes by managing and preserving the patient's own blood.

PBM extends beyond deciding whether someone should receive red blood cells. It includes identifying and managing anemia, reducing avoidable blood loss and optimizing physiological tolerance of anemia.

WHO's 2025 publication, Guidance on Implementing Patient Blood Management to Improve Global Blood Health Status, provides a framework for incorporating these principles into healthcare organizations and health systems.

This broader perspective matters because transfusion requirements are influenced by care delivered before a transfusion order is considered. Identifying and treating iron deficiency before elective surgery, where appropriate and feasible, may reduce subsequent transfusion requirements. Effective prevention and management of bleeding can also reduce exposure to donor blood.

An AI system designed to review transfusion orders addresses only one part of this larger process. Its contribution should be evaluated accordingly.

A reduction in blood utilization may indicate improved clinical practice, but it could also reflect changes in patient characteristics, surgical activity or delayed treatment. The more meaningful objective is appropriate blood use, supported by reliable clinical decisions and safe administration.

Why Hemoglobin Alone Cannot Determine the Transfusion Decision

Hemoglobin concentration is an important component of transfusion assessment, but it does not fully describe a patient's clinical need.

A person with longstanding anemia may tolerate a hemoglobin concentration that would be more concerning in another patient experiencing acute blood loss, myocardial ischemia or physiological instability. Symptoms, cardiovascular status, the rate of hemoglobin decline, bleeding, anticipated procedures and available treatment alternatives all influence the decision.

The 2023 AABB International Guidelines provide substantial evidence supporting restrictive red blood cell transfusion strategies in many hospitalized patients.

For most hemodynamically stable adults, the guidelines recommend considering transfusion when hemoglobin is below 7 g/dL. In accordance with thresholds used in relevant clinical trials, clinicians may select 7.5 g/dL for patients undergoing cardiac surgery and 8 g/dL for those undergoing orthopedic surgery or with preexisting cardiovascular disease.

For hospitalized adults with hematologic or oncologic disorders who are hemodynamically stable, the guidelines suggest considering a restrictive threshold of 7 g/dL, although this recommendation is conditional and supported by low-certainty evidence.

These recommendations provide an evidence-based foundation for clinical practice. They do not establish a mandatory threshold for every patient or replace individualized clinical assessment.

An important development occurred in August 2025, when AABB published international clinical practice guidelines specifically addressing red blood cell transfusion in patients with acute myocardial infarction.

For hospitalized patients with acute myocardial infarction and hemoglobin below 10 g/dL, the panel suggested a more liberal transfusion strategy. This was a conditional recommendation based on low-certainty evidence, reflecting uncertainty about the precise balance of benefit and harm.

The recommendation does not mean that every patient with acute myocardial infarction and hemoglobin below 10 g/dL must automatically receive red blood cells. Decisions should incorporate clinical circumstances, potential adverse effects, available alternatives and responsible use of blood resources.

The distinction between the 2023 general guideline and the 2025 acute myocardial infarction guideline has direct implications for AI-assisted decision support.

An algorithm that automatically treats every transfusion above 7 g/dL as inappropriate could generate misleading recommendations in patients for whom a higher threshold may be clinically justified. Conversely, routinely recommending transfusion at a higher threshold without confirming the relevant diagnosis and clinical circumstances could encourage unnecessary treatment.

Different international guidelines may also address overlapping patient populations using different recommendations.

For example, NICE guideline NG24 includes a red blood cell transfusion threshold of 8 g/dL for acute coronary syndrome, whereas the newer AABB guideline specifically addressing acute myocardial infarction suggests a more liberal strategy when hemoglobin is below 10 g/dL.

These recommendations should not be treated as interchangeable or automatically applied outside their intended populations. Their evidence base, scope, publication dates and the patient's clinical circumstances need to be considered when developing locally approved transfusion protocols.

General restrictive thresholds should also not be extrapolated without qualification to patients with active major hemorrhage, specialized chronic transfusion requirements or other situations requiring distinct management strategies.

For AI governance, the implication is that a system must recognize the limits of the clinical guidance it applies. The responsibility for interpreting that guidance and making the transfusion decision remains with appropriately qualified healthcare professionals.

Defining the Role of AI in Transfusion Decision Support

Not every electronic system that identifies a potentially inappropriate transfusion order uses artificial intelligence.

A computerized provider order entry system may apply a predetermined rule, such as displaying a warning when a transfusion is ordered above a specified hemoglobin threshold. Such a function may be valuable, but conventional rule-based decision support does not automatically constitute AI.

Machine-learning systems may perform different tasks. They can analyze multiple clinical variables to estimate perioperative transfusion requirements, predict substantial blood use or identify patterns associated with potentially inappropriate orders.

Large language models may help extract information from clinical notes, summarize previous transfusion history or identify documentation that appears inconsistent with a proposed order.

These functions should not be confused with one another.

A model that predicts how many units a patient may require during surgery has not necessarily demonstrated that it can determine whether a specific transfusion order is clinically appropriate.

A system that accurately extracts hemoglobin results from electronic records does not establish the patient's need for transfusion. Similarly, an algorithm designed to identify possible adverse reactions after transfusion performs a different function from one reviewing the clinical indication before transfusion begins.

Healthcare organizations should define these distinctions before procurement, clinical validation and implementation.

For transfusion indication review, a reasonable initial function may be identifying orders that warrant closer professional assessment.

