Preventing Wrong-Patient Errors in Clinical AI in Viet Nam: Patient Identification, Record Matching and Safe Attribution of AI Outputs
PRACTICAL FRAMEWORKS & TOOLS
Preventing Wrong-Patient Errors in Clinical AI in Viet Nam: Patient Identification, Record Matching and Safe Attribution of AI Outputs
PRACTICAL FRAMEWORKS & TOOLS
Last reviewed: 10 October 2026
A hospital introduces an AI application to help identify abnormalities on chest X-rays. The system analyzes an examination and produces a clinically plausible result. The result is transferred to the hospital information system, where a physician can review it alongside other clinical information.
There is one problem: the result has been associated with the wrong patient record.
The AI may have correctly analyzed the image it received. The error occurred because the examination, patient record and AI output were not linked correctly when information moved between systems.
This distinction matters. Healthcare organizations often focus on whether AI produces accurate findings, recommendations or risk scores. Yet even an accurate output can create a serious patient-safety risk if it is attributed to the wrong person, the wrong encounter or the wrong diagnostic order.
Similar problems can arise when an AI application prepares a clinical summary using information from another patient's record, or when an AI agent retrieves the correct information but writes its output into an incorrect electronic medical record.
For healthcare organizations in Viet Nam, safe AI implementation therefore requires more than evaluating clinical performance. It also requires confidence that the information entering the AI system belongs to the intended patient and that its output remains correctly associated with the relevant clinical activity.
The question is not simply whether the AI result is correct. It is whether that result belongs to the patient for whom it is being used.
Patient identification and clinical record matching are different responsibilities
Accurate patient identification has long been a fundamental patient-safety concern. Wrong-patient errors can affect medication administration, laboratory testing, diagnostic procedures and treatment decisions, whether AI is involved or not.
AI introduces additional points at which information may become incorrectly associated with a patient's care.
A typical clinical AI workflow may involve several distinct elements: the patient, the clinical encounter, a diagnostic or treatment order, the underlying clinical information and the AI-generated output. These elements are related, but they are not interchangeable.
A patient identifier helps establish whose information is being considered. An encounter represents a particular visit or episode of care. An order identifies the investigation or intervention requested. The AI output represents information generated from specified inputs for a defined clinical purpose.
A hospital may correctly identify a patient while inadvertently linking an AI result to an earlier encounter. An imaging examination may belong to the correct person but be associated with the wrong order. A result may also be generated from the correct examination and subsequently displayed within another patient's record because of an integration error.
These situations become particularly relevant when a patient has several encounters within a short period, receives multiple examinations of the same type or moves between inpatient and outpatient services.
A reliable process must therefore establish more than whether the correct name or patient number appears on a screen. It should support verification that the patient, encounter, relevant order and AI output belong together.
In this article, patient identity integrity refers to the reliability of those relationships throughout an AI-enabled clinical workflow. It is used as a practical patient-safety concept, not a separate legal category defined under Vietnamese law.
Where wrong-patient errors can enter an AI-enabled workflow
Patient identification errors do not necessarily originate within an AI application.
An error may begin during registration, when two patients have similar demographic information, when duplicate records are created or when clinical information is associated with an incorrect encounter.
These problems can occur in conventional healthcare information systems. Introducing AI does not automatically create them, but it may allow incorrect information to pass through additional processing steps without being recognized.
For example, an AI application may receive laboratory information with a valid patient identifier but an incorrect encounter reference. An imaging system may process a study correctly while the receiving application associates the result with another diagnostic order.
The problem becomes harder to detect when information moves automatically among several systems.
A hospital may use separate hospital information, laboratory, radiology, image archiving and electronic medical record systems. An AI application may connect to one or several of these systems. Each transfer creates an opportunity for the relationship between patient, encounter, order and result to be preserved correctly or disrupted.
AI agents introduce another concern. Unlike applications that simply analyze a predefined input and return an output, some agents can retrieve records, select documents or execute actions through connected software tools.
If an agent receives an incorrect record identifier, it may continue processing without recognizing the inconsistency. Depending on its authorized capabilities, it might then write information into the wrong record or initiate a downstream action using the incorrect patient context.
