Clinical AI Evaluation and Research Ethics in Viet Nam: Distinguishing Routine Monitoring, Quality Improvement and Research Governance
EVIDENCE & IMPLEMENTATION READINESS
Clinical AI Evaluation and Research Ethics in Viet Nam: Distinguishing Routine Monitoring, Quality Improvement and Research Governance
EVIDENCE & EVALUATION
Last reviewed: 6 October 2026
Once clinical AI is deployed, healthcare organizations need to evaluate what happens in real-world use. They may monitor drift, review incidents, compare outcomes, examine override rates, assess subgroup performance, test a revised workflow or analyze whether an AI-assisted pathway actually improves care. These activities are essential to responsible implementation, but they raise an increasingly important question as organizations move toward continuous learning: when is an AI evaluation part of routine monitoring or quality improvement, and when has it become research requiring formal research governance and ethics oversight?
The distinction matters because the same clinical data can be used for very different purposes. A hospital may review deterioration alerts each month to make sure a deployed system remains safe. A quality team may change an escalation workflow and measure whether response time improves. A research group may use the same clinical data to test a predefined hypothesis about whether AI-assisted care changes ICU transfers. The datasets, statistical methods and clinicians involved may overlap, but the activities do not necessarily have the same governance status.
AI makes the boundary less obvious. Continuous monitoring can produce findings worth publishing. Quality-improvement projects can use sophisticated statistical methods. Research can be embedded directly in clinical workflows. A Learning Health System deliberately shortens the cycle between care, evidence and improvement. None of these characteristics, by itself, determines whether an activity is formal research.
For healthcare organizations in Viet Nam, the practical question should therefore not be “Are we calling this monitoring, quality improvement or research?” It should be “What are we actually doing, why are we doing it, what happens to people and their data, and what governance does this activity require?”
The label does not decide the governance pathway
Calling an activity a pilot, local validation, quality-improvement project, real-world evaluation, service evaluation, audit or Learning Health System initiative does not, by itself, determine its legal or ethical status. Conversely, sophisticated statistical analysis or an intention to publish findings does not automatically convert every quality-improvement activity into biomedical research.
The substance of the activity matters more than the label. A useful assessment should consider the primary purpose; whether care is deliberately being changed to generate knowledge; whether people are assigned to different interventions or workflows; whether additional data, procedures or interactions are introduced specifically for the evaluation; whether the activity is prospectively structured around a research question; whether patients or staff face additional risk or burden; whether identifiable health data are being reused, linked or transferred; and whether the project is mainly designed to improve an immediate local system or to systematically answer a broader scientific question.
These are governance indicators, not a statutory checklist. As of this review, the Vietnamese framework does not provide one universal operational test that cleanly separates every clinical-AI monitoring activity, QI project and research study. Where the boundary is uncertain, healthcare organizations should seek an appropriate institutional research-governance determination, and where relevant Ethics Committee input, rather than relying solely on the project team's preferred description.
Routine monitoring begins with operational responsibility
Routine monitoring is part of normal governance when a healthcare organization is checking whether a deployed AI system continues to operate safely and as intended. This may include monitoring model performance, alert volume, system availability, response times, missing data, clinician overrides, complaints, incidents, subgroup signals or changes following a software update.
The primary purpose is usually operational: identify problems, maintain safe performance and determine whether the organization needs to investigate, correct, restrict or suspend the system.
For high-risk AI systems, this lifecycle approach is already reflected in Viet Nam's current AI framework. Decree No. 142/2026/NĐ-CP, effective since 1 May 2026, requires deployers to organize risk management during deployment and operation, monitor for deviations, risks or incidents, maintain appropriate human supervision and intervention, and apply risk controls within their authority [6].
This type of monitoring does not become research merely because the organization analyzes data, calculates performance indicators or identifies a clinically interesting pattern. Healthcare organizations need to be able to observe technologies on which care depends.
