Governing AI-Assisted Venous Thromboembolism Prevention in Viet Nam: VTE and Bleeding Risk Assessment, Prophylaxis Decisions and Safety Monitoring

Governing AI-Assisted Venous Thromboembolism Prevention in Viet Nam: VTE and Bleeding Risk Assessment, Prophylaxis Decisions and Safety Monitoring

A patient is admitted to hospital for abdominal surgery. The medical record contains several factors associated with venous thromboembolism (VTE), including reduced mobility, older age and a history of cancer. An artificial intelligence (AI) system identifies an elevated VTE risk and prompts the surgical team to review the need for preventive treatment.

The alert draws attention to a legitimate concern, but the patient's postoperative condition presents another. Pharmacological prophylaxis may reduce the likelihood of thrombosis, while anticoagulation started at an inappropriate time could increase the risk of serious bleeding. The surgical team decides to reassess the bleeding risk before introducing an anticoagulant, considers suitable mechanical prevention measures and reviews the plan as the patient's condition improves.

The AI system may have identified the risk correctly. Whether it contributes to safer care, however, depends on what happens next. Clinicians must weigh competing risks, select appropriate preventive measures, ensure those measures are delivered and recognize when changes in the patient's condition require a different approach.

For healthcare organizations in Viet Nam, this is the central challenge of AI-assisted VTE prevention. The potential value lies in supporting timely, clinically appropriate prevention while avoiding unnecessary anticoagulation, missed contraindications and preventable bleeding. Identifying patients at risk is important, but it is only one part of that process.

Why VTE Prevention Requires More Than Risk Prediction

Venous thromboembolism includes deep vein thrombosis (DVT) and pulmonary embolism (PE). Both may occur during or after hospitalization, particularly among patients with reduced mobility, major surgery, serious medical illness or other recognized risk factors.

Hospital-associated VTE can cause substantial morbidity and mortality. Some patients also experience longer-term complications, including post-thrombotic syndrome following DVT and chronic thromboembolic pulmonary hypertension following PE. Preventing avoidable VTE is therefore an important hospital patient-safety responsibility.

Unlike some preventive interventions, VTE prophylaxis involves competing clinical risks. Anticoagulant medicines can reduce the likelihood of thromboembolism but may increase bleeding. Mechanical approaches, including intermittent pneumatic compression, avoid anticoagulant-related bleeding but are not appropriate or sufficient in every situation.

The right prevention strategy depends on the reason for hospitalization, the patient's current condition, underlying thrombotic risk, bleeding risk, contraindications and anticipated course of treatment.

International clinical guidance consistently recognizes this balance. The National Institute for Health and Care Excellence (NICE) guideline NG89 recommends assessing both VTE and bleeding risk, with reassessment when clinically indicated. The American Society of Hematology (ASH) guidelines distinguish between medical and surgical patients and address circumstances in which pharmacological or mechanical prophylaxis may be appropriate.

An elevated VTE risk score does not automatically establish that anticoagulation should be prescribed. A system that recognizes thrombotic risk without adequately accounting for bleeding risk, contraindications and clinical circumstances could encourage inappropriate treatment.

The purpose of AI-assisted prevention should therefore be to strengthen clinical decision-making and the delivery of preventive care, rather than simply increase the number of patients classified as high-risk or receiving anticoagulants.

The Clinical Framework Already Available in Viet Nam

Healthcare organizations in Viet Nam have an established national professional reference for VTE prevention.

Ministry of Health Decision No. 3908/QD-BYT, dated 20 October 2023, issued Guidelines for Prophylactic Treatment of Venous Thromboembolism (Hướng dẫn điều trị dự phòng thuyên tắc huyết khối tĩnh mạch). The guidance applies to healthcare examination and treatment facilities nationwide and addresses VTE prevention across several clinical populations.

Its recommendations cover general prevention strategies, acute medical illness, intensive care, cancer, general surgery, neurosurgery, orthopedic trauma and obstetric care, among other situations. The document also addresses the management of bleeding complications associated with anticoagulant treatment.

For hospitals considering AI-assisted prevention, this guidance provides an important foundation. The immediate task is not to develop a new algorithmic standard, but to understand whether existing clinical processes consistently translate applicable professional recommendations into daily care.

For acute medical inpatients, the Ministry of Health guidance calls for assessment of VTE risk, bleeding risk and anticoagulant contraindications within 24 hours of admission, followed by reassessment every 48–72 hours or when the patient's clinical condition or treatment objectives change.

