Integrating Clinical AI into Healthcare Quality and Patient Safety Systems in Viet Nam: From AI Governance to Routine Quality Management
IMPLEMENTATION INSIGHTSHEALTHCARE IMPLEMENTATION, MONITORING & EVALUATIONRESPONSIBLE AI, DATA GOVERNANCE & DIGITAL SAFETY
Integrating Clinical AI into Healthcare Quality and Patient Safety Systems in Viet Nam: From AI Governance to Routine Quality Management
IMPLEMENTATION INSIGHT
Last reviewed: 4 October 2026
A healthcare organization can have an AI committee, an AI inventory, a risk register, vendor assessments and model-monitoring dashboards, and still fail to integrate AI into the way it actually manages quality and patient safety.
Consider a clinical AI system that produces an unsafe recommendation. The AI governance team may record a model issue, the IT team may open a vendor ticket and the clinical department may remind staff to exercise caution. But if the event never enters the organization's patient-safety reporting system, is not examined alongside other care failures, generates no meaningful corrective action and is absent from quality review, the organization has governed the technology without fully governing the care affected by it.
That distinction matters as AI moves from pilot projects into routine clinical workflows. Once AI influences triage, diagnosis, prioritization, documentation, treatment decisions, communication or care coordination, it is no longer only a digital or technology issue. It becomes part of the system through which healthcare quality is produced.
The practical goal should therefore not be to replace existing quality management with a new AI bureaucracy. It should be to make clinical AI visible inside the organization's ordinary mechanisms for managing risk, learning from harm, measuring performance and improving care.
International practice is moving in the same direction. Joint Commission launched its voluntary Responsible Use of AI in Healthcare Certification in June 2026 for eligible U.S. healthcare organizations and health systems. Its standards are organized around governance, effective data management, risk and bias reduction, monitoring, evaluation and validation, and transparency, education and training [13]. Earlier Joint Commission-CHAI guidance also emphasized ongoing quality monitoring, voluntary reporting of AI-related safety events, risk and bias assessment, privacy, security, transparency and staff education [12]. These materials are not Vietnamese legal requirements, but the direction is important: responsible AI is increasingly being treated as part of healthcare quality and patient safety rather than as a separate technology-management problem.
Viet Nam already has a quality and patient-safety architecture into which AI can be integrated
Clinical AI does not enter Vietnamese healthcare organizations in an organizational vacuum.
Law No. 15/2023/QH15 on Medical Examination and Treatment, as amended and consolidated in Integrated Document No. 26/VBHN-VPQH dated 26 February 2026, establishes the current statutory foundation for healthcare quality and patient safety [1]. Article 57 defines categories of quality standards for healthcare establishments, while Article 58 makes maintenance and improvement of quality an explicit purpose of quality assessment and requires annual self-assessment against applicable basic quality standards.
Article 71 is particularly relevant to clinical AI. It provides that prevention of medical incidents is based on identifying and reporting incidents, analyzing their causes, developing recommendations and preventive measures, and preventing recurrence in order to improve quality and patient safety [1]. That is already a learning cycle. AI does not require the organization to invent one from the beginning.
For hospitals, Circular No. 19/2013/TT-BYT remains in force and provides a detailed organizational framework for hospital quality management [2]. It establishes structures including a Quality Management Council, a quality-management department or unit, dedicated quality personnel and a quality network extending into hospital departments and units. Its responsibilities include identifying quality problems and patient-safety risks, measuring quality indicators, managing incidents, supporting improvement activities, conducting internal quality assessment, monitoring compliance and developing patient-safety programs.
Circular No. 35/2024/TT-BYT has applied basic hospital quality standards since 1 January 2025 [3]. Under Article 120(6) of the Law on Medical Examination and Treatment, application of the basic quality standards began for hospitals on 1 January 2025 and is scheduled to extend to other organizational forms of healthcare establishments from 1 January 2027 [1]. Circular No. 35 itself is hospital-specific, so the applicable implementing requirements for other organizational forms should be assessed when that stage takes effect.