The AI model may have interpreted the medical information correctly. The failure may instead lie in how the information was selected, associated, transferred or stored.
This is why clinical AI implementation needs to consider the reliability of the entire workflow rather than treating the model as an isolated component.
What recent evidence tells us about identity-related risks
A study by Klang and colleagues, published in the International Journal of Medical Informatics in September 2026, examined whether clinical AI agents could recognize inconsistencies in patient identity information [1].
The researchers developed a simulated electronic health record environment using publicly available MIMIC-IV emergency department data. Six AI models were evaluated while performing a task involving the extraction of ICD-10-CM codes from visit information and writing those codes into patient records.
The evaluation involved approximately 1.2 million tool calls across several experimental conditions. Each run included ten visits from one patient, with one visit altered in the relevant experimental conditions. The alterations included replacing an entire patient header with another patient's information, changing one digit in a medical record number or modifying the patient's age.
The agents were permitted to withhold writing when uncertain. Nevertheless, they frequently continued processing records containing the introduced inconsistencies.
For example, GPT-4.1 withheld writing in 17.4% of runs involving a fully exchanged patient header. Withholding was much less frequent when the researchers introduced subtler changes to medical record numbers or patient ages.
In this experiment, withholding a write was used as an operational measure of possible mismatch detection. It does not necessarily establish that the agent recognized the identity problem or correctly reasoned about its significance.
The researchers emphasized that the central risk was misbinding, associating otherwise correctly processed information with an incorrect patient record, rather than necessarily extracting the wrong clinical codes.
These findings are important, but their limitations must be recognized. The study evaluated AI agents in a controlled, simulated environment. It did not measure actual wrong-patient incidents, clinical harm or the performance of all AI applications used in hospitals.
Its results therefore cannot be used to estimate how often AI causes patient identification errors in Vietnamese healthcare organizations.
Nevertheless, the study demonstrates why healthcare providers should not assume that an AI agent will recognize identity inconsistencies simply because it performs its underlying information-processing task successfully.
Earlier research on electronic health records provides additional context.
In a randomized controlled trial published in the Journal of the American Medical Informatics Association in 2013, Adelman and colleagues evaluated interventions designed to reduce electronic orders being placed in the wrong patient's record. Requiring clinicians to re-enter patient identification information reduced wrong-patient ordering events more than a simpler confirmation alert [2].
A later randomized clinical trial by Adelman and colleagues, published in JAMA in 2019, examined whether restricting clinicians to one open patient record would reduce wrong-patient orders. The researchers found no statistically significant reduction compared with allowing several records to remain open. Most ordering sessions in the unrestricted group involved only one open record, limiting what the study could establish about the effect of simultaneous record access [3].
These studies examined electronic ordering rather than AI-generated outputs. They remain relevant because they show how patient identification safety depends on the interaction between software design, workflow conditions and human behavior.
Together, the evidence supports the need for explicit identity safeguards. Neither AI capability nor a particular interface design should be assumed to prevent wrong-patient errors without evaluation.
Distinguish the original error from the failure to detect or contain it
When an AI output is discovered in the wrong patient record, one of the most important questions is where the incorrect association began.
Not every wrong-patient event involving AI is caused by the AI model.
An error may already exist in the source information. For example, a laboratory specimen may have been registered under the wrong patient, or an examination may have been associated with an incorrect encounter before the data reach the AI application. The AI may then process that information exactly as supplied.
A second possibility is that the source information was correct, but an interface or integration process linked it incorrectly while transferring data between systems.
In this situation, the AI may generate an accurate result from the correct examination, while the surrounding software associates that result with another patient's record.
A third possibility involves the AI application or agent itself. An agent may retrieve information using an incorrect reference, fail to recognize conflicting identifiers or select the wrong destination when executing a write action.
The distinction between the source of an error and the failure to detect it is particularly important.
An incorrect identifier may originate in a registration system. The AI application may not have created the error, but it may still fail to identify a detectable inconsistency and allow the information to continue through the workflow. A receiving system may then release the result without recognizing the mismatch.
These failures occur at different points and may involve different responsibilities.