The guidance incorporates the Padua Prediction Score for VTE risk assessment in medical patients and the IMPROVE bleeding risk score. For general surgical patients, it includes the Caprini risk assessment model and procedure-related prevention recommendations.

These models and assessment intervals should be applied within their intended clinical populations. A pathway designed for acute medical patients should not automatically be extended to neurosurgical, obstetric, trauma or other specialized populations without considering their particular clinical requirements.

International guidance provides additional context. The Australian Commission on Safety and Quality in Health Care's Venous Thromboembolism Prevention Clinical Care Standard emphasizes individualized risk assessment, documented prevention planning, patient participation, implementation, reassessment and continuity of care. The 2020 standard is undergoing review, with an updated version anticipated in late 2027.

Safer Care Victoria's guidance for VTE prevention in hospitalized adults, with its guidance page updated in June 2026, also provides a contemporary reference for reviewing local practices.

For Vietnamese hospitals, the practical priority is to establish a clinically coherent, locally approved prevention pathway grounded in national professional guidance and informed by relevant international evidence.

AI should be evaluated against that pathway rather than allowed to determine the pathway independently.

What AI Can Add to Conventional Risk Assessment

Not every digital VTE risk assessment tool uses artificial intelligence.

Padua and Caprini are conventional clinical risk assessment models that assign points to defined patient characteristics. IMPROVE includes tools for assessing VTE risk or bleeding risk, depending on the specific model used.

These scores can be incorporated into an electronic health record and calculated automatically using programmed rules. Automating an established risk score does not, by itself, make the software an AI system.

Machine-learning models may perform a different function by analyzing combinations of clinical variables to estimate the probability of VTE over a specified period. Large language models may help identify relevant information within narrative medical records, particularly when risk factors are not consistently documented in structured fields.

For example, a previous episode of DVT may be mentioned in a surgical assessment but absent from the dedicated risk-factor section of the electronic record. An AI-assisted system could potentially recognize that history and prompt a clinician to verify it before completing the assessment.

Such a function may improve the completeness of clinical information. However, extracting risk factors, calculating a conventional score and predicting future clinical outcomes are different tasks.

A system that accurately reconstructs a Caprini score from medical records has not necessarily demonstrated better VTE prediction than the original Caprini model. Similarly, a model developed to predict pulmonary embolism should not automatically be used to recommend preventive anticoagulation.

Before implementation, healthcare organizations need to define precisely what the technology is intended to accomplish. It may identify missing risk factors, calculate an established score, estimate future VTE risk or notify a clinician that a prevention decision requires review.

Each function has different clinical implications and requires appropriate validation.

Prediction models should also specify their intended clinical population, target outcome and time horizon. Estimating VTE risk during hospitalization differs from predicting events within 30 or 90 days after discharge.

Without these distinctions, even a statistically accurate model may produce information that is difficult to interpret or use safely in practice.

Balancing Thrombosis and Bleeding Risks

Safe VTE prevention requires simultaneous consideration of thrombotic risk and the potential harms of preventive treatment.

Previous thromboembolism, active malignancy, major surgery, acute medical illness, immobility and certain thrombophilic conditions may increase the likelihood of VTE. Their relative importance varies according to the patient's clinical circumstances.

A patient undergoing elective surgery may have a different risk profile from someone admitted with acute medical illness. Patients receiving intensive care may experience rapidly changing risks that are not adequately represented by information recorded at admission.

Bleeding risk requires equally careful attention. Active bleeding, recent significant hemorrhage, severe thrombocytopenia, selected intracranial conditions and other clinically important findings may influence whether pharmacological prophylaxis is appropriate.

Renal impairment, extremes of body weight, concurrent medicines and planned invasive procedures may affect anticoagulant selection, dosing, monitoring or administration timing. Patients undergoing spinal or epidural anesthesia require particular attention to anticoagulant timing and procedural safety.

Not all these factors represent absolute contraindications. Some require treatment modification, closer monitoring or coordination between clinical teams. These distinctions cannot reliably be reduced to a single risk score.

Incomplete information introduces another source of risk.

An AI system may recognize immobility and cancer but overlook a recent bleeding event documented only in an operative note. Laboratory measurements may be outdated, or medication records may not accurately show whether a patient is already receiving therapeutic anticoagulation.

When clinically important information is unavailable, the system should make that uncertainty visible and prompt verification rather than assume that no contraindication exists.

The prevention pathway must also distinguish patients at risk of VTE from those with suspected or confirmed thrombosis.

A patient with symptoms suggesting acute DVT or pulmonary embolism requires appropriate diagnostic assessment and, when indicated, treatment. Routine prophylaxis scoring must not delay investigation of a potentially established thromboembolic event.