For example, the system might recognize a request for multiple red blood cell units in a stable, nonbleeding adult or identify an order for which the documented clinical indication appears incomplete.

These findings may justify review, but they do not prove that the order is inappropriate. The treating clinician may possess relevant information that has not yet been entered into the electronic record, or the patient's condition may have changed since the latest laboratory result was obtained.

An AI-generated alert should therefore be regarded as a signal for verification rather than a definitive determination.

The hospital should be able to explain the system's intended purpose, the population in which it has been evaluated, the information it uses and the decisions that remain outside its validated function. These boundaries become especially important when different AI capabilities are integrated into the same electronic health record or transfusion management platform.

Reliable Clinical Information Is Essential

Transfusion decisions are particularly sensitive to the timing and interpretation of clinical information.

A hemoglobin sample may have been collected before recent bleeding, during substantial intravenous fluid administration or before a transfusion that has already been completed. The latest laboratory result visible in an electronic record may therefore no longer represent the patient's current physiological condition.

An AI system that uses the most recently reported hemoglobin value without considering when the sample was collected and what occurred afterward may produce an inappropriate recommendation.

Consider a patient who has already received a red blood cell unit following an earlier laboratory measurement. If post-transfusion assessment remains pending, the system may incorrectly recommend additional blood because it has not recognized the completed intervention.

The opposite problem can arise during rapidly evolving hemorrhage. A patient may be losing blood significantly before a new hemoglobin measurement reflects the severity of the situation.

In acute bleeding, hemoglobin concentration may not immediately indicate the magnitude of blood loss. Clinical assessment and established emergency hemorrhage protocols remain essential.

Other data quality problems may involve inconsistent laboratory units, missing diagnoses, incomplete medication information, duplicate records or errors in patient identification.

For AI-assisted review to be clinically useful, the hospital needs reliable connections among relevant clinical records, laboratory information, transfusion service records and treatment documentation.

The system should distinguish an order entered by a clinician from a blood component that has been reserved, issued, started, completed, canceled or returned unused. These events have different clinical and operational meanings.

Missing information also requires careful handling. The absence of documented bleeding does not necessarily establish that the patient is not bleeding. Incomplete clinical records should not automatically be interpreted as evidence supporting a restrictive transfusion recommendation.

Where important information is unavailable, prompting clinical verification may be safer than presenting a definitive conclusion.

An AI system that communicates uncertain findings with excessive confidence can create new safety risks, even when its overall predictive performance appears satisfactory.

Clinician Approval Must Remain a Meaningful Decision

AI-assisted review is useful only when it supports professional decision-making rather than reducing approval to routine acknowledgment of an electronic alert.

The treating clinician must consider the patient's condition, underlying cause of anemia, relevant laboratory findings, bleeding status, available alternatives and expected benefits and risks of transfusion. Patient preferences and applicable requirements for information, consent and documentation must also be considered.

A clinician may appropriately accept, modify or reject an AI-generated recommendation.

For example, an order may be flagged because the hemoglobin concentration is above the usual restrictive threshold. The physician may nevertheless determine that transfusion is justified because of documented cardiovascular or other patient-specific circumstances.

That decision should not automatically be classified as failure to follow AI.

Similarly, a clinician may reject an AI-generated suggestion to transfuse because the patient is stable, the anemia can be addressed through another treatment or the expected benefit is insufficient.

The hospital should support meaningful documentation of clinically important decisions without creating unnecessary administrative burdens.

Requiring repeated explanations for numerous low-value alerts may contribute to alert fatigue and superficial acknowledgment. Governance arrangements should distinguish clinically significant exceptions from ordinary situations in which further documentation adds little value.

Emergency transfusion decisions require particular safeguards.

A system designed to review routine orders in stable patients should not create barriers to emergency blood release or delay resuscitation during life-threatening hemorrhage.

In Viet Nam, Article 45 of Circular No. 26/2013/TT-BYT addresses immunohematological compatibility assurance in specified emergency circumstances when the usual testing process cannot be completed in time or compatible components cannot be selected through the normal pathway.

These exceptional circumstances require the relevant professional authorization, conditions and subsequent testing specified in the Circular. They do not provide a general exemption from transfusion safety requirements.

An AI system should not independently block an emergency transfusion because a hemoglobin measurement does not meet a routine threshold.

The purpose of AI oversight is to improve clinical decisions without interfering with necessary emergency treatment.

The Single-Unit Approach and Clinical Reassessment

One practical element of patient blood management is avoiding routine multiple-unit transfusions when a smaller initial intervention followed by reassessment would be appropriate.

For adults without active bleeding, NICE guideline NG24 recommends considering single-unit red blood cell transfusion, followed by clinical reassessment and a hemoglobin check before determining whether further transfusion is needed.

Patients with active major hemorrhage or other urgent circumstances require separate clinical pathways. Patients receiving regular transfusions for chronic anemia may also need individualized treatment plans and thresholds.

NICE NG24 was updated on 26 February 2026. That update primarily addressed the use of tranexamic acid to reduce transfusion requirements in surgical settings. It did not introduce a new universal red blood cell threshold or fundamentally change the established single-unit reassessment principle.

For hospitals considering AI, this provides a relatively well-defined opportunity for decision support.

A system could identify orders for multiple units in stable, nonbleeding patients and prompt clinicians to consider whether initial single-unit transfusion followed by reassessment would be appropriate.