The 2024–2025 national hospital quality-assessment cycle also illustrates the broader quality environment. Decision No. 3652/QĐ-BYT and subsequent implementation guidance required hospitals to assess the Basic Quality Standards under Circular No. 35 while continuing assessment under the Viet Nam Hospital Quality Criteria Version 2.0 issued through Decision No. 6858/QĐ-BYT, with specified adjustments [4,5]. These instruments were not designed as AI governance standards, but they demonstrate the quality-management infrastructure into which clinical AI increasingly needs to fit.
Circular No. 43/2018/TT-BYT provides another important operational reference for medical-incident prevention and learning [6]. Its text includes the concept of a near miss and an incident-management framework covering reporting, analysis, feedback and preventive action. As of this review, the national legal database records the Circular as partially expired, so individual provisions should be applied only to the extent currently effective and read together with the current Law on Medical Examination and Treatment and subsequent applicable regulations. The broader implementation principle remains useful: healthcare organizations already have mechanisms for learning from incidents and near misses, and AI-related events can be connected to those mechanisms rather than automatically placed in an entirely separate system.
AI governance and healthcare quality management answer different questions
AI governance and healthcare quality management overlap, but they are not interchangeable.
AI governance asks questions such as: What AI systems are being used? Who approved them? What is the intended use? Which data do they rely on? How were they validated? What risk classification applies? What human oversight is required? How are changes controlled? What are the vendor's responsibilities?
Healthcare quality and patient-safety management asks a different set of questions: Is care safe? Are processes reliable? Are outcomes improving? What went wrong? Why did it happen? Could the same problem happen again? Are corrective actions working? Are patients experiencing harm, delay, inequity or poor communication?
Once AI enters clinical work, those questions intersect. A model can meet a technical performance threshold while creating a poor workflow. A vendor can meet its service-level agreement while clinicians become increasingly over-reliant on the system. A model can remain statistically stable while the clinical pathway around it changes enough to create new risk.
IHI's 2026 work on AI monitoring makes the distinction useful: governance establishes what needs oversight, while monitoring helps determine what is actually happening in use. Effective monitoring needs to look beyond statistical model performance to patient outcomes and user adoption [14].
The implication is that AI governance should connect to quality management, but neither should disappear into the other. Cybersecurity still requires cybersecurity expertise. Model performance still requires analytical capability. Personal-data protection still requires appropriate data governance. Medical-device questions still require regulatory assessment. But when those issues affect care, they also need to become visible to the quality and patient-safety system.
Do not create a second patient-safety system just for AI
A separate technical pathway for investigating AI problems can be useful. A separate patient-safety system usually is not.
Suppose a clinician follows an AI-generated recommendation and treatment is delayed. Several factors may have contributed: an inaccurate AI output, incomplete source data, poor interface design, automation bias, inadequate training or an escalation process that was never clearly defined.
Calling the event simply an “AI incident” can hide the larger system failure. It belongs in the same organizational safety architecture that would examine other failures in clinical decision support, communication, diagnosis or workflow.
For hospitals in Viet Nam, Circular No. 19 already assigns the quality-management function responsibilities for detecting, aggregating, analyzing and reporting errors and incidents and proposing corrective solutions [2]. Circular No. 43 provides a more detailed incident-management architecture, subject to its current legal status, including incident reporting and analysis of contributing factors and causes [6].
Clinical AI can therefore be captured as a contributing factor or relevant technology within an existing patient-safety event, rather than requiring every AI-related problem to become a new and isolated category of healthcare harm.
The incident record may need additional information: which AI system was involved; what system or model version was active; what output was generated; which important inputs were available; whether the clinician saw the output; what action followed; whether the AI was overridden; and whether technical or vendor-related factors contributed.
Those details support AI governance. The event itself remains part of patient safety.