Patients already receiving therapeutic anticoagulation for another indication generally do not require an additional prophylactic anticoagulant. Treatment interruption, changing procedural risks and other special circumstances nevertheless require individualized review.

The clinical objective is not to maximize anticoagulant use. It is to identify the preventive approach that offers an appropriate balance of expected benefit and potential harm.

Choosing the Right Prevention Plan and Ensuring It Is Delivered

Once VTE and bleeding risks have been assessed, the clinical team needs an individualized prevention plan.

For some patients, pharmacological prophylaxis will be appropriate. For others, mechanical prophylaxis may be preferred, particularly when anticoagulant-related bleeding risk is temporarily unacceptable. Selected patients may require a combination of approaches, while those at sufficiently low risk may not need pharmacological prophylaxis.

Recommendations differ between clinical populations. ASH guidance for acutely ill medical patients does not necessarily support the same approach as recommendations for major surgical patients. Orthopedic, neurosurgical, obstetric and cancer-related situations require attention to their specific clinical guidance.

An AI system should not convert every elevated VTE risk score into the same treatment recommendation.

Medication selection and administration require professional review of kidney function, body weight, platelet count, bleeding history, concurrent medicines and planned procedures. Where heparin products are used, appropriate monitoring and assessment for complications such as heparin-induced thrombocytopenia may also be necessary.

Mechanical prophylaxis is not free of potential harm. Intermittent pneumatic compression devices must be appropriately selected, fitted and used, with monitoring for discomfort, skin injury or other complications. Compression stockings may be unsuitable for patients with certain vascular, neurological or skin conditions.

Early mobilization and appropriate hydration can support prevention where clinically suitable, but they should not be assumed to replace indicated pharmacological or mechanical prophylaxis.

Selecting an appropriate prevention plan is only one part of the process. The next challenge is ensuring that the plan is carried out correctly.

A patient may have an appropriate anticoagulant order, yet the medicine may not be administered at the intended time. It may be withheld because of a procedure, missed during transfer between departments or discontinued without a documented reassessment plan.

Similarly, a mechanical device may be ordered but remain unavailable, incorrectly fitted or unused for substantial periods.

A digital system that confirms only the existence of a prophylaxis order may therefore overestimate how reliably preventive care is being delivered.

Hospitals need to distinguish among an appropriate clinical decision, an appropriate prescription or order, and the actual delivery of prevention.

Consider a patient whose anticoagulant is withheld because of postoperative bleeding concerns. The initial decision may be clinically justified. However, if bleeding risk subsequently decreases while VTE risk remains elevated, continued withholding without reassessment may become a missed prevention opportunity.

The original decision may not have been wrong. The problem may lie in the absence of a reliable process for reviewing that decision as the patient's condition changes.

AI-supported monitoring could potentially identify patients whose planned prophylaxis has not been administered or whose temporary contraindications have not been reassessed. However, this requires reliable information about why treatment was withheld, whether alternative measures were used and whether the clinical team has already reviewed the situation.

Otherwise, repeated alerts could challenge appropriate decisions or encourage unsafe treatment changes.

Medication administration records, nursing documentation, pharmacy review and clinical notes all contribute to understanding whether the prevention plan has been carried out. No single electronic field should automatically be treated as a complete representation of the patient's care.

Reassessment During Hospitalization and at Discharge

VTE and bleeding risks change throughout hospitalization.

A patient who is mobile on admission may become immobilized after surgery or clinical deterioration. Another patient may recover mobility more quickly than expected. A bleeding concern that initially prevents anticoagulation may resolve, while the underlying thrombotic risk remains substantial.

Clinical teams therefore need to revisit prevention decisions at meaningful points during care.

Changes in mobility, surgery, invasive procedures, significant bleeding, kidney function, admission to intensive care and transfer between departments may all justify reassessment.

The Ministry of Health's 2023 guidance and NICE NG89 recognize the importance of reassessment, although recommended intervals and clinical considerations vary across patient populations.

AI may help identify relevant changes in recorded information, but the organization must determine which changes require clinical review and how urgently they should be addressed.

Repeated low-value notifications can contribute to alert fatigue. A useful notification should identify what has changed, explain why the information matters and allow the clinician to examine the supporting evidence.

A message stating only that a patient is at high risk offers limited value when the clinical team cannot determine what remains unresolved.

An AI system should not independently modify anticoagulant treatment simply because a calculated risk score changes.

When a patient moves between departments, the receiving team should have access to the current prevention plan, reasons for withheld treatment and outstanding reassessment needs. This supports continuity of care rather than requiring each team to restart assessment without understanding earlier decisions.