It could also identify situations in which a first unit has been administered but the expected clinical or laboratory reassessment has not been documented before additional blood is requested.

However, the system should not treat every multiple-unit order as inappropriate.

Patients with major hemorrhage, rapidly evolving instability or other circumstances requiring urgent resuscitation may need a different strategy. The routine single-unit approach must not interfere with massive hemorrhage protocols or other clinically justified treatment.

The value of AI lies in helping clinicians recognize when reassessment may prevent unnecessary transfusion, not in enforcing a numerical rule independently of the patient's condition.

Safe Administration Requires More Than an Appropriate Order

Even when the clinical indication is appropriate, transfusion can cause harm if subsequent safety controls fail.

The transfusion process involves patient identification, specimen collection, blood group testing, antibody screening when indicated, compatibility assessment, component selection, blood issue, bedside verification, administration and monitoring.

Errors may occur at any stage.

A blood sample may be collected from the wrong patient or labeled incorrectly. A blood component may be issued for the wrong recipient. A discrepancy between the patient, blood unit and accompanying documentation may remain undetected.

The appropriateness of the original transfusion order does not eliminate these risks.

In Viet Nam, Circular No. 26/2013/TT-BYT provides detailed requirements governing blood transfusion activities.

Article 42 addresses requirements for treatment units before receiving blood components, including assessment of anticipated transfusion needs and relevant pretransfusion testing.

Article 50 establishes requirements for verification and monitoring at the treatment unit. These include checking patient and blood-unit information, performing bedside ABO grouping and verification as specified in the Circular, monitoring the patient's condition and promptly identifying and managing transfusion-related abnormalities.

The Circular also includes specific requirements for observing clinical signs during transfusion, with particular attention to the initial period after transfusion begins.

Other provisions address compatibility testing, selection of suitable blood components, emergency circumstances and the management of transfusion-related incidents.

These are established professional safeguards, not optional activities that may be replaced by AI-generated conclusions.

Electronic identification, barcode matching and automated compatibility controls may help prevent certain errors. However, these technologies are not necessarily AI systems and should be assessed according to their actual functions.

An algorithm reviewing transfusion indications and an electronic system preventing wrong-patient or wrong-component errors address different safety problems.

Both may contribute to safer care, but neither substitutes for the other.

For healthcare organizations, the objective is to maintain reliable safeguards throughout the transfusion pathway rather than depend on a single technology to compensate for weaknesses elsewhere.

Hemovigilance Extends Beyond Recording Transfusion Reactions

Hemovigilance is the systematic surveillance of the transfusion chain, including recognition, reporting, investigation and analysis of adverse reactions and incidents.

The International Society of Blood Transfusion describes hemovigilance as encompassing activities from blood collection through transfusion and recipient follow-up.

For hospitals, effective hemovigilance is not limited to documenting reactions after a blood unit has been administered. It also involves identifying errors and near misses, understanding contributing factors and implementing measures to prevent recurrence.

Some of the most useful safety information comes from incidents detected before they reach the patient.

For example, a mismatch between a patient's identity and a blood sample label may be recognized before compatibility testing or blood issue. Although no transfusion reaction occurs, the incident reveals a weakness in the process that could have led to serious harm.

During and after transfusion, clinical teams must remain alert to possible adverse reactions, including hemolytic, allergic and febrile reactions, transfusion-associated circulatory overload and transfusion-related acute lung injury.

A clinical event occurring after transfusion does not automatically establish that transfusion caused it.

Assessment requires consideration of timing, the patient's underlying condition, clinical findings, investigation results and recognized case definitions.

This distinction is particularly important when AI is used to identify possible transfusion reactions from electronic records.

An algorithm may recognize fever, increasing oxygen requirements or documentation suggesting a possible reaction. Such findings may help identify cases requiring review, but the system should not independently establish a final diagnosis or causal relationship.

Qualified professionals must assess the event through the appropriate clinical and transfusion service processes.

Recent international evidence reinforces the importance of fundamental safety controls.

An international forum summary published in Vox Sanguinis in August 2026 examined practices for preventing ABO-incompatible red blood cell transfusion across participating blood establishments and hospitals.

The report highlighted the continuing importance of patient identification, specimen collection and labeling, blood issue and bedside verification. It also documented variation in how institutions combine manual and electronic safety measures.

The forum was descriptive and did not establish the comparative effectiveness of individual preventive interventions. Nevertheless, it illustrates why sophisticated decision support cannot compensate for weaknesses in basic transfusion safety practices.

WHO's Global Status Report on Blood Safety and Availability 2025, published in June 2026, also reinforces the importance of well-governed blood systems, quality assurance, appropriate clinical blood use and hemovigilance.

The WHO report draws on information reported by countries to assess blood system governance, availability, safety and clinical use. It does not provide evidence that AI-assisted transfusion indication review has improved outcomes in Vietnamese hospitals.

For AI governance, clinical indication review, compatibility assurance, administration monitoring and hemovigilance must remain connected but distinct responsibilities.

An AI system may support selected parts of the process, but organizational safety depends on the reliability of the entire transfusion pathway.

What the Evidence Shows About AI in Transfusion Medicine

AI applications in transfusion medicine have expanded across anemia detection, bleeding risk estimation, prediction of transfusion requirements, inventory management and clinical decision support.

However, the available evidence varies considerably according to the function being studied.