Where investigation requires processing or sharing prompts, outputs, logs, recordings, medical records or other information relating to identifiable patients, the applicable requirements of Viet Nam's Personal Data Protection Law and Decree No. 356/2025/NĐ-CP also need to be considered [10,11]. Safety investigation does not suspend privacy, confidentiality or personal-data obligations.
Near misses and weak signals should become part of organizational learning
Healthcare organizations should not focus only on cases in which AI clearly causes serious harm. Near misses may reveal weaknesses much earlier.
A clinician may discover an incorrect medication in an AI-generated summary and correct it before signing. A radiologist may disregard an inappropriate AI priority flag. A nurse may notice that an early-warning system is using stale observations. A physician may identify an irrelevant or outdated source retrieved by a generative AI assistant before acting on its recommendation.
No harm occurred, but the control system was tested.
Circular No. 43 defines a near miss as a situation that has occurred without causing consequences, or nearly occurred but was detected and prevented before harming the patient [6]. Its text also establishes incident-reporting mechanisms, although individual provisions need to be read in light of the Circular's current partially expired status.
This concept is particularly useful for AI because severe harm may be uncommon while weak signals appear much earlier: repeated overrides, unusual workarounds, recurrent correction of AI-generated documentation, staff complaints, unexplained differences between departments or patterns of outputs that clinicians consistently distrust.
An organization that waits for a severe AI-related injury before investigating its controls is effectively using patient harm as its detection mechanism. A stronger quality system learns from near misses, weak signals, repeated overrides, unusual workarounds and emerging patterns.
Root-cause analysis should not stop at “the AI was wrong”
When an AI-related event occurs, “model error” is rarely a sufficient root-cause conclusion.
The model may have produced an incorrect answer, but the safety question is why that answer was able to influence care in the way it did. The contributing factors may include data quality, workflow design, alert placement, inadequate escalation, poor interface design, an inappropriate threshold, outdated knowledge, a vendor change, insufficient training or overconfidence in AI output.
The same principle applies when a human makes the final decision. Human review should not become the explanation that ends the investigation.
If clinicians repeatedly accept unsafe AI recommendations, the organization should ask why. Was the system presented as more authoritative than intended? Were users trained to understand its limitations? Was workload too high for meaningful review? Were alternative clinical data difficult to access? Had repeated prior experience taught clinicians that the AI was usually correct?
Research on operational safety in clinical AI reinforces this systems perspective. Conventional model-performance metrics do not by themselves define the real-world conditions under which an AI system can be used safely [17].
The useful question is therefore not simply “Who made the error?” It is:
“What combination of technology, people, process and environment allowed this risk to reach the patient?”
Corrective action should address the care system, not just the model
Corrective action should follow the cause identified. If an event arises from a data-interface mapping problem, the solution may be technical. If excessive false-positive alerts are the problem, threshold review or workflow redesign may be required. If clinicians misunderstand the system's intended use, education and interface communication may need to change. If an upstream model changed without adequate notification, vendor-management and change-control processes may need strengthening.
Retraining the AI is therefore only one possible response. Other actions may include revising the clinical procedure, modifying escalation criteria, reducing automation, adding confirmation steps, improving source-data quality, changing role responsibilities or restricting the system to a narrower use case.
Corrective action also needs an effectiveness check. A closed vendor ticket does not prove that the patient-safety problem has been solved. Publishing a revised policy does not prove that staff follow it. Completing a training session does not demonstrate that behavior changed.
The quality system should define what evidence would show that the intervention actually reduced the risk. This is where established quality-improvement practice adds something that technology governance alone often lacks.
AI monitoring should become part of the quality indicator architecture
Clinical AI dashboards frequently focus on model statistics such as sensitivity, specificity, calibration, AUC, latency, drift and uptime. These measures can be important, but quality leadership needs a broader view.