Routine monitoring can, however, evolve into a different kind of activity. Suppose monitoring shows that an AI deterioration system behaves differently across wards. If the organization subsequently designs a prospective comparative study, deliberately applies different thresholds across units and collects additional patient-level variables to test a predefined hypothesis, the nature of the activity has changed. Its origins in routine monitoring do not determine the governance requirements of what comes next.
Quality improvement asks how local care can be made better
Quality improvement, or QI, usually begins with a recognized problem in current care and an intention to improve the local system. A hospital may identify slow responses to AI alerts, redesign the escalation pathway, provide training and measure response times before and after the intervention. Another organization may identify recurrent errors in AI-generated documentation, change its review process and examine whether those errors decline.
These activities can be systematic, data-intensive and methodologically rigorous while remaining fundamentally directed toward improving local care.
Viet Nam already has a hospital quality-management framework that supports this type of work. Circular No. 19/2013/TT-BYT remains recorded as in force and establishes principles including patient-centeredness, continuous quality improvement, use of evidence in quality decisions and staff participation in quality-management activities [7].
Clinical AI should increasingly become part of this existing quality architecture rather than sit in a separate universe of technical model metrics.
At the same time, calling something quality improvement should not be treated as an automatic exemption from research-governance consideration. A project may begin with a legitimate quality problem while using a prospective design, deliberate allocation, additional procedures or data collection, external investigators or other features that justify a formal research-governance determination. The organization should examine what the project actually does rather than assuming that its local-improvement objective settles the question.
Research asks a different kind of question
Research generally involves systematic investigation designed to generate knowledge rather than simply manage an immediate operational problem. In clinical AI, this may include developing or externally validating a prediction model, prospectively comparing AI-assisted and non-AI pathways, studying different AI thresholds, evaluating a new AI intervention before routine deployment, analyzing clinical datasets to test predefined hypotheses or studying how healthcare professionals interact with an AI system.
WHO's July 2026 report Artificial Intelligence-related Health Research: Ethics Review and Oversight distinguishes three broad categories of AI-related health research: health-related data science using AI; research conducted with AI tools and technologies; and health-related research on AI tools and technologies [2]. WHO also highlights challenges involving transparency, bias, fairness, accountability, privacy and potential harms that existing ethics-oversight systems may not always be prepared to address.
These categories are useful for understanding the breadth of AI-related health research but are not Vietnamese statutory classifications.
For example, using machine learning to analyze retrospective clinical data for a scientific investigation may fall within health-related data science using AI. Researchers using generative AI to perform or support aspects of a study may fall within research conducted with AI tools and technologies. A prospective study examining the performance, safety or clinical effect of a new AI diagnostic system may constitute health-related research on an AI tool or technology.
Viet Nam has a current ethics-committee framework
Circular No. 43/2024/TT-BYT has been effective since 1 February 2025 and remains in force. It replaced Circular No. 04/2020/TT-BYT and establishes the current framework for the establishment, organization and operation of national and institutional Ethics Committees in Biomedical Research [1].
Under Article 5, an Ethics Committee reviews the ethical and scientific aspects of biomedical research involving human subjects as a basis for competent authorities to approve, inspect, evaluate and accept research. The Circular provides for national and institutional Ethics Committees. Where an organization does not establish its own institutional Ethics Committee, biomedical research involving human subjects may be reviewed by an appropriate institutional Ethics Committee in the relevant field [1].
This matters for AI because research ethics is not relevant only when a study involves an experimental drug, implanted device or invasive procedure. AI-related biomedical research may involve human subjects through interventions, clinical decisions, observation, interaction, identifiable health information or changes to care pathways.
The appropriate institutional question is therefore not simply “Is this an AI study?” It is whether the specific activity falls within biomedical research involving human subjects and what research-governance and ethics oversight are applicable.
