Documenting AI Use in Clinical Records in Viet Nam: A Practical Approach to Provenance, Human Review and Traceability
IMPLEMENTATION INSIGHTS
Documenting AI Use in Clinical Records in Viet Nam: A Practical Approach to Provenance, Human Review and Traceability
IMPLEMENTATION INSIGHT
Last reviewed: 3 October 2026
A clinical note appears in the medical record. It reads clearly, uses appropriate terminology and has been signed or electronically confirmed by a clinician. But how did it get there?
Was it written entirely by the clinician? Was it transcribed from a consultation by an AI tool, summarized from several existing records, drafted by a generative AI system, translated from another language, or influenced by an AI recommendation that the clinician subsequently edited? If the AI system is later found to have a significant defect, can the healthcare organization identify which records it influenced and which patients may need review?
As AI becomes more deeply embedded in documentation and clinical workflows, these questions are moving from technical detail to patient-safety governance. The objective is not to turn the medical record into a technical log or require clinicians to document every interaction with software. It is to preserve enough information to understand when AI materially influenced clinical information, which system was involved, what role the human reviewer played and how the organization can reconstruct that process if questions arise later.
For this article, AI provenance means the traceable history of how information came to exist in its current form: which people and systems contributed, what information or sources were used, what transformations occurred and who ultimately reviewed or accepted the result. For healthcare organizations in Viet Nam, this is increasingly relevant not because current law creates a universal requirement to label every AI-assisted clinical note, but because AI governance, electronic medical records, patient safety, personal-data protection and accountability are beginning to meet in the same workflow.
Clinical information may now have more than one authoring pathway
Traditional clinical documentation gives the reader a relatively straightforward mental model. A clinician assesses a patient, records findings and decisions, and signs or confirms the relevant documentation. AI complicates that model because clinically relevant information can now reach the record through several pathways.
An ambient documentation system may convert a consultation into a draft clinical note. A generative AI assistant may summarize a long history. Software may extract laboratory information and populate structured fields. An imaging system may generate findings that enter a radiology workflow. AI may draft discharge instructions that a clinician subsequently edits, or generate a risk score or recommendation that affects a care plan without producing visible prose at all.
The final record may look conventional even when the process that produced it was not. NIST highlighted this problem in September 2026 through its work on HL7 AI Transparency on FHIR: health data may increasingly be created or modified by AI while appearing indistinguishable to downstream users from conventionally produced data if information about AI involvement is not preserved.
That becomes important when another clinician relies on the information, when a patient questions what was recorded, when a quality team investigates an incident, when a vendor changes or withdraws a model, or when an organization later discovers that a particular model version produced unreliable outputs.
Provenance, authorship, audit logs and AI labels are different things
These concepts are easy to mix together, but they solve different problems. Clinical authorship and responsibility concern the person accountable for documentation or a clinical decision. AI attribution answers the narrower question of whether AI contributed to creating or changing information. Provenance preserves relevant information about how that contribution occurred. Audit logs usually capture technical events such as access, creation, modification or system activity. An AI label may indicate that content was AI-generated or AI-modified without explaining the process behind it.
None of these automatically substitutes for the others. A clinician's signature on an AI-drafted discharge summary may identify who accepted the final document but not which AI system produced the draft. A technical log may show that an AI service was called without showing which part of the final record it influenced. A visible “AI-assisted” label may improve transparency while providing little information about the model version, input sources or human review.
The practical governance question is therefore not simply whether an AI label exists. It is what information may need to be reconstructable later, for what purpose and by whom.
What Viet Nam's current legal framework requires, and what it does not
Viet Nam's medical-record framework already provides an important accountability foundation. Under the current consolidated Law on Medical Examination and Treatment, patients receiving inpatient, day or outpatient treatment must have a medical record created and updated. Paper and electronic medical records have equivalent legal value, and medical records must be stored and kept confidential in accordance with law.
Article 52 of Circular No. 32/2023/TT-BYT requires medical-record information to be recorded accurately, truthfully and completely. It also requires the documentation to clearly show the time of recording and the person making the entry.