A systematic review published in the Journal of Clinical Medicine in November 2025 examined 338 studies involving AI applications across patient blood management and related transfusion activities.

The review identified promising technical performance in several domains, including anemia detection, bleeding risk prediction, transfusion requirement estimation and blood inventory forecasting.

Nevertheless, external validation and routine clinical deployment remained limited. The authors emphasized the need for standardized reporting, independent validation, explainability, prospective evaluation and integration into actual clinical workflows.

A particularly relevant implementation-focused review was published by Maynard and colleagues in Transfusion Medicine Reviews in January 2026.

The researchers examined recent studies reporting prospective AI or machine-learning deployment with integration into clinical workflows. Their literature search, conducted in June 2025, covered publications from January 2022 onward.

Of 1,243 records screened and 31 full-text articles reviewed, only three studies met the review's specified inclusion criteria.

These included a laboratory-integrated system predicting low ferritin in adults with anemia, a patient-facing smartphone application estimating hemoglobin from fingernail images, and a clinician-facing application predicting resuscitation needs in trauma.

The three studies were examples of prospective implementation under the review's specific eligibility criteria. They were not three trials of AI independently approving red blood cell transfusion orders, nor should they be interpreted as representing the entirety of research on clinical AI in transfusion medicine.

Importantly, none demonstrated improvements in clinical outcomes or healthcare costs.

The authors identified persistent challenges involving data quality, governance, interoperability, generalizability, clinical trust and the absence of economic evaluation.

These findings do not establish that AI lacks clinical value. They indicate that evidence supporting technical prediction is substantially broader than evidence demonstrating improved outcomes after routine implementation.

For hospitals considering AI-assisted transfusion indication review, this distinction matters.

A model that predicts transfusion requirements accurately may be useful for planning blood availability. It does not necessarily follow that the same model can determine whether an individual order is appropriate.

Similarly, software that detects possible anemia or predicts hemorrhage may support patient blood management without directly improving transfusion approval decisions.

Evidence from conventional computerized clinical decision support also needs to be distinguished from evidence involving machine learning.

Rule-based alerts, structured electronic order sets and standardized transfusion protocols may improve selected aspects of practice. Their reported effects should not automatically be attributed to AI.

The appropriate conclusion is that AI-assisted transfusion decision support represents a promising area for carefully governed evaluation.

Its clinical value needs to be demonstrated through reliable decisions, safe implementation and meaningful patient outcomes rather than predictive performance alone.

The Vietnamese Context: From Blood Availability to Patient Blood Management

Vietnamese hospitals operate in different clinical and organizational environments.

Large tertiary hospitals may have specialized transfusion medicine services, extensive laboratory capabilities and integrated electronic systems. Other facilities may depend on external blood suppliers, have less comprehensive digital integration or face different constraints in accessing specialist support.

These differences influence what an AI system can realistically achieve.

A decision-support tool may identify a potential transfusion need, but it cannot guarantee that an appropriate blood component is available, compatible and delivered safely.

Similarly, a system may recognize potentially inappropriate blood use, but it cannot independently resolve inconsistent clinical protocols, inadequate documentation or insufficient staff training.

Recent professional discussions in Viet Nam demonstrate increasing attention to patient blood management, digital transformation and multidisciplinary transfusion safety.

In a report published by Bach Mai Hospital on 26 May 2026, hospital leaders and specialists emphasized the importance of moving beyond a narrow focus on blood units toward management of the patient's blood health and optimization of clinical outcomes.

The discussions addressed multidisciplinary coordination, prevention and management of anemia, safer clinical transfusion practices and the potential role of digital systems in strengthening communication among clinical departments, laboratories and blood banks.

The National Institute of Hematology and Blood Transfusion also highlighted transfusion safety and immunohematological challenges during the Third National Blood Transfusion Conference and associated professional sessions held in Hanoi on 15–16 April 2026.

Professional presentations emphasized the need for proactive hemovigilance, stronger clinical governance, improved data quality and more effective learning from transfusion-related incidents.

Discussions about digital transformation also considered the potential application of AI and large-scale data analytics in blood service management and forecasting.

These developments suggest growing professional attention to strengthening transfusion quality and safety in Viet Nam.

However, conference presentations and institutional initiatives should not be interpreted as evidence that AI-assisted transfusion decisions have already been validated or implemented successfully across Vietnamese hospitals.

Each healthcare organization needs to examine its own transfusion practice before determining whether AI is appropriate.

The most important problem may involve inconsistent clinical indications, inadequate reassessment, incomplete documentation, weaknesses in specimen identification or gaps in adverse-event reporting.

Different problems require different solutions.

In some settings, strengthening existing clinical protocols, staff competency and electronic verification may provide greater immediate benefit than introducing a new predictive model.

The Hospital Transfusion Committee as a Governance Mechanism

One advantage of addressing AI-assisted transfusion decisions in Viet Nam is that an established organizational structure already exists for transfusion governance.

Circular No. 26/2013/TT-BYT provides for the Hospital Transfusion Committee.

Articles 59 and 60 describe its composition, functions and responsibilities. The committee advises healthcare facility leadership on safe, appropriate and effective transfusion and supports the development of clinical procedures, training, review of transfusion practices, analysis of adverse events and quality improvement.

Article 67 further establishes responsibilities for healthcare facilities using blood, including relevant organizational and professional arrangements.