Minimal risk does not mean no ethics review
Many AI studies use existing records and may initially appear low-risk. Circular No. 43/2024/TT-BYT defines minimal risk as research in which the probability and magnitude of physical, psychological or social harm, discomfort or other adverse effects are no greater than those ordinarily encountered in daily life or in routine examinations or tests [1].
Article 16 permits expedited review for minimal-risk research and several other specified categories, including certain previously reviewed submissions, protocol modifications, periodic or ad hoc reports, updated study documents, adverse-event reports and protocol-deviation reports [1].
The practical implication is important: minimal risk may affect how a study is reviewed; it should not automatically be interpreted as meaning that research governance is unnecessary.
A retrospective AI study using health records may involve little physical risk while still raising issues of privacy, re-identification, subgroup bias, data linkage, secondary use or commercial access. Risk in AI-related research therefore needs to be understood more broadly than physical injury alone.
Local validation can sit on either side of the boundary
Local validation illustrates why labels are insufficient. Suppose a hospital receives an AI imaging system and, before routine deployment, applies it to historical data to determine whether performance appears acceptable in its own environment. Depending on the purpose, design, data and institutional arrangements, this may be part of technical or organizational evaluation.
Now consider a project in which investigators define hypotheses in advance, calculate a sample size for scientific inference, link imaging findings with outcomes, compare predefined subgroups, collaborate with an external university and structure the project to generate new scientific evidence about performance in Vietnamese patients. Both activities may be called local validation, but their governance profiles are very different.
A third scenario may involve prospective silent deployment: AI generates outputs that do not influence current treatment, while investigators systematically compare those outputs with later clinical outcomes. The absence of an effect on current care may reduce some risks, but it does not by itself establish that the activity is not research.
The relevant question is therefore not only whether AI influences treatment, but why and how the evaluation has been designed and how people and clinical data are involved.
Publication intent is a signal, not a deciding test
An activity does not become research solely because somebody intends to publish the findings. Quality-improvement teams can legitimately publish lessons from improvement work, while a formal research project remains research even if publication ultimately never occurs.
Publication intention can nevertheless be one useful governance signal because it may indicate that an activity is intended to generate knowledge beyond the immediate operational problem.
The more important question is what the activity was designed to do when it began. Was it primarily structured to solve an immediate local problem using routine improvement methods, or was it designed as a systematic investigation around a scientific question? Were comparators, endpoints, sampling and analysis specified to generate scientific evidence? Were people exposed to procedures or allocations introduced for the study?
Those characteristics generally tell the organization more than the eventual existence of a manuscript.
Randomization or deliberate variation deserves particular scrutiny
A particularly important governance signal arises when an organization deliberately creates different conditions in order to compare their effects. Examples include assigning wards to different AI alert thresholds, randomizing clinicians to receive or not receive AI recommendations, or deliberately varying patient-facing AI communications across groups to compare outcomes.
These designs may be scientifically valuable and may even fit within a Learning Health System, but embedding experimentation in routine care does not remove research-ethics considerations.
Where different conditions are deliberately created for evidence generation, healthcare organizations should seek an early research-governance determination. Relevant questions include who bears the risks of the different conditions, whether patients or healthcare workers become research participants, what informed-consent or other participant protections may apply and what independent oversight is appropriate.
Additional procedures and additional data matter
Another useful governance signal is whether an evaluation asks people to undergo something that would not otherwise occur during ordinary care or routine operations. This may include additional imaging, laboratory tests, questionnaires, interviews, recordings, observations, follow-up contacts or collection of additional data solely for the evaluation. It can also include linking clinical records with external datasets in ways not required for ordinary care.
The greater the departure from routine care or routine quality-management activity, the stronger the reason for an early research-governance determination.
The reverse should not be assumed. Using only pre-existing data does not automatically make an activity non-research. Many biomedical research projects are retrospective and introduce no new clinical intervention.
Purpose, design, participant involvement, data use and risk need to be considered together.
Data protection and research ethics are separate governance layers
Research-ethics approval does not replace compliance with data-protection law, and compliance with data-protection law does not replace research-ethics review.