For electronic medical records, Circular No. 13/2025/TT-BYT has been effective since 21 July 2025. An electronic medical record is created, updated, displayed, signed, stored, managed, used and accessed electronically, and must contain the information required under the medical-record framework. The Circular also requires compliance with relevant laws on healthcare, data, electronic transactions, information security, cybersecurity, access to information, personal-data protection and data storage. EMR infrastructure must also support recovery and retrieval of information when needed for treatment, inspection, research and health management.
Circular No. 13 provides several methods by which healthcare workers, patients or their representatives may sign or electronically confirm relevant content, including lawful electronic signatures, biometric techniques and other electronic confirmation methods permitted under the Law on Electronic Transactions. The framework is therefore structured around identifiable human participants who record, sign or confirm relevant medical-record information; it does not provide for an AI system to replace those human actors as the person signing or confirming the record.
At the same time, the Vietnamese medical-record rules reviewed for this article do not impose a general requirement that every AI-assisted or AI-generated clinical note visibly state that AI was used.
The AI legislation should be read separately. The Law on Artificial Intelligence No. 134/2025/QH15 has been effective since 1 March 2026, and Decree No. 142/2026/NĐ-CP since 1 May 2026. Article 6 of the AI Law gives healthcare specific attention, requiring stricter risk management in essential sectors and highlighting three considerations for healthcare: patient safety, reliability under real-world conditions of use and protection of health data.
The AI framework also creates transparency and labeling obligations for AI-generated content, but those obligations are not the same as clinical-record provenance. Article 17 of Decree No. 142 requires providers to apply machine-readable technical marking to AI-generated or AI-modified audio, images and video. Providers are not generally required to apply this machine-readable marking to text outputs, unless another law provides otherwise. Article 18 also provides exceptions from visible labeling for internal organizational content that is not made public and for text processed through functions such as spelling or grammar correction, summarization, paraphrasing or translation where the essential meaning is not distorted.
Those provisions should not be extended into a claim that Vietnamese law currently requires every AI-assisted clinical note to carry a visible “AI-generated” or “AI-assisted” label.
Clinical provenance addresses another problem. A healthcare organization may decide that clinically meaningful AI involvement should remain traceable internally for patient safety, quality assurance or incident investigation even where a public-content labeling obligation does not apply. Conversely, a statutory AI label may tell someone that content originated from AI while providing far less information than is needed to understand how a clinical record was produced.
Patient communication is related but also distinct. Whether and how AI involvement should be explained to a patient will depend on the system, the significance of its role, the clinical context, applicable legal requirements and organizational policy. A spelling assistant and an AI system materially influencing a diagnosis are not equivalent. Provenance supports better communication because the organization knows how AI participated; it should not be confused with a universal disclosure rule.
The broader AI framework nevertheless strengthens the case for traceability. Decree No. 142 defines technical dossiers for high-risk AI systems and defines a model card as documentation describing matters including characteristics, intended use, scope, limitations, potential risks, training information and performance, with the objective of supporting transparency, traceability and reproducibility across model development and use. Providers and deployers also have transparency obligations that include providing information about intended use, scope, conditions and limitations, and retaining information and documentation required for inspection and supervision.
None of this means that model cards, technical dossiers or detailed system logs should be placed in the patient's medical record. It does indicate a wider regulatory direction toward AI systems that can be identified, understood, monitored and investigated rather than used as opaque tools.
It is equally important not to confuse this governance logic with statutory high-risk classification. Decision No. 33/2026/QĐ-TTg, effective from 15 August 2026, currently contains two healthcare entries in its high-risk list, both concerning specified AI-assisted surgical or robotic systems under defined conditions. It does not make every clinical documentation system, ambient scribe, summarization tool or other clinical AI application legally high-risk simply because it is used in healthcare.
Provenance can therefore be appropriate as an organizational patient-safety and governance control even when the particular AI system is not legally classified as high-risk.