Under Article 59, the Hospital Transfusion Committee may be combined with the Drug and Therapeutics Committee when the necessary membership, functions and responsibilities are incorporated in accordance with the Circular.

This existing structure is directly relevant to AI-assisted transfusion decisions.

A hospital may not need to establish a separate transfusion-specific AI committee if its existing governance arrangements can provide appropriate oversight.

Instead, the Hospital Transfusion Committee can work with clinical departments, laboratory and transfusion services, nursing, quality management, information technology and the organization's broader AI governance functions.

The committee can help determine whether a proposed AI system addresses a clinically meaningful problem, whether its outputs are consistent with locally approved practice and whether its performance has been adequately evaluated before clinical use.

It may also review implementation findings, clinically important false alerts, inappropriate treatment delays and the potential effects of AI on transfusion appropriateness and safety.

Responsibilities must nevertheless remain clear.

The committee provides organizational oversight. Individual clinical decisions remain with authorized healthcare professionals, while the transfusion service retains responsibility for its relevant laboratory and blood component activities.

Clinical and nursing teams continue to perform the professional checks, administration and monitoring required within their respective roles.

AI governance should reinforce these established responsibilities rather than create uncertainty about accountability.

Validating AI Before It Influences Transfusion Decisions

An AI-assisted transfusion decision system should not enter routine clinical use solely because its developer reports high predictive accuracy.

The hospital needs to determine whether the system performs reliably in the intended patient population and clinical workflow.

A reasonable initial validation population may consist of stable, nonbleeding adult inpatients for whom restrictive transfusion strategies and reassessment principles are applicable.

This provides a relatively defined clinical use case before considering more complex situations.

Patients with active major hemorrhage, acute myocardial infarction, specialized hematologic conditions or ongoing chronic transfusion requirements may need separate pathways or explicit exclusion from initial deployment.

The organization should determine whether the model correctly identifies orders requiring review, whether relevant clinical information is available and how frequently inappropriate warnings are generated.

Validation should use an appropriate reference assessment involving qualified clinical reviewers who consider the patient's condition, applicable guidance and documented circumstances of the transfusion decision.

The review should not rely solely on whether hemoglobin is above or below a selected threshold.

A low hemoglobin value does not automatically establish that transfusion is indicated, while a higher value does not necessarily make transfusion inappropriate.

The hospital should also test how the model behaves when clinical information is missing, delayed or contradictory.

Recently completed transfusions, pending laboratory results, emergency orders and evolving blood loss may all challenge a system that depends on simplified electronic data.

Cases not flagged by AI also require independent review. Evaluating only orders identified as potentially inappropriate may conceal situations in which the system failed to recognize an order requiring clinical attention.

Performance should be examined across relevant departments and patient groups, and use outside the validated population should be controlled.

Where large language models are used to interpret narrative documentation, validation should also consider local terminology, abbreviations, mixed Vietnamese and English records, unsupported conclusions and changes in model behavior following software updates.

The organization also needs to understand how the system affects clinical behavior.

A large number of low-value alerts may contribute to alert fatigue. Overemphasis on restrictive thresholds may encourage inappropriate hesitation when patients genuinely need blood.

An intervention that reduces unnecessary orders but also delays clinically necessary transfusions cannot be considered safe without further evaluation.

Validation must therefore extend beyond technical accuracy to the clinical consequences of acting, or failing to act, on AI-generated recommendations.

Measuring Appropriate Blood Use and Patient Safety

Meaningful evaluation of AI-assisted transfusion review requires more than counting blood units.

Aggregate red blood cell utilization may provide useful information, but a reduction in units per admission is not sufficient evidence of improved practice.

Changes in blood use may reflect patient case mix, surgical activity, clinical policy or treatment availability. In some circumstances, a reduction could also indicate that patients requiring transfusion are receiving inadequate or delayed treatment.

The first requirement is a reliable definition of appropriate transfusion within the clinical population being evaluated.

For stable, nonbleeding patients, review criteria should consider the applicable hemoglobin threshold, symptoms, clinical indications, available treatment alternatives and the documented reason for the order.

Patients with major hemorrhage or other conditions requiring different strategies should be evaluated using appropriate separate criteria.

The hospital can then examine the proportion of orders meeting locally approved indication criteria, the proportion requiring clarification and the outcomes of those reviews.

For eligible stable patients, additional measures may include appropriate single-unit transfusion practice and whether clinical reassessment occurred before further red blood cells were administered.

These indicators should not be applied indiscriminately to emergency hemorrhage or other excluded populations.

Patient-safety monitoring is equally important.

Hospitals should track clinically important transfusion reactions, wrong-patient and wrong-component incidents, near misses, delays in necessary transfusion and failures in required pretransfusion checks.

Relevant clinical outcomes may include transfusion-associated circulatory overload, acute hemolytic reactions, transfusion-related acute lung injury and other recognized complications, using consistent case definitions and appropriate professional investigation.

Because serious transfusion reactions are relatively uncommon, individual hospitals may require longer observation periods or broader surveillance arrangements to identify meaningful changes.

The quality of reporting must also be considered.

An increase in reported incidents after strengthening hemovigilance may reflect improved detection and a more open reporting culture rather than deterioration in actual patient safety.

Conversely, fewer incident reports do not necessarily indicate improvement if staff are less likely to report concerns.