Discharge presents another important transition.

VTE risk does not necessarily end when hospitalization ends. Some patients remain at increased risk after discharge, particularly following selected major surgical procedures or in other defined high-risk circumstances.

However, extended pharmacological prophylaxis is not appropriate for every hospitalized patient.

For acutely ill medical patients, ASH guidance generally favors prophylaxis during hospitalization rather than routine extended anticoagulant prophylaxis after discharge. Selected patients undergoing major orthopedic procedures or abdominal and pelvic cancer surgery may require different approaches, including consideration of extended prophylaxis when clinically indicated.

Viet Nam's Ministry of Health guidance also addresses extended prophylaxis in specified clinical circumstances.

The discharge decision should consider the indication for ongoing prevention, bleeding risk, contraindications, appropriate treatment duration and the patient's ability to follow the plan safely.

Where anticoagulant medication is prescribed after discharge, the patient should receive understandable information about its purpose, administration, possible adverse effects and circumstances requiring medical attention.

Discharge documentation should clearly communicate whether prevention is to continue, stop or undergo reassessment. Relevant information should also be transferred to the healthcare professionals responsible for subsequent care.

An AI-generated discharge recommendation may support this process, but it should not replace the responsible clinician's assessment.

What the Evidence Shows About AI-Assisted VTE Prevention

Published research provides grounds for cautious interest in AI-assisted VTE prevention, although the evidence remains uneven.

In October 2023, Huang and colleagues published a prospective randomized controlled study in BMJ Open Quality involving 19,785 hospitalized patients at a single center in China.

The intervention incorporated an AI-assisted clinical decision-support system that integrated electronic clinical information, supported VTE and bleeding risk assessment and generated alerts intended to improve prevention.

Hospital-associated VTE occurred in 49 of 9,890 patients receiving conventional care, compared with 27 of 9,895 patients in the AI-assisted intervention group.

The corresponding event rates were 0.50% and 0.27%, representing an observed relative reduction of approximately 46% and an absolute difference of about 0.23 percentage points.

This is an encouraging result because the study evaluated clinical events rather than prediction accuracy alone. However, the findings require careful interpretation.

The reduction in overall VTE was largely associated with fewer DVT events. The trial did not demonstrate a statistically significant reduction in pulmonary embolism considered separately.

The intervention also did not demonstrate statistically significant improvements in the overall proportion of patients receiving prophylaxis or in the proportion receiving appropriate prophylaxis. The study reported changes in mechanical prophylaxis use and medication-use intensity, but these findings do not establish that the intervention improved every aspect of prevention quality.

The trial was conducted at one center, and its results should not automatically be generalized to other hospitals or patient populations. It also does not establish that AI-assisted prevention reduced major bleeding or other relevant adverse outcomes.

The observed difference in VTE incidence therefore supports the potential benefit of the specific intervention studied, while leaving questions about reproducibility, mechanisms of benefit and overall safety.

A different type of evidence emerged in July 2026, when Ma and colleagues published a multicenter validation study in npj Digital Medicine examining large language models for automated Padua and Caprini risk scoring.

The researchers used anonymized electronic health records from 30 hospitals in China. Fifty cases from 10 hospitals were used during development, while 200 cases from the remaining 20 hospitals were used for testing.

Expert knowledge-augmented prompting improved performance compared with simpler prompting approaches.

In the test dataset, the reported F1 score for overall risk stratification was 0.96 for Padua and 0.64 for Caprini. Performance on individual score components was generally higher than performance on the final risk classification, illustrating the difference between recognizing individual risk factors and assigning the correct overall risk category.

These findings suggest that large language models may assist with extracting clinical information and reconstructing established risk scores. They also demonstrate that performance can vary substantially between assessment tasks.

The study did not establish that LLM-assisted scoring improved prophylaxis decisions, reduced VTE or prevented bleeding complications after routine clinical deployment.

In August 2026, Moreno and colleagues published a review in Journal of Vascular Surgery: Venous and Lymphatic Disorders examining AI applications across VTE prediction, detection and clinical interventions.

The review identified 23 studies, including 20 prediction or detection studies and three clinical-impact intervention studies.

Among the 20 prediction or detection studies evaluated using PROBAST+AI, 18 were judged to have a high risk of bias. Only seven demonstrated meaningful external or prospective validation, and five of those seven still had a high risk of bias. Calibration was not reported in 14 studies.

These findings are important for clinical implementation. Strong discrimination statistics may create an impression that a model is ready for routine use, even when questions about validation, calibration and generalizability remain unresolved.