Documentation should be proportionate to clinical consequence
Not every use of AI needs the same level of documentation. If an approved tool corrects spelling in a generic administrative sentence without changing its meaning, detailed model provenance may add little value. If AI drafts a patient-specific discharge summary, summarizes prior clinical information or translates instructions that enter the medical record, traceability becomes more important. If an AI system generates an interpretation, prediction or recommendation that may influence diagnosis, treatment, triage, medication or another clinical decision, the case becomes stronger again.
A useful internal distinction is between cosmetic assistance, substantive documentation assistance and clinically consequential AI contribution. Cosmetic assistance changes presentation without materially changing clinical meaning. Substantive assistance creates, summarizes, translates, extracts or restructures patient-specific information. Clinically consequential contribution occurs when an AI interpretation, prediction or recommendation may influence a care decision.
These are practical governance categories proposed for organizational use, not statutory classifications under Vietnamese law. Their purpose is to help the organization decide how much traceability is proportionate. The more consequential the AI contribution, the stronger the case for preserving reliable attribution, human review and technical provenance.
Human review needs the same proportionality. “Human in the loop” can sound reassuring, but the phrase means little unless the organization defines what the human is expected to review and what authority that person has to change, reject or disregard the AI output. A clinician clicking “accept” after a superficial glance is not equivalent to reviewing an AI-drafted note against the consultation, available patient information and clinical judgment.
For relatively low-consequence documentation assistance, review may focus on factual accuracy, patient identity and preservation of intended meaning. For clinical summaries, the reviewer may need to confirm that important information has not been omitted or distorted. For recommendations that could influence care, the reviewer needs to decide whether the output is clinically appropriate in light of the patient's actual circumstances before it contributes to a decision.
Where these distinctions matter, the workflow can differentiate states such as AI-generated but not yet reviewed, reviewed, materially edited after review, and accepted or approved by the responsible human. These states need not all appear as visible text in the medical record, but the organization should be able to determine them when patient safety, quality review or accountability requires it.
The purpose is not to create another box for clinicians to click. It is to avoid a false impression of purely human authorship when AI materially shaped clinically significant information and the nature of human review matters.
The medical record should not become the AI audit log
Healthcare organizations may need to preserve significant information about AI use, but not all of it belongs in the visible clinical record. One practical design is to separate information into three layers.
The clinical-record layer contains the information needed for care and required under the medical-record framework, including identifiable human responsibility and, where clinically material, an appropriate indication that AI contributed. A structured provenance layer can link an affected record element to information such as the AI system or deployment, relevant version, time of use, type of AI contribution and human reviewer. A more restricted technical and governance layer can retain system logs, configuration information, model-card references, input or output records where justified, and links to the approved use case.
This three-layer approach is a governance design option, not a structure prescribed by Vietnamese legislation. Its value is that it can preserve useful traceability without turning the clinical record into a technical engineering record.
The distinction also matters for privacy. Raw prompts may contain extensive health information, duplicate information already held elsewhere or include context that does not need to be retained as another copy. Viet Nam's Personal Data Protection Law, effective from 1 January 2026, specifically addresses personal-data processing in AI environments and requires AI-based personal-data processing to be classified by risk so that appropriate protective measures can be applied. Decree No. 356/2025/NĐ-CP also classifies health status as sensitive personal data and requires access restrictions, processing procedures and security measures for sensitive data.
For that reason, “store every prompt” is not a sound universal provenance policy. A healthcare organization should determine whether a particular workflow requires the exact prompt, a prompt-template identifier and version, an input-category record, a technical reference or another less data-intensive method. The same reasoning applies to intermediate model outputs: retaining them may sometimes be important for investigation, but indefinite retention without a defined purpose can create additional privacy, access-control and security exposure.
Provenance should improve accountability without creating unnecessary copies of sensitive data.
A practical minimum provenance model for organizational use
For governance purposes, a healthcare organization may choose to make a minimum set of information reconstructable when AI materially contributes to clinical documentation or other clinically consequential information. Depending on the use case, that may include the patient and encounter concerned, the affected record element, the nature of the AI contribution, the AI system or deployment involved, a model or configuration version where meaningful and available, the time of the contribution, the clinician or authorized professional who reviewed it, the review status, whether material edits occurred, the general category of input or source information and the location of more detailed technical records if an investigation is required.