Viet Nam's Law No. 91/2025/QH15 on Personal Data Protection and Decree No. 356/2025/NĐ-CP have both been effective since 1 January 2026 [3,4]. Health-status information falls within the current framework for sensitive personal data. Decree No. 356 expressly recognizes that personal data may be used for research and development of self-learning algorithms, AI systems and other automated systems provided that personal-data protection requirements are complied with. It also provides that data generated through AI inference must receive personal-data protection where they can identify or help identify a particular individual [4].
An AI research project may therefore receive ethics approval while still requiring separate attention to data-access controls, processing arrangements, applicable impact assessments, contractual safeguards, security and cross-border data requirements. A QI activity may not constitute biomedical research requiring formal Ethics Committee review, yet still involve sensitive health information governed by personal-data law.
De-identification or anonymization measures can materially reduce privacy risk, but organizations should confirm what has actually been done to the dataset and whether the resulting information remains personal data under the applicable framework rather than assuming that removal of names alone makes data anonymous.
The question “Do we need ethics review?” should therefore never replace the separate question “Are we appropriately governing the personal data?”
Existing clinical data do not become ethically neutral because they already exist
Clinical AI evaluation frequently relies on information originally collected during routine care. The absence of a new blood test, scan or intervention can create the impression that ethical risk is minimal, but clinical records may contain highly sensitive information and new analytical methods can create new forms of risk.
Linking datasets can increase identifiability. AI may infer characteristics that were never explicitly entered into a record. Rare subgroups can become more identifiable even after obvious identifiers have been removed. External researchers or commercial technology companies may gain access to information that originated within the care relationship.
WHO's 2026 report highlights privacy, fairness, bias, accountability, unequal benefit and power imbalances as relevant concerns in AI-related health research, with particular attention to risks and inequities affecting low- and middle-income countries [2].
For Viet Nam, these considerations become especially important when health data or analyses involve overseas universities, multinational technology companies, cloud platforms or external research collaborators. Scientific value does not remove obligations relating to personal-data protection, institutional control, contracts, cybersecurity or applicable cross-border requirements.
Consent questions require more than one answer
Whether individual consent is required in an AI evaluation cannot be answered merely by deciding whether the project is called research. Several distinct questions may arise: consent to clinical care; consent or another applicable legal basis or condition for personal-data processing; informed consent to participate in biomedical research; consent associated with recordings or biological materials; and authorization for particular secondary uses.
These should not be treated as interchangeable.
Circular No. 43/2024/TT-BYT gives an Ethics Committee authority to approve or decline a waiver of the research-information document and voluntary participation form in specified circumstances, including where absolute confidentiality for the participant is necessary or where consent cannot be obtained from the participant or legally authorized representative, after consideration of the benefits and risks of the research and measures to protect participants' rights and safety [1].
This is a specific power under the research-ethics framework, not a general authority to waive every type of consent or authorization that may apply. Project teams should therefore not assume that retrospective data automatically remove the need for consent or, conversely, that every AI research activity necessarily requires a newly signed consent form.
The appropriate question is: what consent, authorization, waiver or other governance basis applies to this specific activity under the relevant clinical, research and personal-data frameworks?
Vendor-sponsored evaluation deserves independent scrutiny
Healthcare organizations may be invited by AI companies to provide data, validate technologies, participate in post-deployment studies or publish local performance results. These collaborations can generate valuable evidence, but commercial relationships create additional governance questions.
Who designed the protocol? Who controls the analysis? Who owns or controls access to the resulting data? Can the healthcare organization publish unfavorable findings? What happens when performance is worse than expected? Are negative findings reported? Can identifiable or pseudonymized data be accessed outside the organization? Do investigators or organizational decision-makers have financial interests related to the vendor?
Circular No. 43/2024/TT-BYT treats conflict of interest as an important ethics-governance issue, and persons with relevant conflicts should not participate in the review of the affected research in a manner inconsistent with the Circular [1].
