Governing AI-Assisted Atrial Fibrillation Screening in Viet Nam: ECG-Based Risk Stratification, Rhythm Confirmation and Safe Follow-Up

Governing AI-Assisted Atrial Fibrillation Screening in Viet Nam: ECG-Based Risk Stratification, Rhythm Confirmation and Safe Follow-Up

IMPLEMENTATION INSIGHTS

Last reviewed: 28 October 2026

A 72-year-old patient visits an outpatient clinic for routine management of hypertension and diabetes. The patient has no known history of atrial fibrillation and has not experienced noticeable palpitations. A 12-lead electrocardiogram (ECG) recorded during an earlier appointment showed sinus rhythm.

The hospital has introduced an artificial intelligence (AI) application that analyzes existing ECG recordings to estimate the likelihood of atrial fibrillation being diagnosed over the following two years. When the patient's earlier ECG is analyzed, the system produces an elevated risk score.

The result raises several questions. Should the physician arrange another ECG, recommend longer rhythm monitoring, or consider the prediction alongside the patient's cardiovascular risk factors? If additional testing does not identify an arrhythmia, what should happen next?

These are practical questions for healthcare organizations considering AI-assisted atrial fibrillation screening. Predictive algorithms may help identify people who deserve closer attention, but their usefulness depends on how the information influences subsequent clinical decisions.

A risk score is not the same as a diagnosis, and a confirmed diagnosis does not automatically determine treatment. Healthcare professionals must interpret the available evidence, decide whether additional investigations are justified, and arrange appropriate follow-up.

For hospitals and clinics in Viet Nam, the challenge is therefore not simply whether AI can recognize patients at increased risk. It is whether that information helps clinicians make better decisions without creating unnecessary investigations, incorrect diagnoses, or inappropriate treatment.

Why atrial fibrillation screening requires careful decisions

Atrial fibrillation (AF) is a common cardiac arrhythmia associated with ischemic stroke, heart failure, hospitalization, and reduced quality of life. Its presentation varies considerably. Some patients experience palpitations, fatigue, or breathlessness, while others have no recognizable symptoms.

AF may also occur intermittently. A patient can have sinus rhythm during a routine consultation despite experiencing episodes at other times. This makes additional rhythm assessment potentially useful, particularly among older adults and people with cardiovascular risk factors.

The 2024 European Society of Cardiology (ESC) guidelines recommend routine heart rhythm assessment during healthcare contacts in individuals aged 65 years or older. They also recommend considering population-based screening using prolonged, noninvasive ECG monitoring in people aged 75 years or older, or those aged 65 years or older with additional CHA₂DS₂-VA risk factors [1].

These recommendations do not mean that every older adult requires prolonged ECG monitoring. The appropriate approach depends on the patient's circumstances, the screening method, available resources, and the potential consequences of detecting an arrhythmia.

International recommendations are not entirely uniform. In its 2022 recommendation, the US Preventive Services Task Force concluded that the available evidence was insufficient to determine the overall balance of benefits and harms of AF screening in asymptomatic adults aged 50 years or older without a history of stroke or transient ischemic attack [5].

The difference between these positions reflects an important unresolved issue. Detecting additional cases of AF is not necessarily equivalent to preventing more strokes or improving other clinical outcomes. The benefit of screening depends on which patients are assessed, which arrhythmias are identified, and whether subsequent management is appropriate.

AF screening should therefore be considered a structured clinical service rather than an isolated activity intended simply to increase the number of detected cases.

AI prediction, rhythm detection, and diagnosis are different activities

The expression "AI-assisted AF screening" can describe technologies with different clinical purposes.

Some systems analyze a 12-lead ECG recorded during sinus rhythm and estimate the likelihood of AF being diagnosed over a future period. Others examine a current ECG to identify an arrhythmia that may already be present. Another group analyzes longer recordings, including ambulatory or Holter ECG data, to detect possible intermittent AF episodes.

Although these applications all use ECG information, they answer different clinical questions and require different forms of evidence.

An AI model that predicts future AF risk from a sinus-rhythm ECG does not demonstrate that the patient was experiencing AF when the recording was obtained. Its output represents an estimate derived from patterns associated with subsequent AF diagnoses in the data used to develop the model.

This distinction is particularly important because a model predicting documented AF during follow-up does not necessarily identify the exact biological onset of the arrhythmia. Some patients may already have undetected intermittent AF, while others may develop the condition later.

By contrast, software that identifies possible AF in a current rhythm recording is making a classification about the recorded signal. That classification may still be incorrect. Frequent premature atrial contractions, recording artifacts, and other rhythm abnormalities can produce irregular patterns that require professional interpretation.