This is not a legally prescribed minimum dataset in Viet Nam. It is a practical organizational model for deciding what should remain traceable.
Most of this information does not need to appear visibly beside every clinical sentence. A discharge summary, for example, may remain easy for clinicians to read while structured metadata records that an approved AI documentation system contributed to the draft, identifies the relevant deployment version, records when the event occurred, identifies the clinician who reviewed and edited the content and links the signed record to a protected technical audit trail.
The principle matters more than the exact fields: when clinically important information needs to be investigated later, the organization should be able to move backward from that information to the process that produced it.
Provenance becomes most valuable when something changes later
The value of traceability may not become obvious until months after deployment. A healthcare organization may discover that a model version performed poorly in a particular patient group. A vendor may report a defect. A prompt template may be found to omit important context. An update may materially change system behavior, or an incident investigation may identify an AI recommendation that contributed to delayed or inappropriate care.
Without provenance, the organization may know that an AI system was in use but still be unable to identify which patients, records or decisions were actually affected.
This is also why version information matters. “AI was used” is rarely sufficient when models, retrieval sources, prompts and system configurations can change over time. For clinically consequential uses, organizations should preserve enough information to distinguish changes that may matter to performance or safety. The useful identifier might be a product version, model version, deployment identifier, validated configuration or another stable internal reference, depending on the technology.
Vendor agreements and change-control processes should support that capability. Provenance tells the organization what produced a past output; change control governs what may change before future outputs are produced.
The emerging HL7 AI Transparency on FHIR work includes exactly this kind of retrospective use case, including identifying outputs from a model later determined to be problematic or outputs associated with a poorly constructed input or prompt.
That is where provenance stops being an abstract transparency principle and becomes a patient-safety capability.
HL7 AI Transparency on FHIR offers an emerging technical approach
One of the most relevant technical developments in 2026 is the HL7 AI Transparency on FHIR Implementation Guide. FHIR, Fast Healthcare Interoperability Resources, is a widely used interoperability standard for structuring and exchanging health information. The emerging guide is intended to make AI involvement observable when AI creates, updates or otherwise influences health data represented in FHIR.
The guide currently supports several levels of observability. At the simpler end, coded tagging can identify data as AI-influenced. Richer provenance can record information about the AI model, context and input data and the process of human-machine interaction. The AI Provenance profile is specifically designed to record the use of an AI model in generating or enhancing FHIR resources, including information about the model, inputs and outputs.
The guide also provides examples in which an AI system authors clinical information and a human subsequently verifies it. That is a useful conceptual distinction for healthcare governance: the system or person that generated information does not necessarily have to be the same person who verified and accepted it.
Its maturity should not be overstated. As of 3 October 2026, HL7 identifies version 1.0.0-ballot as STU1 Ballot 1, with Trial-use status and Maturity Level 2, and states that no current official final version has yet been published.
Healthcare organizations may therefore study or pilot the approach, particularly where FHIR already forms part of their health-information architecture. It should not be presented as a finalized Vietnamese technical requirement or as a standard that every healthcare organization is currently required to implement.
A practical implementation approach for healthcare organizations in Viet Nam
Healthcare organizations do not need to wait for another regulation or a final interoperability standard before improving traceability. The first step is to identify where AI already creates, transforms, summarizes, extracts or contributes to information that enters the clinical record or materially affects a clinical workflow.
The organization can then determine which uses are significant enough to justify provenance and what information needs to remain reconstructable. This design is best treated as a multidisciplinary governance issue involving clinical leadership, quality and patient safety, health-information management, IT, cybersecurity, personal-data protection and whoever is responsible for AI governance. The objective is to avoid both a purely technical solution that clinicians cannot use and a purely clinical solution that cannot support later investigation.
Human-review states should be built into the workflow rather than assumed. If software creates a draft, the system should distinguish an unreviewed draft from information that has actually been reviewed and accepted. Clinically consequential AI-generated information should not move automatically into a finalized medical record merely because the technology is capable of doing so.