The broader organizational principle is straightforward: commercial sponsorship does not make an evaluation inappropriate, but financial or contractual interests should not substitute for independent assessment of scientific validity, participant protection and data governance.
Multi-site evaluation creates additional governance questions
AI systems are increasingly deployed across hospital groups, specialty networks and multi-site healthcare organizations. Evaluation may therefore involve sites with different workflows, populations, equipment and information systems.
A finding at one hospital should not automatically be treated as evidence for an entire network. Likewise, ethics or institutional authorization at one site should not automatically be assumed to resolve every governance requirement at another participating organization.
Circular No. 43/2024/TT-BYT recognizes institutional Ethics Committees and provides that organizations without their own institutional committee may have biomedical research involving human subjects reviewed by an appropriate institutional Ethics Committee in the relevant field [1].
Multi-site AI research therefore needs explicit coordination around the protocol, institutional responsibilities, data processing, investigators, participant protection, safety reporting and access to results. These questions become particularly important when evidence from a limited number of sites will later be used to support system-wide deployment.
Quality improvement still requires ethical discipline
Correctly determining that an activity is QI rather than formal research does not make ethical considerations disappear.
A QI intervention may change care, increase staff burden, affect patient groups differently, involve sensitive information or produce unexpected harm. Organizations should therefore apply proportionate ethical reasoning even when formal research-ethics review is not required.
Questions may include whether the intervention is justified, whether foreseeable risks are proportionate, whether patient interests are protected, whether inequitable effects are being monitored, whether staff understand changed responsibilities and whether the organization has a mechanism to modify or stop an intervention when harm emerges.
The distinction between QI and research helps determine part of the governance pathway. It should not determine whether the organization thinks ethically at all.
Routine monitoring can generate research questions
The relationship among monitoring, QI and research is dynamic. Routine AI monitoring may reveal an unexpected subgroup difference. That observation may lead to a QI intervention, and the improvement activity may subsequently generate a scientific question worth studying formally.
There is nothing inherently problematic about this progression. The governance problem occurs when the purpose and design of the activity change without the organization recognizing that its governance requirements may also need to change.
Healthcare organizations should therefore establish transition points. When routine monitoring evolves into a predefined scientific hypothesis, when new interventions or additional data collection are introduced, when external investigators become involved or when different care strategies begin to be compared prospectively, the governance pathway should be reconsidered.
The status of an activity should be reassessed when the activity materially changes rather than decided once at the beginning and left unchanged indefinitely.
Learning Health Systems make the boundary more important, not less
A Learning Health System intentionally connects routine care, data, evidence and improvement. This is one of its strengths, but it can make boundaries between routine operations, QI and research less visible.
The solution is not to slow learning. It is to build governance capable of distinguishing different activities while allowing them to connect.
A healthcare organization can continuously monitor AI, rapidly improve workflows and conduct formal research when appropriate. Each activity needs governance arrangements proportionate to what it actually involves.
WHO's 2026 AI-research ethics work is particularly relevant because it argues that ethics oversight needs to address AI across the research lifecycle and that Research Ethics Committees may require additional expertise, training and institutional support as AI-related research becomes more complex [2].
Healthcare organizations in Viet Nam will increasingly need this capability as routine monitoring, quality improvement, real-world evidence generation and formal scientific evaluation begin to overlap.
A practical Clinical AI Evaluation Governance Screen
Healthcare organizations do not need to turn every AI evaluation into a legal exercise. A structured governance screen can help identify when closer institutional review is warranted.
The first question is purpose: is the primary objective to operate the AI safely, improve a local care process or systematically generate new knowledge? The second is change to care: is the evaluation observing existing practice, or are interventions, thresholds, messages, workflows or decisions being deliberately changed for evaluation? The third is participants and data: are identifiable patients or staff involved, are sensitive health data, recordings, biological material or linked datasets being used, and are additional data collected specifically for the project?