Systems analyzing prolonged recordings introduce further considerations, including recording quality, episode duration, and the clinical significance of detected abnormalities.

The consequences of an incorrect result also differ. An inaccurate risk prediction may lead to unnecessary monitoring or missed opportunities for assessment. An incorrect rhythm classification may result in a false diagnosis. Misinterpreting a brief device-detected episode may influence treatment decisions that require more careful evaluation.

Healthcare organizations should ensure that each application is used within its intended purpose. Evidence supporting a future-risk model should not automatically be generalized to software designed to diagnose current AF or evaluate arrhythmia burden.

The distinction must also be maintained in the medical record. A prediction of future AF should be documented as a risk estimate, not automatically entered as an established diagnosis.

What the 2026 VITAL-AF analysis adds to the evidence

An important contribution to this field was published in April 2026, when researchers reported a secondary analysis of the VITAL-AF trial in the Journal of the American College of Cardiology [3].

The original VITAL-AF trial was a cluster-randomized study involving 16 primary care practices in the United States. It examined whether single-lead ECG screening during routine visits increased the diagnosis of AF among adults aged 65 years or older.

Among 30,715 participants without previously diagnosed AF, new AF was identified during one-year follow-up in 1.72% of the screening group and 1.59% of the usual-care group. The difference was not statistically significant [4].

A prespecified subgroup analysis suggested greater AF detection among participants aged 85 years or older. However, this finding did not establish that the same benefit would occur across all age groups.

The 2026 secondary analysis addressed a different question. Rather than prospectively testing an AI-guided screening program, the investigators examined whether existing clinical and ECG information could identify individuals more likely to receive an AF diagnosis.

Participants were eligible for this analysis if they had a 12-lead ECG recorded within the three years before trial enrollment. Of 30,630 VITAL-AF participants without prevalent AF considered in the secondary analysis, 16,937 had the necessary ECG and clinical information.

Researchers compared three previously developed risk models: the established CHARGE-AF clinical risk model, an AI model using a prior 12-lead ECG, and a combined model incorporating both clinical and AI-derived ECG information.

The ECG-AI model achieved an area under the receiver operating characteristic curve (AUROC) of 0.784 for predicting AF diagnosed within two years. The CHARGE-AF model achieved an AUROC of 0.711, while the combined model achieved 0.788 [3].

These findings suggest that AI-derived ECG information may improve the ability to distinguish people who are subsequently diagnosed with AF from those who are not.

However, the combined model provided only a modest numerical improvement in AUROC compared with ECG-AI alone. The results should not be interpreted as proof that combining the models necessarily produces a meaningful improvement in clinical care.

The investigators also identified a subgroup with a relatively high observed rate of AF diagnosis. In the highest risk decile of the combined model, AF diagnoses occurred at 10.07 per 100 person-years in the screening group, compared with 7.76 per 100 person-years in the control group.

This finding suggests that risk-guided screening may be more productive among certain higher-risk individuals. Nevertheless, participants were not originally randomized according to AI-generated risk scores. The observed subgroup difference therefore cannot establish what would happen if an AI-directed screening program were introduced prospectively.

The analysis also highlighted an important trade-off. Restricting screening to patients with higher predicted risk may improve detection efficiency, but it can reduce the number of people reached and potentially exclude others who would have benefited from assessment.

Most importantly, the study did not establish that AI-guided screening reduces stroke, mortality, or other long-term adverse outcomes.

The findings provide evidence that ECG-based AI may help identify populations in which additional assessment could be useful. Whether that information improves patient outcomes remains a separate question.

Better prediction does not automatically mean better screening

A model with a higher AUROC may distinguish patients who later receive an AF diagnosis more accurately than a conventional clinical score. However, this does not establish that its predicted probabilities are accurate for a particular hospital population.

A model may rank patients reasonably well while systematically overestimating or underestimating their absolute probability of receiving an AF diagnosis. This distinction between discrimination and calibration becomes important when risk thresholds determine who receives additional monitoring.

If a hospital adopts a threshold developed in another population without appropriate evaluation, some patients may undergo unnecessary investigations while others may be excluded from potentially useful assessment.

The outcome being predicted also deserves careful interpretation. In many studies, the endpoint is a documented AF diagnosis during follow-up rather than the actual biological onset of the arrhythmia.

Whether AF is documented depends partly on whether the patient undergoes ECG testing, receives further clinical assessment, and remains engaged with healthcare services. A model predicting documented AF may therefore reflect both underlying cardiovascular risk and patterns of healthcare access and detection.

The availability of historical ECG recordings introduces another consideration. Patients who previously underwent ECG testing may differ from those who did not. In the VITAL-AF secondary analysis, only a subset of participants had the information required for ECG-based modeling.