For clinically consequential AI, monitoring may need to connect four levels:
model performance → workflow performance → patient-safety signals → patient outcomes
A deterioration model might be monitored for sensitivity and calibration, but the organization may also need to understand alert acknowledgment, time to clinical response, unnecessary escalation, missed deterioration, ICU transfer and relevant patient outcomes.
A documentation AI system may require monitoring of note accuracy, while quality management may also need to examine correction burden, delayed documentation, medication-list errors, clinician workload and patient complaints.
This approach fits the existing Vietnamese hospital quality framework. Circular No. 19 requires hospitals to develop and monitor quality indicators, collect and analyze quality information and use those results for improvement [2]. IHI's 2026 AI-monitoring work similarly emphasizes connecting statistical performance with outcome performance and user adoption [14].
The objective is not to put every AI metric on the executive dashboard. It is to determine which AI-related measures are clinically meaningful enough to become quality indicators, and which thresholds should trigger review, investigation or escalation.
Audit and case review should include AI when AI influenced the pathway
Healthcare organizations often inspect an AI system intensively before deployment and much less intensively once it becomes routine. Quality review provides an opportunity to correct that imbalance.
Internal quality assessments, clinical audits, tracer activities, documentation review, serious-case analysis, mortality review and other case-based processes can examine AI when it materially affected the pathway being assessed. Reviewers may need to establish whether the AI was used within its intended purpose, whether adequate human oversight occurred, whether important data were available, whether the AI output was documented appropriately, whether escalation policy was followed and whether an AI-related deviation contributed to the outcome.
This does not mean every chart review requires an AI checklist. AI should be examined when it is relevant to the care being reviewed.
Retrospective traceability becomes especially important. If a serious event occurred six months earlier, can the organization reconstruct which AI system and version were active, what output was presented, what important information the system received, who saw the output and what action followed?
If not, meaningful quality investigation may become impossible.
Where organizations use morbidity and mortality review, serious-case review or other multidisciplinary case-review processes, the same principle applies. If AI influenced diagnostic reasoning, prioritization, timing or confidence, its role should be visible. If AI helped prevent harm, that should also be recognized.
The objective is not to turn case review into a technical model audit. It is to understand the whole system of care.
Competency, patient feedback and clinical policy should make AI operationally visible
AI governance documents often state that users must be trained. Quality management should go further and ask whether users are competent to use the system safely.
Training may explain what a tool does. Competency assessment should examine whether staff can apply that knowledge in realistic situations. Can users recognize when AI is outside its intended use? Do they know what to do when AI conflicts with the clinical picture? Can they escalate a suspected AI error? Can they continue safely during downtime? Do they understand what changed after a major update?
Circular No. 19 already connects hospital quality management with education and capacity building [2]. Joint Commission's 2026 RUAIH framework similarly includes education and training as part of responsible organizational AI governance [13].
Competency may also need to be version-sensitive. A material change in system behavior, workflow or permitted use may justify renewed education or reassessment rather than assuming that training remains valid indefinitely.
Patient feedback provides another source of quality information that technical monitoring may miss. Patients may notice that documentation contains something they never said, that an AI-enabled interface makes access to a clinician more difficult, or that AI-generated material contains inaccurate information. Those signals should be routed appropriately when they suggest clinical, quality, privacy or governance risk.
An organization-wide AI policy is useful, but it cannot carry the entire burden of implementation. If AI prioritizes radiology studies, relevant radiology procedures should describe how it affects workflow. If AI supports medication reconciliation, medication-management procedures should address its role. If ambient AI produces draft clinical notes, documentation procedures should define the required review and authentication process.
That is the difference between having an AI policy and embedding AI safely into clinical policy.
Change control should connect technical updates with clinical-quality review
Clinical AI can change even when the healthcare organization's clinical protocol does not. Models are updated. Data sources change. APIs change. Thresholds are adjusted. Workflow integrations are redesigned. Foundation models may be replaced.