There is also an important limitation to checking identifiers alone. Information may be consistently labeled with the wrong patient's identifiers from the beginning. In that situation, a system could find no technical mismatch even though the underlying clinical information belongs to someone else.
This is why reliable source identification, specimen and examination verification, record matching and downstream attribution controls need to complement one another.
A patient-safety investigation should establish where the information first became incorrect, which systems subsequently handled it, what safeguards were available and why the error reached the point at which it was discovered.
Corrective action should then address the actual causes rather than automatically attributing every wrong-patient event to the AI model.
Verify the relationship between patient, encounter, order and AI output
A central requirement for safe AI-enabled workflows is the ability to establish that an output belongs to the correct patient and clinical activity.
A patient identifier provides part of that assurance, but it is not sufficient by itself.
For diagnostic imaging, relevant references may include the patient identifier, accession number, examination date and unique imaging study identifier, depending on the systems in use. Laboratory workflows may rely on order and specimen identifiers that connect the original request, collected sample and reported result.
The DICOM standard includes identifiers such as Patient ID, Accession Number and Study Instance UID, each serving a different purpose in identifying and organizing medical imaging information [4].
Similarly, HL7 FHIR provides resources and references that can represent relationships among patients, encounters, service requests and diagnostic reports [5].
These standards support the representation and exchange of clinical information. They do not automatically guarantee that a healthcare organization has linked the correct records, nor are they universal statutory requirements for every AI application used in Viet Nam.
Healthcare organizations should understand which identifiers their systems actually use, how those identifiers are assigned and whether they remain consistent when information crosses system boundaries.
An identifier may be unique within one healthcare system without being unique across an entire network. Local medical record numbers, encounter identifiers and order numbers can also have different scopes of validity.
A reliable workflow should therefore depend on verified relationships among the relevant records, not simply on matching one visible number.
The same principle applies to AI-generated summaries, risk scores or recommendations that do not originate from a diagnostic order. The organization should be able to establish which patient records and clinical period formed the basis of the output and where that output is intended to be used.
Detect inconsistent information before it affects care
Identity discrepancies may not appear as obvious differences between patient names.
Two people may have similar names and dates of birth. A name may be recorded differently across systems because of spelling variations, Vietnamese diacritics or differences in how name components are entered.
Patients may also have separate identifiers at different healthcare facilities. Duplicate records can arise within one organization and may remain undetected until a subsequent encounter or record reconciliation process.
Healthcare organizations should therefore consider how their systems identify inconsistencies when patient, encounter or order information does not match the expected relationship.
A mismatch may involve an order associated with another patient, a result returned for a different encounter or an AI output that cannot be reliably associated with the examination originally submitted for analysis.
Sometimes the problem is not contradictory information but missing information. An output may be returned without the reference needed to determine which of several similar examinations it belongs to.
The absence of a visible contradiction does not establish that the association is correct.
Where the required relationship cannot be confirmed, the workflow should support clarification rather than automatically assigning the result to the most plausible record.
This is particularly important in automated processing, where numerous outputs may pass through interfaces without individual review.
Organizations should also consider the effect of delayed messages, repeated transmissions, corrected orders, record merges and changes made while an AI task is still running.
A result may have been associated with an order when processing began, but additional verification may be necessary if the relevant record or encounter has changed before the result returns.
International guidance provides useful reference points. The World Health Organization's Patient Identification patient-safety solution emphasizes reliable identification practices, including the use of at least two appropriate patient identifiers in relevant care processes [6].
The US 2025 SAFER Guide: Patient Identification also recommends practices to reduce the risk of duplicate records, patient selection errors and incorrect associations between patients and electronic health information [7].
These documents are international patient-safety guidance, not binding Vietnamese requirements for AI systems. Their relevance is that patient identification must be supported by reliable organizational processes as well as technology.
A practical Patient Identity and AI Output Attribution Review
Healthcare organizations in Viet Nam can organize their safeguards through a proposed Patient Identity and AI Output Attribution Review.
This approach is intended to support patient safety in AI-enabled workflows. It is not a prescribed Ministry of Health procedure, an internationally validated assessment instrument or an independent certification standard.
The review concentrates on three points: before clinical information enters AI processing, when the AI output is associated with a clinical record, and before that output is used in a patient-specific decision.