These differences are relevant when considering whether an international model can be applied to Vietnamese patients.

A hospital needs to understand whether the model is appropriate for its intended population, whether the predicted probabilities are sufficiently reliable, and whether using them would change clinical decisions in a beneficial way.

A high score should not automatically become a mandatory indication for prolonged monitoring. Equally, a low score must not prevent clinically indicated investigations.

The aim is not simply to select fewer patients for testing. It is to use available information to make better decisions about who needs assessment, what type of assessment is appropriate, and how the results will influence care.

From an elevated AI score to appropriate rhythm assessment

When an AI system identifies elevated future AF risk, the physician should interpret the result alongside the patient's clinical circumstances.

Symptoms, medical history, cardiovascular risk factors, previous ECG findings, and current examination results may all influence whether additional assessment is appropriate.

For an asymptomatic patient, an elevated risk score may support consideration of further rhythm evaluation. However, prolonged monitoring should not become an automatic response to every high-risk prediction, particularly when the clinical benefit of such monitoring remains uncertain.

A different approach may be required when the patient presents with symptoms suggesting an arrhythmia. Palpitations, unexplained syncope, or concerning breathlessness may justify diagnostic investigation regardless of the AI score.

This distinction between screening and diagnostic evaluation should be clear within the hospital's clinical pathway.

Screening generally concerns individuals without a known diagnosis who are assessed according to a defined program. Diagnostic evaluation responds to symptoms, examination findings, or other clinical evidence suggesting that a disorder may be present.

A model developed to prioritize screening should not be used to restrict access to investigations that are otherwise clinically indicated.

The pathway should also allow physicians to make decisions that differ from the algorithm's classification when supported by the patient's circumstances. Clinical judgment should remain available rather than being replaced by a fixed risk threshold.

ECG confirmation remains essential

When AF is suspected, appropriate electrocardiographic evidence and professional interpretation remain central to establishing the diagnosis.

The 2024 ESC guidelines recommend ECG confirmation of clinical AF before proceeding with AF-specific risk stratification and management. A standard 12-lead ECG can establish the diagnosis when it demonstrates AF. Suitable single-lead or multiple-lead recordings may also provide diagnostic evidence [1].

The minimum duration required for diagnosis on monitoring devices is less straightforward than a single universal cutoff suggests. A recording of approximately 30 seconds or longer has commonly been used as a diagnostic convention for single-lead or multiple-lead ECG devices, although the ESC guidance acknowledges limitations in the supporting evidence.

For suspected AF identified through screening, the ESC guidelines recommend appropriate ECG documentation interpreted by a physician with expertise in ECG rhythm assessment.

Recording duration alone does not establish diagnostic reliability. Signal quality, rhythm characteristics, artifacts, and the clinical context remain important.

The 2023 ACC/AHA/ACCP/HRS guideline, published in Circulation in 2024, recommends that an initial AF diagnosis in a person without known AF be established by a clinician through visual interpretation of the electrocardiographic signal, regardless of the type of recording device [2].

This safeguard is particularly important when software generates an automated rhythm label.

For example, frequent premature atrial contractions may create an irregular rhythm that is incorrectly classified as AF. Reviewing the underlying ECG can prevent an inaccurate diagnosis and the unnecessary treatment decisions that might follow.

Devices using photoplethysmography may also identify irregular pulse patterns, but such findings alone do not provide the ECG evidence required to establish clinical AF.

For predictive models analyzing sinus-rhythm ECGs, the distinction is even more important. An elevated future-risk estimate may prompt consideration of further assessment, but it does not prove that AF was present during the original recording or at the time the result was generated.

Healthcare organizations should preserve this distinction in clinical documentation, patient communication, and the interpretation of automated results.

Choosing monitoring and follow-up without unnecessary testing

A normal ECG does not necessarily exclude intermittent AF. If an arrhythmia occurs only occasionally, a brief recording during a routine consultation may show sinus rhythm.

When clinical suspicion remains, ambulatory ECG monitoring or another suitable method of rhythm assessment may be considered.

The selection of a monitoring method and its duration should reflect the patient's symptoms, suspected frequency of episodes, clinical risk, and the information needed to guide subsequent care.

Longer monitoring may identify episodes missed by a single ECG. Nevertheless, this does not mean that every patient with an elevated AI risk score needs prolonged monitoring.

Additional investigations can create financial burdens, inconvenience, uncertain incidental findings, and greater demand for specialist interpretation. These consequences become particularly relevant when a healthcare organization has limited ambulatory monitoring capacity.

The significance of a negative result also depends on the investigation performed.

A normal ECG provides information about the rhythm during the recording. Even prolonged monitoring cannot guarantee that AF has never occurred or will not develop later.