Some changes are technical and low consequence. Others can alter the assumptions under which the system was originally validated.
For high-risk AI systems, Article 13 of Decree No. 142/2026/NĐ-CP requires conformity reassessment where specified significant changes affect the basis of the original conformity assessment [8]. Article 15 establishes risk-management responsibilities for providers and distinct operational responsibilities for deployers, including monitoring deviations, risks and incidents, maintaining human oversight, applying risk controls within the deployer's authority and coordinating with the provider when significant changes or incidents require risk-management measures to be reviewed [8].
Healthcare quality management can support that regulatory framework by asking a practical question:
Does this technical change also require a clinical-quality review?
A change may not require formal conformity reassessment or complete revalidation yet still justify targeted evaluation of workflow, documentation accuracy, incident patterns, user behavior or clinical outcomes.
The organization's internal quality threshold can therefore be broader than the statutory threshold for formal conformity reassessment.
Legal high-risk classification should not define the quality-management threshold
Decision No. 33/2026/QĐ-TTg, effective from 15 August 2026, establishes Viet Nam's current formal list of high-risk AI systems [9]. As of this review, its healthcare section identifies specified AI-assisted surgical and robotic-surgery systems under defined conditions.
Most clinical decision-support, predictive and generative-AI tools therefore do not become legally high-risk simply because they are used in healthcare.
That distinction matters legally, but it should not determine whether quality teams pay attention.
A discharge-summary generator outside the formal high-risk list can still introduce a medication error. A triage assistant can still contribute to delayed care. A documentation model can still place incorrect clinical information into the medical record.
Healthcare quality management should therefore respond to clinical consequence as well as regulatory classification.
A risk does not need a special statutory label before a healthcare organization should manage it.
AI-related risk should connect to the organization's broader risk system
An AI-specific risk register can contain important technical information: system owner, model version, vendor, data dependencies, validation, controls, identified hazards and monitoring arrangements.
Clinically significant AI risks should also become visible in the organization's broader clinical or enterprise risk-management structure where appropriate.
If an AI-supported triage pathway creates a known risk of delayed escalation for a particular group, this is not only an “AI risk.” It is a patient-safety risk in the triage pathway.
The AI risk register and the clinical or enterprise risk register should therefore reference each other rather than compete as separate sources of truth.
This also matters in multi-site healthcare systems. Central AI governance can standardize procurement controls, minimum validation requirements, vendor management, change notification and core monitoring. Local quality teams still need ownership of whether the AI works safely within their own staffing, patient population, equipment, workflow and response capacity.
An organization-wide average can appear acceptable while one facility experiences a recurring local failure. Responsible multi-site governance therefore requires central standards with local quality accountability.
AI can strengthen quality and patient-safety work, but it also needs governance
The relationship between AI and quality management is not only about protecting patients from AI. AI may itself support quality and safety work.
A 2026 npj Digital Medicine study developed an AI-based incident analysis and learning system using 1,548 expertly curated training incidents and evaluated the system on 350 real-world radiation-oncology incidents. In the real-world evaluation, agreement with expert reviewers reached 88% when cases considered reasonably concordant despite limited report detail were included, and processing was substantially faster than manual analysis [15].
A separate 2026 BMJ Quality & Safety study analyzed 9,357 patient-safety report narratives using a large language model to extract safety problems and develop a structured taxonomy, illustrating how AI may help expose patterns across large volumes of narrative safety data [16].
These studies demonstrate potential, not a case for automating safety investigation.
Incident reports are incomplete. Causality is difficult. Organizational context matters. AI may identify patterns while missing local knowledge understood by experienced quality and safety professionals.
The principle should therefore remain consistent:
AI can support the quality-management system, but AI used by the quality-management system must itself be governed.
A practical Clinical AI–Quality Integration Framework
Healthcare organizations do not need to redesign their entire quality-management system around AI. A more practical approach is to connect AI to the systems already responsible for safe care.