The fourth question is study design: is there a predefined scientific question, hypothesis, comparator, protocol, sampling plan or analysis intended to support scientific inference? The fifth is additional risk or burden: does participation expose people to procedures, decisions, data uses or inconvenience they would not otherwise face? The sixth is allocation: are patients, clinicians, wards or sites intentionally assigned to different AI or non-AI conditions?
The seventh is external involvement: are universities, vendors, funders or other organizations involved in study design, data access, analysis or publication? The eighth is intended use of the resulting knowledge: is the evaluation primarily supporting an immediate operational decision, or is it structured to answer a broader scientific question?
No single answer automatically determines legal status. Taken together, however, these factors can help identify when routine operational governance may be insufficient and when a formal institutional research-governance or ethics determination should be sought.
A practical routing model is therefore:
routine monitoring pathway → quality-improvement governance pathway → potential research pathway requiring formal research/ethics determination
When the status is genuinely uncertain, the organization should not guess. The appropriate institutional research-governance function, and where relevant the Ethics Committee, should determine the applicable pathway before the activity advances beyond what has already been authorized.
Do not use the governance screen to avoid ethics review
Any practical framework can be misused if people begin with the outcome they want. A team seeking rapid deployment may call a prospective experiment “quality improvement.” A vendor may describe a formal product evaluation as “post-implementation monitoring.” Investigators may call secondary analysis of sensitive clinical information an “audit” simply because no new patient contact occurs.
The purpose of a governance screen is the opposite: to identify the appropriate oversight pathway early enough that evaluation can proceed responsibly, not to find terminology that keeps a project outside ethics review.
The useful governance question is therefore not “How can this project stay outside research?” but “What level and type of oversight are appropriate for what we are actually planning to do?”
Ethics review should add value, not become a procedural formality
Research ethics should not be reduced to obtaining a committee approval document. For AI-related studies, meaningful review should consider whether the research question justifies the involvement of participants and use of their data, whether the scientific design is credible enough to justify the risks involved, whether relevant populations are represented, whether subgroup effects and bias have been considered, whether privacy and re-identification risks are appropriately controlled, whether human oversight is adequate and how unexpected findings or harms will be handled.
WHO's 2026 report emphasizes that Research Ethics Committees remain central to participant protection but may require additional expertise, training and resources for increasingly complex AI-related research. It also emphasizes that responsible research governance extends beyond Ethics Committees to researchers, regulators, funders, data-access bodies, journals and other institutions [2].
In Viet Nam, an Ethics Committee reviewing a complex AI study may therefore benefit from expertise in areas such as AI evaluation, data science, cybersecurity, human factors, healthcare quality or algorithmic bias. Circular No. 43/2024/TT-BYT already provides a mechanism for obtaining specialist input: where necessary, Ethics Committee leadership may invite an independent expert adviser who has no conflict of interest with the research. The adviser may participate in discussion but does not vote [1].
This does not mean every AI study requires a specialist AI panel. It means the review process should have access to expertise proportionate to the scientific and ethical questions raised by the study.
Research governance should continue after initial approval
Ethics approval is not the end of responsible AI research. AI systems can change during a study. Vendors may release updates. Input data may shift. New safety signals can emerge. Investigators may amend protocols or introduce new analyses.
Circular No. 43/2024/TT-BYT provides mechanisms for ongoing ethics oversight, including review of protocol changes, periodic and ad hoc reports, adverse-event information, protocol deviations and other updates. Certain categories may qualify for expedited review under Article 16 [1].
AI-related research therefore needs change control just as deployed AI does. Investigators and healthcare organizations should know which changes require Ethics Committee notification or review, which version of an AI system is being evaluated and whether an update changes the scientific assumptions, participant risks or validity of previously approved procedures.
A study evaluating “the AI system” becomes difficult to interpret if the technology changes during the study without adequate version documentation and governance.