For a patient with persistent concerning symptoms, further evaluation may be appropriate despite an initially negative recording. For an asymptomatic patient whose clinical assessment is otherwise reassuring, immediate additional testing may not be necessary.

An earlier high-risk prediction should not automatically place someone into an indefinite cycle of repeated investigations.

The follow-up plan should reflect the patient's clinical circumstances, findings from previous assessments, any new symptoms, and the expected value of further testing.

It is also important to communicate negative findings clearly. Patients should understand what the examination assessed, what the result means, and when further medical attention may be needed.

A reliable follow-up process should help patients receive appropriate care without creating the impression that an AI risk classification represents a permanent diagnosis.

AF prediction should not independently determine anticoagulation

AF is associated with thromboembolic risk, and oral anticoagulation is an important component of stroke prevention in appropriately selected patients.

However, an elevated AI-predicted probability of future AF is not equivalent to confirmed AF and does not independently establish an indication for anticoagulant treatment for AF-related stroke prevention.

A patient with sinus rhythm and a high AI risk score should not automatically receive an AF diagnosis or anticoagulant prescription solely because of that prediction.

When AF is confirmed, clinicians must separately assess stroke risk, bleeding-related considerations, relevant comorbidities, patient preferences, and other factors according to applicable professional guidance.

The 2024 ESC AF-CARE framework emphasizes management of comorbidities and risk factors, prevention of stroke and thromboembolism, rate and rhythm control, and continuing evaluation [1].

The ACC/AHA/ACCP/HRS guideline similarly supports individualized, risk-based management [2].

Additional distinctions may be necessary when short atrial high-rate episodes are detected by implanted devices in people without previously documented clinical AF. These findings should not automatically be treated as equivalent to every form of ECG-confirmed AF, and their management requires appropriate interpretation.

The separation between prediction, diagnosis, and treatment is particularly important because anticoagulants can cause clinically significant bleeding.

An AI risk score may provide a reason to consider further investigation. It should not independently determine whether anticoagulation is started for AF-related stroke prevention, which medication is prescribed, or how long treatment continues.

Explaining AI results without creating unnecessary anxiety

How an AI result is communicated can influence the patient's understanding of their condition and willingness to undergo further assessment.

A label such as "high risk" may sound definitive, particularly when accompanied by a numerical score. Some patients may interpret it as confirmation that they already have a serious heart disorder. Others may assume that a low-risk score means AF cannot develop.

Neither interpretation is appropriate.

Clinicians should explain the result according to the model's actual purpose. If the software estimates the likelihood of AF being diagnosed during the next two years, the discussion should make clear that this is a prediction over a defined period, not a statement that AF is currently present.

The distinction between relative and absolute risk may also require explanation. Someone classified as higher-risk than other patients may still have a relatively low absolute probability of receiving an AF diagnosis.

Where numerical probabilities are provided, they should be interpreted in light of the model's calibration and applicability to the patient population.

Patients should understand why additional monitoring may be considered and what the possible results could mean. The discussion should also address the burden of testing, including further appointments, possible costs, and the possibility that no arrhythmia will be identified.

Some patients may prefer additional evaluation, while others may be concerned about monitoring or its practical consequences.

Appropriate patient participation helps ensure that decisions are both clinically justified and realistic.

The purpose of communication is not to convince every patient with an elevated score to undergo further testing. It is to support an informed decision based on the patient's circumstances and the available evidence.

Emerging evidence from Viet Nam

Research conducted in Viet Nam provides useful context for the development of AI-assisted rhythm assessment, although different technologies must be evaluated according to their specific functions.

A study published in Digital Health in June 2026 described the development of an AI-supported ECG monitoring system using data from Nguyen Trai Hospital in Ho Chi Minh City [6].

The investigators analyzed 1,489 Holter ECG recordings and developed a deep learning model for AF detection. The model was trained using 1,089 recordings and evaluated on a separate held-out dataset of 400 recordings.

In that evaluation dataset, which contained 25 AF-positive recordings, the model achieved reported case-level sensitivity of 100%, specificity of 88%, and positive predictive value of 35.7%.

These figures require interpretation in context. The reported sensitivity was obtained from a relatively limited number of AF-positive recordings, and the positive predictive value indicates that many positive classifications required further review. The evaluation was conducted within the study's data environment rather than through independent external validation across multiple healthcare organizations.

A prospective pilot involving 167 higher-risk patients examined a workflow combining AI analysis with wearable ECG monitoring. AF was detected in 24 participants, including 20 whose AF was identified after the first 24 hours of monitoring.

The study suggested that integrating AI analysis with wearable ECG technology and professional review was feasible in the participating setting.