The first domain is leadership and governance integration. Which AI risks need escalation to executive leadership, the Quality Management Council or an equivalent quality structure? How does the AI-governance function communicate with quality, patient safety, clinical leadership, IT, cybersecurity, data protection and other relevant functions? Which problems remain primarily technical, and which have become clinical-quality issues?
The second domain is risk integration. Are clinically significant AI hazards connected to the organization's wider clinical or enterprise risk system? Is the potential consequence for patients visible outside the AI team? Are risk owners identified? Can the organization distinguish technical, clinical and regulatory risk without allowing them to become disconnected?
The third domain is incident reporting and learning. Can staff report AI-related adverse events and near misses through ordinary patient-safety channels? Can investigators identify which AI system was involved and reconstruct the relevant context? Are events routed to quality, clinical, technical, privacy and vendor teams where appropriate?
The fourth domain is monitoring and quality measurement. Which model metrics matter clinically? Which workflow, safety and patient-outcome indicators should accompany them? What thresholds trigger investigation? Does monitoring include adoption, overrides, complaints and unintended consequences rather than model performance alone?
The fifth domain is audit and assurance. Do relevant internal assessments, clinical audits, tracer activities and case reviews examine AI where it affects the pathway being assessed? Can the organization demonstrate that human oversight, escalation, documentation and fallback controls work in practice rather than merely existing in policy?
The sixth domain is competency and policy integration. Are AI-related responsibilities embedded into the clinical procedures people actually use? Are staff trained and assessed for safe use? Do significant changes trigger renewed education or competency review where appropriate? Are downtime and escalation arrangements understood before they are needed?
The seventh domain is corrective action and improvement. When an AI-related problem occurs, does the response address underlying causes across technology, workflow, people and policy? Is responsibility for corrective action clear? Is effectiveness checked? Can recurring signals become formal improvement work rather than repeated isolated fixes?
The eighth domain is leadership review and organizational learning. Does leadership receive a concise view of material AI risks, incidents, clinical outcomes, quality trends, major system changes and unresolved actions? Can lessons from one department or site be shared across the organization? Does accumulated evidence influence decisions to continue, modify, expand, restrict, suspend or retire an AI system?
The Clinical AI–Quality Integration Framework is a practical organizational framework proposed for healthcare quality, patient safety and responsible AI implementation. It is not a statutory checklist prescribed by Vietnamese law.
Routine quality management is where AI governance becomes real
An AI governance committee can approve a policy, but it cannot by itself make a clinical workflow safe.
That happens in everyday operations: when staff recognize an unsafe output, an incident is reported, a quality team notices a pattern, an audit shows that a control is not working, a department redesigns a process, or leadership checks whether corrective action actually reduced risk.
The mature state is therefore not one in which every AI problem is automatically sent to a central AI committee. It is one in which staff understand that an AI-related documentation error belongs in documentation quality review; an AI-related clinical near miss belongs in the patient-safety learning system; an AI-related workflow problem belongs in improvement work; and an AI change that could alter care triggers the appropriate technical and clinical review.
Specialist AI governance remains necessary.
Quality management makes that governance operational.
What this means for healthcare organizations in Viet Nam
Viet Nam already has many of the building blocks needed to integrate clinical AI responsibly: statutory quality standards, annual quality assessment, hospital quality-management structures, quality indicators, incident reporting, near-miss learning, cause analysis, corrective action, workforce education and continuous improvement [1–6].
The newer AI framework adds another layer. Law No. 134/2025/QH15 and Decree No. 142/2026/NĐ-CP establish risk-based AI governance, with specific additional requirements for high-risk AI systems [7,8]. These requirements do not replace healthcare quality obligations, and healthcare quality systems do not replace compliance with AI law.
The opportunity is to connect them intelligently.
Healthcare organizations should not need one isolated system to govern “AI safety” and another unrelated system to govern “patient safety” when the same event can involve both.