Research findings should feed back into operational governance
The purpose of evaluation should not end with publication. If a study identifies poor performance in a patient subgroup, a new safety risk, excessive clinician workload, poor adherence or lack of meaningful clinical benefit, those findings should reach the people responsible for the deployed system.
Research governance, AI governance and quality governance therefore need a connection. A healthcare organization should not generate evidence showing that an AI-enabled pathway has an important limitation while leaving routine operations unchanged simply because research and operational governance do not communicate.
Negative or inconclusive findings can also be valuable. They may support a decision not to scale the system, to narrow its intended use, change workflow, strengthen monitoring or suspend or retire the technology.
Responsible evaluation should therefore create a loop:
question → appropriate governance → evidence generation → interpretation → operational decision → monitoring and learning
That is how research can contribute directly to health-system improvement rather than operate as an activity parallel to care.
What this means for healthcare organizations in Viet Nam
Healthcare organizations in Viet Nam will increasingly generate evidence about AI through routine clinical operations. That is a positive development. The challenge is to ensure that the speed of learning does not erase meaningful distinctions among operational monitoring, quality improvement and formal research.
Routine monitoring asks whether a deployed AI system remains safe and reliable. Quality improvement asks how a local care process can perform better. Research asks systematic questions intended to generate new knowledge and may require formal scientific and ethical oversight.
In practice, these activities overlap. The appropriate response is not to depend on labels but to examine purpose, design, participants, data, interventions, risk and the intended use of the knowledge generated.
Healthcare organizations should also avoid treating ethics review as the only governance issue. A single AI evaluation may simultaneously involve research ethics, personal-data protection, cybersecurity, clinical governance, medical-device regulation, contractual responsibilities and AI lifecycle obligations.
The practical transition is therefore:
from “we are only evaluating the AI” → to “what kind of evaluation are we conducting?” → to “what governance does this specific activity require?”
That question becomes increasingly important as healthcare organizations in Viet Nam move from isolated AI pilots toward real-world monitoring, comparative evaluation, evidence generation and Learning Health System approaches.
Responsible healthcare AI requires organizations to generate evidence. It also requires them to generate that evidence responsibly.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, research-ethics, clinical, statistical, data-protection, medical-device, cybersecurity or institutional-governance advice.
The distinctions among routine monitoring, quality improvement and research described in this article are practical governance concepts. They are not presented as a definitive statutory classification test under Vietnamese law.
Calling an activity a pilot, validation, quality-improvement project, audit, service evaluation, real-world evaluation or Learning Health System activity does not, by itself, determine whether formal research governance or Ethics Committee review is required. The appropriate pathway depends on the activity's purpose, design, participants, data, interventions, risks and applicable legal and institutional requirements.
The Clinical AI Evaluation Governance Screen proposed in this article is an organizational decision-support framework. It is not an official framework issued by the Ministry of Health, an Ethics Committee rule, a WHO classification system or a substitute for an institutional research-governance determination.
Research-ethics oversight and personal-data protection are separate but overlapping governance requirements. Approval, waiver or determination under one framework should not be assumed to satisfy requirements under another.
References
Ministry of Health of Viet Nam. Circular No. 43/2024/TT-BYT Regulating the Establishment, Organization and Operation of Ethics Committees in Biomedical Research. Issued 12 December 2024; effective 1 February 2025; currently recorded as in force.
World Health Organization. Artificial Intelligence-related Health Research: Ethics Review and Oversight. Published 21 July 2026.
National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Personal Data Protection. Issued 31 December 2025; effective 1 January 2026.
National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Artificial Intelligence. Issued 30 April 2026; effective 1 May 2026.
Ministry of Health of Viet Nam. Circular No. 19/2013/TT-BYT Providing Guidance on Management of Quality of Medical Examination and Treatment Services in Hospitals. Issued 12 July 2013; effective 15 September 2013; currently recorded as in force.
World Health Organization. Ensuring Ethical Standards and Procedures for Research with Human Beings. Current WHO research-ethics resource.