Where AI is implemented, the hospital should also examine the number of alerts, their clinical relevance, the proportion reviewed, changes made to orders and situations in which clinicians appropriately override recommendations.

The effects on workload and treatment timeliness should be assessed.

A system requiring extensive additional review while producing little improvement may not represent an efficient use of hospital resources.

The purpose of monitoring is to establish whether the complete transfusion pathway is becoming safer, more appropriate and more reliable.

Learning From Transfusion Errors and Near Misses

Hemovigilance becomes most valuable when findings lead to practical improvements.

Consider a near miss in which a blood sample is identified as belonging to the wrong patient before a blood unit is issued.

The incident should prompt examination of specimen collection, patient identification, labeling and the systems supporting those activities.

An AI model designed to review transfusion indications may have no direct role in preventing that error.

The appropriate response may involve strengthening identification procedures, staff training or electronic matching controls rather than modifying the AI prediction system.

In another situation, a patient may receive an unnecessary additional red blood cell unit because no reassessment occurred after the first transfusion.

That event may reveal an opportunity to improve clinical decision support and the review of outstanding orders.

Where AI is involved, the hospital should determine whether the system generated a relevant warning, whether it reached the responsible clinician and whether the information was interpreted correctly.

The review should also examine whether the model relied on outdated information, failed to recognize an important clinical exception or presented uncertainty inadequately.

Not every disagreement between AI and a clinician represents an error.

A clinician may appropriately override a general recommendation because of patient-specific circumstances. Such decisions may reveal important limitations in the system or the scope of the clinical guidance being applied.

A useful review process distinguishes justified clinical exceptions from situations in which important safety concerns were overlooked.

It should encourage staff to report near misses, communicate uncertainty and contribute to improvement without assuming that the algorithm's recommendation is necessarily correct.

The objective is to strengthen the reliability of transfusion care, not to assign responsibility solely on the basis of an AI-generated output.

Legal, Professional and Data Governance in Viet Nam

AI-assisted red blood cell transfusion decision support must operate within Viet Nam's applicable healthcare, blood transfusion, artificial intelligence, medical device and personal data protection frameworks.

The Law on Medical Examination and Treatment No. 15/2023/QH15, as amended, provides the broader legal framework for healthcare practice, patient rights, medical records and professional responsibilities.

Circular No. 26/2013/TT-BYT, dated 16 September 2013, remains a directly relevant specialized legal instrument governing blood transfusion activities in Viet Nam.

It covers blood collection and processing, component management, compatibility testing, clinical preparation, administration, monitoring, risk surveillance and institutional transfusion governance.

As of 1 November 2026, the Circular remains recorded as effective in Viet Nam's national legal database. Healthcare organizations should apply it together with other applicable legal provisions and locally approved professional procedures.

The Circular does not establish a dedicated AI-specific transfusion decision-support framework.

Consequently, existing mandatory transfusion procedures should be distinguished from additional AI validation, monitoring and governance measures proposed by healthcare organizations.

An AI system does not replace the responsibility of the treating clinician to assess the indication for transfusion, nor does it replace the duties of laboratory, transfusion service and clinical staff in ensuring safe administration.

Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on 1 March 2026. Decree No. 142/2026/ND-CP, effective from 1 May 2026, provides implementing provisions, while Decision No. 33/2026/QD-TTg, effective from 15 August 2026, establishes the list of high-risk AI systems under the applicable framework.

An AI application supporting transfusion decisions should not automatically be assigned a particular legal risk classification solely because it is used in healthcare.

Regulatory classification requires assessment of the system's intended purpose, actual functions, degree of automation, deployment circumstances and applicable legal criteria.

Clinical risk must also be considered independently.

A system that delays necessary transfusion, encourages inappropriate blood use or overlooks important clinical exceptions may create substantial patient-safety risks regardless of whether it is expressly included in a particular AI risk listing.

Healthcare organizations should also determine whether the software falls within Viet Nam's medical device regulatory framework.

Relevant provisions may include Decree No. 98/2021/ND-CP on medical device management, as amended, and Ministry of Health Circular No. 24/2026/TT-BYT, effective from 1 July 2026, concerning risk determination and management measures for medical devices and related regulatory amendments.

Software used solely for administrative reporting may have a different intended purpose from a system designed to influence clinical treatment decisions.

The use of AI alone does not automatically establish medical device status. However, describing a product as a clinical workflow assistant or requiring a clinician to approve its outputs does not remove the need to assess its actual medical functions.

A product-specific regulatory assessment is therefore necessary before commercialization, procurement or routine clinical deployment.

Personal data protection presents another important responsibility.

AI-assisted transfusion review may involve identifiable information concerning diagnoses, laboratory results, blood groups, transfusion histories, clinical conditions and treatment decisions.

The Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/ND-CP, both effective from 1 January 2026, provide relevant requirements for personal data processing and protection.

Healthcare organizations should establish the applicable lawful basis and purposes for processing, restrict access to authorized personnel and implement appropriate security, retention and accountability arrangements.

Where an external technology provider processes identifiable patient information, responsibilities for data access, hosting, maintenance, secondary use and incident management should be clearly defined.

Using clinical records for AI development, training, validation, retraining or commercial model improvement requires separate consideration of the relevant purposes, lawful grounds and applicable safeguards.

The existence of a clinical record does not automatically authorize unrestricted use of identifiable patient information for unrelated model development.

Cross-border processing, outsourcing and external hosting may create additional legal, documentation and impact-assessment obligations where applicable.