Taken together, the available evidence suggests that AI may contribute to improved risk recognition and selected clinical outcomes, but prediction accuracy alone does not establish that implementation will improve patient safety.

For hospitals considering AI-assisted VTE prevention, local evaluation should examine clinical workflow, thrombosis outcomes, bleeding-related safety, unintended consequences and the reliability of prevention delivery.

What Vietnamese Implementation Research Reveals

Evidence from Viet Nam reinforces the importance of examining how prevention is delivered in everyday clinical practice.

A mixed-methods study published in Vascular Health and Risk Management in April 2025 examined VTE prophylaxis practices among patients undergoing abdominal and pelvic surgery.

The researchers reviewed 240 medical records from April 2023 and conducted interviews with 16 surgeons to investigate factors influencing prevention practices.

Only 11.7% of the reviewed cases received a prophylaxis method considered appropriate under the study's assessment criteria.

Among patients receiving anticoagulant prophylaxis, adherence was relatively high for medicine selection and dosing, at 100% and 89.3%, respectively. However, adherence was lower for the timing of initiation, at 50%, and treatment duration, at 28.6%.

These findings illustrate an important difference between selecting an appropriate anticoagulant and delivering a complete prevention regimen.

The qualitative component identified concerns about bleeding, resource limitations, insufficient knowledge and training, uncertainty about preventive benefits and gaps in institutional protocols as barriers to consistent practice.

Potential enablers included leadership commitment, multidisciplinary collaboration, training, computerized clinical support and formal anticoagulation stewardship activities.

The findings suggest that prevention gaps may persist even when clinicians have access to appropriate medicines and recognized risk assessment tools.

The difficulty may lie in translating recommendations into consistent decisions, timely administration, appropriate treatment duration and reliable reassessment.

This study reflects a particular clinical setting and patient sample. Its findings should not be interpreted as the national rate of appropriate VTE prophylaxis in Viet Nam or as a direct estimate of practice across all hospitals in 2026.

Nevertheless, it provides a useful example of why implementation conditions need to be understood before introducing additional technology.

If a hospital's main problems involve inadequate protocols, inconsistent training or incomplete reassessment, a sophisticated AI prediction model may not address the underlying causes. Improving existing prevention processes may be necessary before AI can add meaningful clinical value.

Evaluating AI-Assisted Prevention in Vietnamese Hospitals

A hospital considering AI-assisted VTE prevention should begin by defining the intended clinical function and identifying the population in which the system will operate.

A model designed to extract risk factors from narrative documentation requires evaluation of information extraction accuracy and the resulting risk classification. A predictive model requires assessment against a clinically meaningful outcome over a defined period. A system generating reminders about prevention decisions needs evaluation of whether those reminders are appropriate, timely and useful.

These functions should not be treated as interchangeable.

Local validation should use representative patient records from the intended departments and clinical settings. Vietnamese hospitals may have documentation containing Vietnamese and English terminology, inconsistent abbreviations, incomplete medical histories and information distributed across several clinical systems.

Errors in patient identification, medication records, laboratory measurements or previous VTE history may substantially influence the system's output.

Evaluation should therefore include both patients identified by AI and patients the system does not flag. Reviewing only AI-positive cases may conceal clinically important missed risks.

For predictive models, discrimination is only one aspect of performance. Calibration, event rates, clinically relevant thresholds, false-positive results, missed cases and the workload associated with alerts are also important.

A model may identify elevated VTE risk correctly but still be unsafe if it overlooks bleeding-related contraindications or encourages unnecessary prophylaxis.

Population differences require careful consideration. A model validated in general adult medical wards should not automatically be applied to pregnancy, pediatrics, major trauma, neurosurgery or other specialized populations without appropriate evidence.

Performance should also be examined across clinically relevant groups and when important information is incomplete.

An initial supervised evaluation can compare AI outputs with independent professional assessment without allowing the system to alter routine clinical decisions. This may reveal problems with data extraction, risk classification, inappropriate alerts and potential safety concerns before broader deployment.

Technical validation alone, however, cannot establish that implementation will improve care.

The hospital also needs to determine whether alerts result in appropriate clinical action and whether selected preventive measures are actually delivered.

A system that generates more alerts may increase workload without improving prevention if clinical teams lack time, resources or clearly assigned responsibilities to respond.

The evaluation should therefore extend across the full pathway from risk assessment to patient outcomes.

Timeliness and completeness of VTE and bleeding risk assessment are useful process measures. The appropriateness of prevention decisions is equally important, including justified decisions not to prescribe anticoagulants.