However, its findings should not be interpreted as evidence validating AI prediction of future AF from an earlier sinus-rhythm ECG.

The Vietnamese study focused on detecting AF within rhythm recordings, whereas the 2026 VITAL-AF analysis examined prediction of subsequently diagnosed AF using historical ECG and clinical information.

These are different clinical applications.

The Vietnamese study also does not establish that introducing the proposed monitoring system into routine care reduces stroke, hospitalization, or mortality. Its relatively limited evaluation population and the need for broader independent assessment remain important considerations.

Nevertheless, the research provides an example of how AI-assisted cardiovascular monitoring is being explored in Viet Nam and why local workflow evaluation matters.

For hospitals considering similar technologies, the relevant question is whether the evidence supports the specific application, patient population, equipment, and clinical decisions proposed for their own organization.

Data quality and local validation in Vietnamese healthcare settings

The reliability of ECG-based AI depends on more than the predictive algorithm.

Lead placement, recording artifacts, electrical noise, sampling frequency, equipment characteristics, and preprocessing methods may influence the signals supplied to the model.

A system developed using ECG recordings from particular devices or acquisition settings may not perform identically when introduced into a different environment.

Before routine use, healthcare organizations should understand the supported ECG formats, recording conditions, and technical limitations of the proposed software.

The system should also have an appropriate process for handling incomplete, invalid, or unsupported recordings. A technically unsuitable ECG should not silently generate a reassuring risk classification.

Patient identification is another important consideration, particularly when historical ECGs are retrieved from electronic medical records or transferred between systems.

The original recording, examination date, patient identity, and generated AI result should remain reliably linked. Even a technically accurate prediction can be unsafe if it is associated with the wrong patient.

The timing of the ECG also matters. A model designed to analyze recordings obtained within a particular period should not automatically be applied to substantially older recordings without considering whether that use is supported by evidence.

Local evaluation should examine the model's performance in the intended patient population, including discrimination, calibration, and clinically relevant differences across patient groups.

Differences in age, comorbidities, access to follow-up, and patterns of ECG use may influence performance. A model predicting documented AF may behave differently in a setting where patients receive fewer opportunities for rhythm monitoring.

Hospitals should also consider whether applying AI only to patients with available historical ECGs creates unintended selection effects.

A model can perform acceptably among patients with complete data while leaving out others who may benefit from clinical assessment.

Retrospective evaluation requires particular care. Information used to generate a prediction should be limited to what would actually have been available at the intended clinical decision point. Otherwise, apparent performance may be inflated by information that would not have been accessible in real practice.

These considerations support safer implementation, but they should not be presented as a universally mandated Vietnamese validation procedure or a standardized national threshold for AF prediction software.

Establishing a workable clinical pathway

Before introducing AI-assisted AF risk stratification, a hospital should understand the problem it intends to address.

One organization may wish to improve recognition of AF among older adults who already undergo routine ECG examinations. Another may have limited access to ambulatory monitoring and want to use those resources more selectively.

A different hospital may find that its main problem is not detecting potential AF but ensuring that patients complete recommended investigations and receive their results.

These situations require different responses.

If delayed ECG interpretation is the main concern, improving existing reporting and review processes may be more useful than introducing a predictive model.

If patients frequently fail to complete clinically indicated monitoring, strengthening appointment coordination and follow-up communication may provide greater immediate benefit.

Where AI is selected, its intended use should be clearly defined. The hospital needs to establish which patients are eligible, which ECG recordings can be analyzed, who reviews the prediction, and what additional assessment may follow.

A high score should not automatically generate an AF diagnosis, medication order, or mandatory specialist referral.

Instead, the result should be available to the responsible clinician as additional information when determining whether further rhythm assessment is appropriate.

The organization should establish who arranges monitoring, who interprets recordings, who communicates results, and who is responsible for subsequent clinical decisions.

Cardiology or electrophysiology consultation may be appropriate when the diagnosis or management is complex. However, an elevated AI prediction alone should not automatically require referral for every patient.

The pathway should also function when AI is unavailable, generates an invalid result, or is unsuitable for a particular clinical situation.

The aim is to integrate useful predictive information into established cardiovascular care without creating a separate process in which an algorithm effectively determines patient management.

Vietnamese legal and professional responsibilities

Introducing ECG-based AI into clinical practice requires consideration of the applicable Vietnamese legal framework.

Clinical evidence supporting an AI model does not automatically establish that a particular product may be deployed for every proposed purpose in Viet Nam.

Regulatory requirements depend on the software's actual functions, intended medical purpose, the arrangements under which it is supplied and used, the information being processed, and the clinical decisions its outputs may influence.