The leadership question should therefore move beyond “Do we have an AI governance program?”
A more useful question is:
“When AI affects care, can our quality and patient-safety system detect the risk, understand what happened, improve the process and verify that the improvement actually worked?”
That is where AI governance becomes routine healthcare quality management.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, accreditation, quality-management, data-protection, cybersecurity or medical-device advice.
The terms Clinical AI–Quality Integration Framework, AI-related incident, AI-related near miss, quality integration, corrective action and effectiveness check are used here as practical healthcare quality, patient-safety and AI-governance concepts. The organizational framework proposed in this article is not a statutory checklist under Vietnamese law.
Circular No. 19/2013/TT-BYT and Circular No. 35/2024/TT-BYT specifically address hospitals. Article 120(6) of the current consolidated Law on Medical Examination and Treatment schedules application of basic quality standards to other organizational forms of healthcare establishments from 1 January 2027, but the applicable implementing requirements should be assessed for each type of establishment.
As of this review, the national legal database records Circular No. 43/2018/TT-BYT as partially expired. References in this article to its near-miss and medical-incident framework should therefore be understood only to the extent the relevant provisions remain applicable and together with the current Law on Medical Examination and Treatment and subsequent regulations.
Similarly, Articles 13 and 15 of Decree No. 142/2026/NĐ-CP establish specific obligations relating to high-risk AI systems. The broader quality-integration practices proposed here are organizational recommendations and should not be interpreted as statutory requirements for every AI system.
Joint Commission and IHI materials cited below are international healthcare quality, patient-safety and responsible-AI references. They are not Vietnamese legal requirements and should be adapted to the organization's clinical setting, risk profile and applicable Vietnamese law.
References
[1] Office of the National Assembly of Viet Nam. Integrated Document No. 26/VBHN-VPQH consolidating Law No. 15/2023/QH15 on Medical Examination and Treatment, as amended. Issued 26 February 2026. Original Law effective 1 January 2024.
[2] 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.
[3] Ministry of Health of Viet Nam. Circular No. 35/2024/TT-BYT on Basic Quality Standards for Hospitals. Issued 16 November 2024; effective 1 January 2025.
[4] Ministry of Health of Viet Nam. Decision No. 6858/QĐ-BYT Promulgating the Viet Nam Hospital Quality Criteria, Version 2.0. Issued 18 November 2016.
[5] Ministry of Health of Viet Nam. Decision No. 3652/QĐ-BYT on Hospital Quality Assessment and Patient and Healthcare Worker Satisfaction Survey Content for 2024–2025. Issued 3 December 2024.
[6] Ministry of Health of Viet Nam. Circular No. 43/2018/TT-BYT Providing Guidance on Prevention of Medical Incidents in Medical Examination and Treatment Establishments. Issued 26 December 2018; effective 1 March 2019. Currently recorded in the national legal database as partially expired.
[7] National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
[8] 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.
[9] Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. Issued 30 June 2026; effective 15 August 2026.
[10] National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
[11] 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.
[12] Joint Commission and Coalition for Health AI. Guidance on Responsible Use of AI in Healthcare. 2025.
[13] Joint Commission. Responsible Use of AI in Healthcare Certification. Launched June 2026.
[14] Zier L, Weckman A, Martinez N. AI Monitoring: From Model Metrics to Patient Outcomes. Institute for Healthcare Improvement; 4 March 2026.
[15] Jinia AJ, Chapman K, Liu S, et al. Artificial intelligence-based incident analysis and learning system to enhance patient safety and improve treatment quality. npj Digital Medicine. 2026;9:254.
[16] Chen K, Rogers K, Haberkorn W, et al. AI-driven analysis of patient safety reports using large language models: an exploratory multiple methods study. BMJ Quality & Safety. Published 2026.
[17] Kim YT, Kim H, Bahl M, et al. Defining operational safety in clinical artificial intelligence systems. npj Digital Medicine. 2026;9:281.