Implementation measures should examine whether prescribed medicines are administered, whether mechanical devices are correctly used and whether withheld or interrupted treatment is reassessed when indicated.

These indicators require carefully defined denominators and clinical criteria. A higher anticoagulant prescribing rate is not necessarily evidence of better prevention, particularly when some patients have low VTE risk or clinically important bleeding contraindications.

Outcome monitoring should include clinically confirmed hospital-associated VTE and relevant bleeding complications.

Where feasible, hospitals may separately examine DVT, pulmonary embolism, major bleeding and other clinically important anticoagulant-related adverse events, using consistent definitions.

A thrombosis present on admission should not be counted as a new hospital-associated event. Similarly, suspected VTE should not automatically be treated as a confirmed diagnosis.

The organization should define how events detected after discharge will be captured and how surveillance periods and attribution rules will be applied.

Changes in diagnostic testing can also influence observed event rates. More frequent imaging may increase detection, while reduced testing may conceal missed cases. Trends in VTE incidence should therefore be interpreted alongside diagnostic practices, patient characteristics and surveillance quality.

Major bleeding deserves particular attention. A reduction in thrombosis accompanied by an unacceptable increase in serious bleeding would not necessarily represent an improvement in overall patient safety.

Balancing measures should include unnecessary anticoagulation, inappropriate mechanical prophylaxis, excessive alerts, additional clinical workload and delays in decisions requiring urgent attention.

Changes observed after AI deployment cannot automatically be attributed to the technology, particularly when other aspects of the prevention program have changed at the same time.

Ongoing performance monitoring is necessary because patient populations, clinical practice, data sources and model versions may change. Hospitals should have procedures for investigating errors, reviewing unintended consequences and modifying or suspending an AI function when necessary.

The central question is whether the AI-supported prevention pathway delivers safer care than the previous process, not merely whether the model generates technically accurate predictions.

Learning From Disagreements and Missed Prevention Opportunities

Disagreements between AI outputs and clinical decisions can provide valuable information about both the technology and the hospital's prevention pathway.

Consider a patient with an elevated AI-generated VTE risk score but active bleeding. The clinician may appropriately decide against pharmacological prophylaxis at that time.

Such a decision should not automatically be interpreted as failure to follow the AI recommendation. The important questions are whether the bleeding concern is clinically justified, whether suitable alternative measures have been considered and whether a reassessment plan is in place.

In another situation, the AI system may classify a patient as low-risk while the clinical team identifies important information missing from the electronic record.

The clinician should be able to act on the available clinical evidence and investigate the discrepancy. The system's output should not be regarded as inherently more reliable simply because it was generated automatically.

Hospitals should review clinically important missed cases, unnecessary alerts and situations in which staff accepted AI recommendations despite conflicting clinical evidence.

These reviews should distinguish problems arising from incomplete data, incorrect information extraction, inappropriate model assumptions, workflow failures and genuine differences in clinical interpretation.

The objective is to improve the reliability of prevention rather than penalize clinicians for appropriately exercising professional judgment.

Where serious errors recur, the organization should have authority to restrict or suspend the relevant AI function. Essential clinical assessment and prevention responsibilities must continue when the technology is interrupted or unavailable.

Legal, Professional and Data Governance in Viet Nam

AI-assisted VTE prevention must operate within Viet Nam's applicable healthcare, artificial intelligence, medical device and personal data protection frameworks.

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

For VTE prevention, Ministry of Health Decision No. 3908/QD-BYT dated 20 October 2023 is a directly relevant national professional reference. Healthcare organizations should ensure that their locally approved prevention protocols are consistent with applicable Vietnamese requirements and the patient's clinical circumstances.

Introducing AI does not transfer prescribing authority or professional accountability from appropriately authorized healthcare practitioners to software.

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.

The healthcare-specific entries expressly identified in that list concern certain AI-assisted surgical and robotic surgical applications. AI-assisted VTE risk assessment and prevention is not separately named as a healthcare category in those entries.

However, this observation does not mean that every VTE-related AI application is automatically classified as low-risk or exempt from applicable obligations.

The regulatory position of a particular product depends on its intended purpose, functions, degree of automation, deployment arrangements and relevant legal requirements.

A system that extracts clinical information for professional review may have a different regulatory profile from software that independently generates treatment recommendations or directly influences anticoagulant prescribing.

Clinical risk and statutory AI risk classification also require separate consideration. A system that recommends inappropriate anticoagulation, overlooks active bleeding or fails to recognize important contraindications may create serious patient-safety concerns regardless of its classification under a particular regulatory list.