Hospitals should maintain appropriate records of clinically important AI outputs and related professional decisions while protecting patient confidentiality.

International guidelines from AABB, WHO, NICE and other professional organizations can inform local clinical practice, but they do not automatically create binding obligations under Vietnamese law.

Similarly, practical governance arrangements proposed for AI implementation should be distinguished from specific statutory requirements unless an applicable legal provision establishes them.

A Practical Starting Point for Vietnamese Hospitals

A hospital considering AI-assisted transfusion indication review should begin by examining its existing transfusion practice.

An initial assessment may focus on stable adult inpatients in one or two departments, where the clinical population and applicable transfusion guidance can be defined clearly.

The hospital can review a representative sample of red blood cell orders and examine the documented indication, pretransfusion hemoglobin, symptoms, bleeding status, relevant comorbidities and number of units requested.

The review should also determine whether appropriate reassessment occurred, whether necessary treatment was delayed and whether any safety concerns were identified.

The Hospital Transfusion Committee, clinical departments, laboratory and transfusion services, nursing leadership and quality management should contribute to interpreting the findings.

The results may reveal problems that can be addressed without AI.

Inconsistent application of restrictive transfusion guidance may respond to updated local protocols, clinician education and structured electronic order sets.

Inadequate reassessment after an initial unit may require clearer clinical responsibilities and workflow changes.

Incorrect specimen labeling or patient identification may require improvements in verification procedures rather than an AI model assessing transfusion indications.

If AI is expected to provide additional benefit, the organization should define a narrow clinical function.

One possible initial use case is identifying red blood cell orders in stable, nonbleeding adults that warrant review against locally approved indication criteria.

During an initial silent evaluation, the system could analyze historical or current records without directing treatment decisions.

Its outputs should be compared with independent clinical assessment to determine whether the system identifies meaningful opportunities for improvement and handles clinically important exceptions appropriately.

The evaluation should include false alerts, missed cases, delayed data, emergency scenarios and the potential impact on clinical workload.

If performance is acceptable and relevant regulatory requirements have been addressed, a limited supervised clinical pilot may be considered.

During the pilot, AI should provide review information to qualified clinicians while established prescribing, compatibility testing and transfusion safety processes remain fully operational.

The hospital should monitor whether the intervention improves the appropriateness of clinical decisions, whether it introduces unnecessary delays and whether clinicians can respond to alerts without excessive workload.

The Hospital Transfusion Committee and relevant governance functions should review these findings and recommend adjustments where necessary.

The organization should also establish responsibility for monitoring system performance, managing model changes, investigating incidents and restricting or suspending use when necessary.

Expansion to other departments or patient populations should depend on additional clinical validation and evidence of safe implementation.

A system evaluated in stable adult medical patients should not automatically be extended to trauma resuscitation, acute myocardial infarction, obstetric hemorrhage or specialized hematologic care without addressing the different clinical requirements.

Throughout implementation, the hospital must preserve its ability to assess patients, authorize transfusions and deliver blood components safely when the AI system is unavailable.

The objective is to strengthen an existing transfusion safety program, not to make clinical care dependent on one software product.

From AI Recommendations to Safer Transfusion Practice

AI-assisted red blood cell transfusion decision support offers a potential opportunity to improve how hospitals review clinical indications and use blood resources.

Emerging research suggests that AI can analyze complex clinical information, estimate transfusion requirements and support selected decision-making tasks.

However, evidence demonstrating improved patient outcomes after routine clinical implementation remains limited.

For healthcare organizations in Viet Nam, the most appropriate starting point is the existing professional and legal framework governing transfusion practice, supported by relevant international evidence and a realistic assessment of local clinical processes.

The decision to transfuse must reflect the patient's condition and the evidence applicable to that situation. Appropriate alternatives should be considered, and necessary transfusion must not be delayed by inflexible automated rules.

Once transfusion is authorized, safe care depends on reliable compatibility testing, accurate patient and blood-unit identification, correct administration and appropriate clinical monitoring.

Hemovigilance provides the mechanism for detecting adverse events, learning from near misses and improving the entire transfusion process.

AI may support parts of this system, but it does not replace these fundamental responsibilities.

A well-governed implementation should help clinicians recognize orders requiring review without discouraging justified treatment. It should provide useful information without creating excessive alerts and support timely decisions without interfering with emergency care.

Success should not be measured simply by the number of orders challenged by AI or the reduction in red blood cell units used.

The more meaningful measure is whether patients receive appropriate transfusion when needed, avoid unnecessary exposure when they do not, and remain protected by reliable clinical decision-making, transfusion safety procedures and organizational learning.

That is where AI-assisted transfusion decision support has the potential to contribute meaningfully to healthcare quality and patient safety.

Selected References and Regulatory Sources

Vietnamese Blood Transfusion and Healthcare Regulations

Ministry of Health of Viet Nam. Circular No. 26/2013/TT-BYT, dated 16 September 2013, Guidance on Blood Transfusion Activities (Hướng dẫn hoạt động truyền máu).

National Assembly of Viet Nam. Law on Medical Examination and Treatment No. 15/2023/QH15, as amended.

International Clinical Guidelines and Patient Blood Management

Carson JL, Stanworth SJ, Guyatt G, et al. Red Blood Cell Transfusion: 2023 AABB International Guidelines. JAMA. 2023;330(19):1892–1902. DOI: 10.1001/jama.2023.12914.