Medical examination, treatment, and professional accountability

Viet Nam's Law on Medical Examination and Treatment No. 15/2023/QH15, effective January 1, 2024, as amended, provides an important legal foundation for healthcare organizations, professional practice, patient rights, and clinical services [7].

Healthcare organizations considering AI-assisted AF risk assessment should ensure that its proposed use is consistent with their authorized activities, the professional scope of the practitioners involved, and applicable technical requirements.

An AI-generated prediction does not independently establish a medical diagnosis or expand a practitioner's legally permitted professional scope.

Decisions about ECG interpretation, further investigation, diagnosis, and treatment remain subject to applicable professional and legal responsibilities.

International clinical guidelines, including ESC 2024 and ACC/AHA/ACCP/HRS 2023, provide evidence-based professional references. However, they should not automatically be treated as legally binding protocols that override Vietnamese requirements.

Any relevant Ministry of Health technical procedures and professional obligations should be assessed separately.

Artificial intelligence regulation

Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on March 1, 2026 [8].

Decree No. 142/2026/NĐ-CP, effective May 1, 2026, provides implementing provisions concerning AI activities, including risk classification and management [9].

Decision No. 33/2026/QĐ-TTg, effective August 15, 2026, establishes a list of high-risk AI systems [10].

The healthcare entries expressly identified in that list concern certain AI-assisted surgical and robotic systems. The list should not be interpreted as automatically assigning every ECG-based prediction or rhythm-analysis application to the same legal category.

Equally, the absence of a specific ECG-AI application from the named list does not, by itself, establish that the system is legally low-risk or exempt from other applicable requirements.

The relevant classification and obligations must be assessed in relation to the actual system, its intended functions, the way it is deployed, and the applicable legal criteria.

A model estimating future AF risk may raise different considerations from software intended to identify an existing arrhythmia or directly influence treatment decisions.

It is also important to distinguish statutory AI risk classification from clinical safety risk. A system may contribute to patient harm if its output leads to missed diagnoses, unnecessary investigations, incorrect clinical records, or inappropriate treatment, regardless of its formal regulatory classification.

Healthcare organizations should therefore consider legal compliance alongside the potential clinical consequences of using the technology.

Medical device regulation

AI software used to analyze ECG information may fall within Vietnamese medical device regulation depending on its intended medical purpose, functions, and applicable regulatory definitions.

Decree No. 98/2021/NĐ-CP on medical device management, as amended by subsequent regulations including Decrees No. 07/2023/NĐ-CP and No. 04/2025/NĐ-CP, forms part of the relevant framework [11].

Circular No. 24/2026/TT-BYT, effective July 1, 2026, further addresses the determination of risk levels and management measures for products classified as medical devices [12].

Hospitals should determine the applicable regulatory status of the specific AI software, ECG equipment, and any integrated functions before routine clinical deployment.

The established regulatory status of an ECG machine does not automatically confirm that newly connected AI software may be used for every proposed medical purpose.

Similarly, international clinical studies or foreign regulatory authorization do not independently establish compliance with all Vietnamese requirements.

The assessment should reflect the actual product and its intended clinical use rather than rely on a general description such as "AI-assisted cardiology" or "AI-enabled ECG analysis."

Personal data protection and secondary data use

ECG recordings and AI-generated risk estimates may contain or be linked to identifiable patient information, including age, examination dates, diagnoses, medical history, and other clinical records.

Viet Nam's Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP took effect on January 1, 2026 [13,14].

Healthcare organizations should assess the applicable grounds and purposes for processing patient information, consent requirements and lawful exceptions where relevant, access controls, information security measures, retention arrangements, and responsibilities of technology providers.

Particular attention is necessary when historical ECG recordings are transferred to external AI platforms or processed using cloud infrastructure.

Hospitals should understand where patient information is processed and stored, who can access it, and whether it is used only for the agreed clinical purpose.

Where personal data are transferred outside Viet Nam, the organization should separately assess the applicable cross-border data transfer requirements and any relevant exceptions or procedural obligations.

Secondary use of ECG information for model development, algorithm training, commercial product improvement, or research requires separate consideration of the relevant legal and institutional requirements.

Removing patient names or identification numbers does not necessarily make ECG data fully anonymous. Associated clinical information and metadata may still permit re-identification.

Where the technology is evaluated through clinical research, relevant research ethics requirements and participant protections should also be considered.

These matters should form part of implementation planning rather than be addressed only after an incident occurs.

Measuring whether AI-assisted screening improves care

A healthcare organization should determine how it intends to evaluate an AI-assisted AF screening program before introducing the technology into routine practice.

Predictive performance is an important consideration, but model discrimination and calibration alone do not establish clinical effectiveness.

The hospital should examine whether the technology changes appropriate clinical actions. Relevant information includes the proportion of eligible patients assessed, the number recommended for additional rhythm evaluation, and whether those investigations are completed.