Healthcare organizations should additionally determine whether the software falls within Viet Nam's applicable medical device regulatory framework, including Decree No. 98/2021/ND-CP, as amended, and Ministry of Health Circular No. 24/2026/TT-BYT, effective from 1 July 2026.

The distinction between administrative support, clinical risk prediction and software with an intended medical purpose requires examination of the product's actual functions and applicable regulatory definitions.

A product's commercial description, or the fact that a healthcare professional reviews its output, does not by itself settle its regulatory classification.

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

Personal data protection is equally important.

AI-assisted VTE systems may process identifiable information concerning diagnoses, surgical procedures, medical histories, laboratory results, medicines and other health-related data.

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 identify the applicable lawful basis and defined purposes for processing, establish appropriate access controls and security safeguards, and clarify responsibilities when external technology providers handle patient information.

Where clinical records are used for AI development, validation, retraining or commercial model improvement, the organization should assess whether those activities are covered by the applicable legal grounds and what additional requirements may apply.

Patient information collected during clinical care should not automatically be made available for unrelated commercial model development.

External hosting, outsourcing, cross-border processing and secondary use of identifiable information also require assessment under applicable Vietnamese requirements, including relevant documentation and impact-assessment obligations where applicable.

These matters should be addressed before identifiable clinical information is transferred to an external AI service.

Hospitals should retain records sufficient to understand the system's role in clinically important decisions while protecting patient confidentiality and complying with applicable documentation requirements.

International recommendations from NICE, ASH and Australian healthcare quality organizations can inform local practice, but they do not automatically create binding obligations under Vietnamese law.

Similarly, organizational approaches proposed for AI governance should be distinguished from specific legal or professional requirements unless a relevant legal provision applies.

Starting With the Existing Prevention Process

For a hospital considering AI-assisted VTE prevention, the most useful first step may be to examine its existing prevention performance in one adult medical or surgical department.

The review should determine whether eligible patients receive timely VTE and bleeding risk assessments, whether prevention decisions are appropriate and whether planned interventions are actually delivered.

Importantly, the review should include situations in which prophylaxis was appropriately withheld. Evaluating only patients who received anticoagulants would provide an incomplete and potentially misleading picture of prevention quality.

Clinical leaders, pharmacists, nursing teams, information technology personnel and quality management staff can jointly examine the findings and identify the most important gaps.

Some hospitals may find that incomplete assessment or inconsistent documentation is the principal problem. Others may discover missed doses, inadequate communication during transfers or a lack of reassessment when temporary contraindications resolve.

These findings should determine the choice of intervention.

A structured electronic assessment or conventional clinical reminder may be sufficient where the main gap is incomplete documentation. More complex AI may be useful where important risk factors are buried in lengthy clinical notes or distributed across systems that are difficult to review manually.

A narrowly defined initial use case could involve AI-assisted extraction of VTE risk factors from clinical documentation, with clinicians verifying the identified information before it contributes to a prevention decision.

The system could first operate in a supervised evaluation setting, with outputs compared against independent professional review. The assessment should examine accuracy, clinically important omissions, inappropriate alerts and additional workload.

Only after satisfactory validation and relevant regulatory assessment should the organization consider a limited clinical pilot in which AI outputs contribute to established decision-making processes.

During the pilot, clinical and quality teams should examine whether alerts lead to appropriate review, whether contraindications are respected, whether preventive measures are delivered and whether unintended consequences occur.

The organization should also determine who is responsible for system monitoring, clinical incident review, model changes and decisions to restrict or discontinue the intervention.

Expansion to other departments should depend on demonstrated safety, clinical usefulness and satisfactory performance in the intended population.

The objective is to strengthen an existing VTE prevention program, not to create a new dependency on AI.

From Risk Prediction to Safer Prevention

Venous thromboembolism prevention illustrates both the potential and the limitations of AI in clinical care.

An AI system may help clinicians recognize risk factors, calculate established assessment scores or identify patients whose prevention plans require review. Published research suggests that selected AI-assisted clinical decision-support interventions can contribute to improved VTE outcomes in particular settings.

However, the broader evidence remains uneven. Many prediction models lack adequate external validation, and technical accuracy alone does not establish that implementation will improve patient outcomes.

For healthcare organizations in Viet Nam, the starting point should remain the national professional guidance, supported by relevant international recommendations and a realistic assessment of existing clinical practice.

Safe implementation requires simultaneous attention to thrombosis and bleeding risks, professional review of preventive decisions, reliable delivery of selected interventions and timely reassessment as patients' conditions change.

It also requires recognition of the potential harms of unnecessary prophylaxis, not simply the benefits of preventing thrombosis.