Pagano MB, Stanworth SJ, Dennis J, et al. Red Cell Transfusion in Acute Myocardial Infarction: AABB International Clinical Practice Guidelines. Annals of Internal Medicine. 2025;178(10):1469–1477. DOI: 10.7326/ANNALS-25-00706.

World Health Organization. Guidance on Implementing Patient Blood Management to Improve Global Blood Health Status. Geneva: WHO; 2025. ISBN: 978-92-4-010466-2.

World Health Organization. The Urgent Need to Implement Patient Blood Management: Policy Brief. Geneva: WHO; 2021. ISBN: 978-92-4-003574-4.

World Health Organization. Global Status Report on Blood Safety and Availability 2025. Geneva: WHO; published 12 June 2026. ISBN: 978-92-4-012154-6.

National Institute for Health and Care Excellence. Blood Transfusion. NICE Guideline NG24. First published 18 November 2015; last updated 26 February 2026.

International Society of Blood Transfusion. Haemovigilance Working Party: Haemovigilance Definitions and Resources.

Clinical AI, Implementation and Transfusion Safety Evidence

Coelho H, Silva F, Correia M, Rodrigues PM. Artificial Intelligence in Patient Blood Management: A Systematic Review of Predictive, Diagnostic, and Decision Support Applications. Journal of Clinical Medicine. 2025;14(23):8479. DOI: 10.3390/jcm14238479.

Maynard S, Farrington J, Raza S, Stanworth SJ. Artificial Intelligence Implementation in Transfusion Medicine: Addressing the Challenges of Clinical Adoption. Transfusion Medicine Reviews. 2026;40(1):150961. DOI: 10.1016/j.tmrv.2026.150961.

Mirrione-Savin A, Murphy MF, Schäfer R, et al. International Forum on the Prevention of ABO Incompatible Red Blood Cell Transfusion: Summary. Vox Sanguinis. 2026;121(9):1440–1450. DOI: 10.1111/vox.70326.

Vietnamese Healthcare Implementation Context

Bach Mai Hospital. Ensuring Clinical Blood Transfusion Safety: A Multidisciplinary Approach (Đảm bảo an toàn truyền máu lâm sàng: Tiếp cận đa chuyên khoa). Published 26 May 2026.

National Institute of Hematology and Blood Transfusion. Experts Recommend Strengthening Transfusion Safety (Các chuyên gia khuyến cáo tăng cường đảm bảo an toàn truyền máu). April 2026.

National Institute of Hematology and Blood Transfusion. Blood Transfusion Activities Continue to Be Implemented Stably and Effectively (Công tác truyền máu tiếp tục được triển khai ổn định, hiệu quả). Published 15 April 2026.

Relevant Vietnamese AI, Data Protection and Medical Device Framework

National Assembly of Viet Nam. Law on Artificial Intelligence No. 134/2025/QH15, effective 1 March 2026.

Government of Viet Nam. Decree No. 142/2026/ND-CP, detailing certain provisions and implementation measures of the Law on Artificial Intelligence, effective 1 May 2026.

Prime Minister of Viet Nam. Decision No. 33/2026/QD-TTg, promulgating the List of High-Risk Artificial Intelligence Systems, effective 15 August 2026.

National Assembly of Viet Nam. Law on Personal Data Protection No. 91/2025/QH15, effective 1 January 2026.

Government of Viet Nam. Decree No. 356/2025/ND-CP, detailing certain provisions and implementation measures of the Law on Personal Data Protection, effective 1 January 2026.

Government of Viet Nam. Decree No. 98/2021/ND-CP on Medical Device Management, as amended.

Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Medical Device Risk Determination, Management Measures and Related Regulatory Amendments, effective 1 July 2026.

Editorial and Clinical Notice

This article is intended for professional education, evidence-informed discussion, responsible healthcare AI governance and healthcare quality improvement. It draws on Vietnamese legal and professional requirements, international clinical guidance, published research and organizational implementation considerations.

The clinical scenarios are illustrative and do not represent documented individual patient cases.

The approaches discussed for AI-assisted transfusion indication review, clinician oversight, transfusion safety monitoring and quality evaluation do not establish a mandatory Vietnamese Ministry of Health AI implementation standard or constitute an independently validated clinical decision instrument.

This article does not provide individualized medical or legal advice, authorize automated transfusion decisions, recommend a transfusion regimen for a particular patient or determine the regulatory status of any specific AI product.

Red blood cell transfusion indications, treatment decisions, patient communication, compatibility testing, administration, monitoring and management of adverse reactions remain the responsibility of appropriately qualified healthcare professionals, consistent with applicable Vietnamese requirements, current clinical guidance and locally approved procedures.

AI recommendations must not delay necessary emergency transfusion or replace required blood safety checks. General hemoglobin thresholds should not be applied indiscriminately to patients with active major hemorrhage, acute myocardial infarction or other circumstances requiring individualized care.

The research discussed does not establish that AI-assisted transfusion indication review necessarily reduces adverse transfusion reactions, mortality, inappropriate transfusion or healthcare costs in Vietnamese healthcare settings. Local validation, appropriate clinical oversight, supervised implementation and ongoing evaluation remain necessary before claims of clinical effectiveness can be made.

Digital Medicine Vietnam (DMVN) | Evidence and regulatory review: 1 November 2026.