The number of patients with ECG-confirmed AF should be distinguished from the number classified as high-risk.

The proportion of patients undergoing monitoring without an AF diagnosis is also relevant, although a negative result does not necessarily mean that monitoring was inappropriate. The clinical justification for the examination matters more than whether it produced a positive finding.

Follow-up outcomes deserve particular attention. A program may generate useful risk estimates yet provide little benefit if patients do not complete indicated assessments or clinicians do not receive and act on the results.

For patients with confirmed AF, evaluation should consider whether appropriate clinical review occurred and whether subsequent management decisions were justified.

Anticoagulation rates should not be used as an isolated measure of success. Some patients may not have an indication for anticoagulation, while others may have contraindications or circumstances requiring individualized decisions.

Hospitals should also examine possible unintended consequences, including false diagnoses, unnecessary referrals, repeated investigations without clear clinical indications, inappropriate reassurance from low AI scores, and treatment decisions made without adequate diagnostic evidence.

Operational effects matter as well. Additional monitoring may increase demand for equipment, ECG interpretation, specialist consultations, and follow-up appointments.

If these services become congested, patients with stronger clinical indications may experience delays.

The program's performance should continue to be evaluated after deployment. Changes in ECG equipment, software versions, recording conditions, patient populations, or patterns of clinical practice may influence both model outputs and the effectiveness of the care pathway.

An increase in detected AF does not necessarily mean that the algorithm has become more accurate. It may reflect changes in patient selection or monitoring intensity. Conversely, a decline in diagnoses may result from incomplete follow-up rather than improved risk prediction.

Healthcare professionals should have a practical way to report concerns when AI outputs appear unreliable or when the resulting workflow creates patient-safety problems.

The appropriate response may involve correcting input data, adjusting clinical procedures, providing additional training, reviewing the software, or suspending a particular application when safety cannot be adequately maintained.

Stroke, major bleeding, hospitalization, and mortality remain clinically important outcomes. However, demonstrating that an AI-guided screening program improves these outcomes requires suitable research designs, adequate follow-up, and careful assessment of other influencing factors.

Routine operational monitoring should not be presented as proof of a causal reduction in adverse clinical outcomes.

The purpose of evaluation is to determine whether AI contributes useful information, supports appropriate assessment, and improves the reliability of care without avoidable harm.

These are proposed organizational evaluation considerations, not a mandatory Vietnamese AI screening scorecard or an independently validated Digital Medicine Vietnam measurement instrument.

A practical direction for healthcare organizations in Viet Nam

For hospitals and clinics in Viet Nam, responsible implementation should begin with a clearly defined clinical need.

Before selecting ECG-based AI software, the organization should understand how AF is currently identified, how ECG recordings are interpreted, which patients receive additional monitoring, and what happens after the results become available.

Clinical leaders from relevant services should participate in defining the intended use and evaluating whether the available evidence supports it.

The proposed technology should be assessed in relation to the hospital's actual ECG equipment, patient population, clinical workflow, regulatory requirements, and capacity to provide follow-up.

Where appropriate and legally permissible, supervised local evaluation may help establish whether the software performs reliably and whether its outputs contribute useful information to clinical decisions.

The organization should also recognize when improvements to existing care processes may provide greater benefit than introducing another predictive system.

A hospital with inadequate access to ambulatory monitoring will not necessarily solve that problem by identifying more people at elevated AF risk. Similarly, a service with inconsistent result communication may need stronger follow-up arrangements before expanding screening activity.

AI should support a functioning clinical pathway rather than create additional demand that the organization cannot safely manage.

From AF prediction to appropriate patient care

ECG-based artificial intelligence offers new opportunities to identify individuals who may have an increased likelihood of atrial fibrillation being diagnosed in the future.

The 2026 VITAL-AF secondary analysis provides encouraging evidence that AI-derived ECG information can improve risk discrimination compared with a traditional clinical model. Research conducted in Viet Nam also illustrates growing interest in using AI to support rhythm monitoring and AF detection.

These developments are promising, but they do not establish that every AI-guided screening approach improves clinical outcomes.

For healthcare organizations, the essential distinction remains between predicting risk, confirming a rhythm disorder, and determining appropriate care.

An elevated AI score may support consideration of additional assessment. A suspected arrhythmia requires appropriate electrocardiographic evidence and professional interpretation. When AF is confirmed, treatment decisions require individualized clinical evaluation.

The value of AI-assisted screening ultimately depends on whether these activities are connected through a reliable clinical pathway.

A hospital should not judge success solely by the number of risk scores generated, the number of ECGs performed, or the number of people selected for monitoring.