The measure of success is not how many patients an AI system identifies as high-risk or how many anticoagulant orders follow an alert.

It is whether the hospital consistently makes appropriate prevention decisions, delivers the planned care and reduces avoidable harm without introducing unacceptable bleeding risks or other unintended consequences.

That is where AI-assisted VTE prevention has the potential to become a meaningful part of healthcare quality and patient safety improvement.

Selected References and Regulatory Sources

Vietnamese Clinical Guidance

Ministry of Health of Viet Nam. Decision No. 3908/QD-BYT, dated 20 October 2023, issuing Guidelines for Prophylactic Treatment of Venous Thromboembolism (Hướng dẫn điều trị dự phòng thuyên tắc huyết khối tĩnh mạch).

International Clinical Guidelines and Quality Standards

National Institute for Health and Care Excellence. Venous Thromboembolism in Over 16s: Reducing the Risk of Hospital-Acquired Deep Vein Thrombosis or Pulmonary Embolism. NICE Guideline NG89. Published 21 March 2018; last updated 13 August 2019.

Schünemann HJ, Cushman M, Burnett AE, et al. American Society of Hematology 2018 Guidelines for Management of Venous Thromboembolism: Prophylaxis for Hospitalized and Nonhospitalized Medical Patients. Blood Advances. 2018;2(22):3198–3225. DOI: 10.1182/bloodadvances.2018022954.

Anderson DR, Morgano GP, Bennett C, et al. American Society of Hematology 2019 Guidelines for Management of Venous Thromboembolism: Prevention of Venous Thromboembolism in Surgical Hospitalized Patients. Blood Advances. 2019;3(23):3898–3944. DOI: 10.1182/bloodadvances.2019000975.

Australian Commission on Safety and Quality in Health Care. Venous Thromboembolism Prevention Clinical Care Standard. Sydney: ACSQHC; 2020.

Safer Care Victoria. Victorian Guideline for the Prevention of Venous Thromboembolism in Adult Hospitalised Patients. Guidance page updated 15 June 2026.

Clinical AI and Implementation Evidence

Huang X, Zhou S, Ma X, et al. Effectiveness of an Artificial Intelligence Clinical Assistant Decision Support System to Improve the Incidence of Hospital-Associated Venous Thromboembolism: A Prospective, Randomised Controlled Study. BMJ Open Quality. 2023;12(4). DOI: 10.1136/bmjoq-2023-002267.

Ma J, Wang D, Zhang Y, et al. Automated Risk Scoring for Venous Thromboembolism Using Large Language Models With Expert Knowledge-Augmented Prompting: A Multicenter Validation Study. npj Digital Medicine. Published 9 July 2026. DOI: 10.1038/s41746-026-02929-3.

Moreno O, Zhang R, Huang A, et al. Integrating Artificial Intelligence Into Venous Thromboembolism Care: Predictive Models, Implementation Challenges, and Future Directions. Journal of Vascular Surgery: Venous and Lymphatic Disorders. Published online 31 August 2026. Article 102608. DOI: 10.1016/j.jvsv.2026.102608.

Nguyen TTT, Tong HT, Nguyen HTL, Nguyen TD. A Call to Action for Anticoagulation Stewardship to Address Suboptimal Thromboprophylaxis Practices for at-Risk Non-Orthopedic Surgical Patients in Vietnam: An Explanatory Sequential Mixed-Methods Study. Vascular Health and Risk Management. 2025;21:305–326. DOI: 10.2147/VHRM.S505100.

Relevant Vietnamese Legal and Regulatory Framework

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

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.

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 Risk Determination and Management Measures for Medical Devices and Related Regulatory Amendments, effective 1 July 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.

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 professional guidance, international clinical recommendations, published research and relevant legal and regulatory sources.

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

The approaches discussed for AI-assisted VTE risk assessment, bleeding risk assessment, prophylaxis decision support, administration monitoring and quality evaluation are organizational implementation considerations. They 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, recommend an anticoagulant regimen for a particular patient, replace professional clinical judgment or constitute a determination of any AI product's regulatory status.

VTE and bleeding risk assessment, prophylaxis decisions, medication prescribing, administration, monitoring and management of complications must remain under the responsibility of appropriately qualified healthcare professionals, consistent with applicable Vietnamese requirements, current clinical guidance and locally approved procedures.

The research discussed does not establish that AI-assisted VTE prevention necessarily reduces thromboembolic events, bleeding complications or mortality in Vietnamese healthcare settings. Appropriate local validation, clinical oversight and ongoing monitoring remain necessary before claims of clinical effectiveness can be made.

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