Organizations should also test retrospective traceability before routine deployment. One useful test is to select a record created during a pilot and ask: Can we identify which AI system contributed, which relevant version or deployment was involved, who reviewed the information, which part of the record was affected and where supporting technical information can be found?
If the organization cannot answer those questions for a clinically consequential use, it may have an operational AI system but still lack adequate provenance.
What this means for healthcare organizations in Viet Nam
The timing is particularly relevant because electronic medical records and healthcare AI are advancing at the same time. Circular No. 13/2025/TT-BYT required hospitals to implement electronic medical records no later than 30 September 2025, while other healthcare facilities within the specified scope have until 31 December 2026. The same framework requires systems to support recovery and retrieval of EMR information when necessary for treatment, inspection, research and health management.
At the same time, Decision No. 1671/QĐ-TTg of 28 August 2026 approved the National Artificial Intelligence Strategy to 2030 with a vision to 2045. Its healthcare agenda includes AI for diagnosis and treatment support, remote care, hospital management, patient-flow optimization, health-record management and coordination of healthcare resources.
As these two directions converge, healthcare organizations will increasingly need to answer a question that traditional medical-record governance was not designed around: when clinical information has been materially shaped by AI, can the organization still explain where it came from and what the human reviewer actually did?
The answer should not require clinicians to fill medical records with technical metadata, and it should not rely on a vague statement that “a human was involved.” A more workable approach preserves clear human responsibility for the clinical record, makes material AI contribution traceable, maintains proportionate structured provenance, links clinical information to richer technical information where needed and avoids storing unnecessary sensitive information simply for the sake of transparency.
The most useful question is therefore not, “Should every AI-assisted note have an AI label?”
It is, “If this information later matters for patient care, quality review or an incident investigation, can we reliably reconstruct how AI contributed and what the human reviewer did?”
That is the level of traceability healthcare organizations should begin designing for.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, health-information-management, cybersecurity, personal-data-protection or medical-device advice.
“AI provenance,” the organizational categories of AI contribution, the three-layer architecture and the practical minimum provenance model described in this article are healthcare-governance concepts used for professional discussion. Current Vietnamese law reviewed for this article does not establish a universal requirement that every AI-assisted or AI-generated clinical note carry a visible AI label or contain the specific provenance fields proposed here.
Viet Nam's medical-record, AI-transparency, AI-risk and personal-data-protection requirements apply according to their respective legal scope. Healthcare organizations should distinguish mandatory legal requirements from internal controls adopted for patient safety, quality assurance, traceability and accountability.
The HL7 AI Transparency on FHIR Implementation Guide is an emerging trial-use interoperability approach. As of 3 October 2026, it is not a finalized Vietnamese regulatory requirement. References to HL7 and NIST are included as technical and governance considerations and should not be interpreted as creating legal obligations in Viet Nam unless separately made applicable through Vietnamese law, regulation, contract, certification requirement or another binding basis.
References
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.
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.
Office of the National Assembly of Viet Nam. Consolidated Document No. 26/VBHN-VPQH on the Law on Medical Examination and Treatment. Consolidated 26 February 2026.
Ministry of Health of Viet Nam. Circular No. 32/2023/TT-BYT Detailing Certain Provisions of the Law on Medical Examination and Treatment. Article 52 addresses medical-record use and documentation.
Ministry of Health of Viet Nam. Circular No. 13/2025/TT-BYT Providing Guidance on Implementation of Electronic Medical Records. Issued 6 June 2025; effective 21 July 2025.
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 Certain Articles and Measures for Implementation of the Law on Personal Data Protection. Issued 31 December 2025; effective 1 January 2026.
Prime Minister of Viet Nam. Decision No. 1671/QĐ-TTg Approving the National Artificial Intelligence Strategy to 2030, Vision to 2045. Issued 28 August 2026.
National Institute of Standards and Technology. HL7 AI Transparency on Fast Healthcare Interoperability Resources (FHIR). Project updated 15 September 2026.
HL7 International. AI Transparency on FHIR Implementation Guide, Version 1.0.0-ballot — STU1 Ballot 1. Trial-use; Maturity Level 2.