It should consider whether the technology helps clinicians identify appropriate patients for assessment, supports accurate diagnosis, and contributes to better decisions without unnecessary investigations or avoidable harm.

The most useful AI prediction is not necessarily the one that identifies the largest number of high-risk patients. It is the one that helps healthcare professionals make better decisions about what those patients need next.

Professional Education Notice

This article is provided by Digital Medicine Vietnam (DMVN) for professional education, evidence interpretation, and knowledge exchange. It does not constitute individualized medical, cardiology, electrophysiology, pharmaceutical, legal, regulatory, or patient-safety advice.

The opening clinical scenario is illustrative and does not describe a documented patient case.

The discussion focuses primarily on ECG-based AI applications estimating the likelihood of future AF diagnosis and the subsequent processes of rhythm assessment, diagnostic confirmation, and clinical follow-up. It does not establish the safety or effectiveness of any particular commercial AI product, autonomous AF diagnosis system, anticoagulant prescribing application, or automated rhythm treatment technology.

Different ECG-based AI applications have different intended purposes, target populations, data requirements, and clinical outputs. Evidence supporting one application should not automatically be generalized to another.

The referenced international clinical guidelines provide professional recommendations and should not be interpreted as automatically superseding applicable Vietnamese laws, regulations, or professional requirements. The legal and regulatory status of individual AI systems requires separate assessment based on their actual intended use and deployment arrangements.

The VITAL-AF findings and other cited studies should be interpreted according to their respective designs, populations, and limitations. Improved prediction, increased AF detection, or more efficient screening does not independently establish reduced stroke, mortality, or other adverse patient outcomes.

The organizational approaches discussed are proposed considerations for responsible healthcare AI implementation and quality improvement. They are not official Vietnamese Ministry of Health AF screening protocols, individual legal compliance determinations, or independently validated DMVN clinical assessment instruments.

References

1. Van Gelder IC, Rienstra M, Bunting KV, et al. 2024 ESC Guidelines for the Management of Atrial Fibrillation Developed in Collaboration With the European Association for Cardio-Thoracic Surgery (EACTS). European Heart Journal. 2024;45(36):3314–3414. DOI: 10.1093/eurheartj/ehae176.

2. Joglar JA, Chung MK, Armbruster AL, et al. 2023 ACC/AHA/ACCP/HRS Guideline for the Diagnosis and Management of Atrial Fibrillation: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation. 2024;149(1)–e156. DOI: 10.1161/CIR.0000000000001193.

3. Vedage NA, Friedman SF, Chang Y, et al. Risk-Guided Atrial Fibrillation Screening With Artificial Intelligence–Enabled Electrocardiogram Models: A VITAL-AF Trial Analysis. Journal of the American College of Cardiology. 2026;87(14):1798–1813. DOI: 10.1016/j.jacc.2026.01.087.

4. Lubitz SA, Atlas SJ, Ashburner JM, et al. Screening for Atrial Fibrillation in Older Adults at Primary Care Visits: VITAL-AF Randomized Controlled Trial. Circulation. 2022;145(13):946–954. DOI: 10.1161/CIRCULATIONAHA.121.057014.

5. US Preventive Services Task Force. Screening for Atrial Fibrillation: US Preventive Services Task Force Recommendation Statement. JAMA. 2022;327(4):360–367. DOI: 10.1001/jama.2021.23732.

6. Nguyen SV, Ho MK, Nguyen DV, et al. Development of a Semi-Real-Time Electrocardiogram Monitoring System Integrating Artificial Intelligence and Wearable Devices for Atrial Fibrillation Screening. Digital Health. 2026;12:20552076261458176. DOI: 10.1177/20552076261458176.

7. National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment. Adopted January 9, 2023; effective January 1, 2024, as amended. Consolidated text No. 26/VBHN-VPQH dated February 26, 2026.

8. National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted December 10, 2025; effective March 1, 2026.

9. Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing Certain Provisions and Measures for Implementing the Law on Artificial Intelligence. Issued April 30, 2026; effective May 1, 2026.

10. Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. Issued June 30, 2026; effective August 15, 2026.

11. Government of Viet Nam. Decree No. 98/2021/NĐ-CP on Medical Device Management, as amended, including by Decrees No. 07/2023/NĐ-CP and No. 04/2025/NĐ-CP.

12. Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Risk Determination and Management Measures for Products Classified as Medical Devices and Amendments to Circular No. 05/2022/TT-BYT. Issued June 30, 2026; effective July 1, 2026.

13. National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted June 26, 2025; effective January 1, 2026.

14. Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing Certain Provisions and Measures for Implementing the Law on Personal Data Protection. Issued December 31, 2025; effective January 1, 2026.