The three points serve different purposes. The first seeks to prevent incorrectly attributed information from entering the process. The second checks whether the output has returned to the intended clinical context. The third provides assurance before the information influences care.
First verification point: Before AI processing begins
The first verification concerns the clinical information selected for AI processing.
The organization should establish that the data belong to the intended patient and relevant encounter. Where the application analyzes a diagnostic examination, the input should also be associated with the correct order or study.
The method of verification should reflect how information enters the AI application.
In an integrated hospital system, some checks may be performed automatically using established patient, encounter and order references. For applications requiring manual uploads, the organization may need a different process to ensure that the selected examination or document belongs to the intended patient.
The workflow should identify when essential references are missing, inconsistent or insufficient to establish the required association.
If a significant discrepancy remains unresolved, patient-specific AI processing that could produce an incorrectly attributed clinical result should not proceed through the normal automated pathway.
Where the AI model does not need direct patient identifiers to perform its task, an appropriately controlled system can maintain the relationship between the input and its originating clinical record without unnecessarily exposing identifiable information to the model.
The important requirement is that the association remains reliable and can be verified when the output returns.
This stage should not depend solely on an AI model deciding whether patient information looks plausible. The organization needs controls that establish the input's relationship with verified source records.
Second verification point: When the AI output returns to the clinical system
The second verification concerns the relationship between the generated output and the original request.
The receiving system should establish that the result corresponds to the intended input and is being associated with the appropriate patient, encounter and order.
This is particularly important for applications that return results after a delay or exchange information through several systems.
An AI result may be clinically accurate but still become unsafe when it is attached to the wrong record.
The organization should therefore examine how the receiving system identifies the original processing request and verifies the destination of the returned output.
Where feasible, this relationship should be maintained through verified references controlled by the clinical information systems rather than relying on patient information generated or inferred by the AI model.
If the necessary association cannot be confirmed, the output should be withheld from automatic clinical release or patient-record updates until the discrepancy is resolved.
This control becomes especially important for AI agents capable of writing clinical notes, updating records, preparing orders or triggering downstream actions.
The successful completion of an AI processing task does not establish that the result has reached the correct clinical destination.
Third verification point: Before the output is used for clinical care
The third verification occurs before an AI-generated output influences a patient-specific clinical decision.
The clinician or other authorized user should be able to establish that the result displayed belongs to the patient and clinical activity under review.
For example, an AI-generated imaging finding should correspond to the relevant examination. A risk score should use the intended patient's information and appropriate clinical period. An AI-generated summary should not contain information from unrelated records.
The verification process should reflect the clinical significance of the output.
An application used for low-consequence administrative support may require different controls from one whose output could influence urgent diagnosis, medication prescribing or treatment decisions.
Where the output has substantial clinical influence, the responsible professional should have access to the relevant underlying information and sufficient context to evaluate both the result and its attribution.
Professional review is important, but it should not be the only safeguard against incorrect patient linkage.
Clinicians may reasonably expect that a result displayed within the patient's electronic record has already been associated with the correct person and clinical activity. The workflow should therefore incorporate appropriate verification before that result becomes available for use.
These three verification points are complementary. They should not be reduced to a single confirmation that the correct name appears on a screen, nor do they eliminate the need for reliable identification during registration, specimen collection or diagnostic examinations.
When an AI output should be withheld from routine clinical use
Healthcare organizations should define the circumstances in which an AI-generated output cannot safely proceed through the normal release or clinical-use pathway.
An unresolved mismatch between patient identifiers is one such circumstance. Others include a result associated with the wrong encounter or order, an output that cannot be linked reliably to its original input, or uncertainty created by duplicate or recently merged patient records.
For AI applications with automated write or action capabilities, uncertainty about the intended destination should prevent the relevant patient-specific action from proceeding until an authorized process has resolved the association.
Missing information should not automatically be treated as a verified match. Nor should the system silently select another patient or encounter simply because it appears to be the closest available match.
The organization may use a controlled status such as identity verification pending to indicate that an output is not yet suitable for routine clinical release or use. This is a proposed operational safeguard, not a status prescribed specifically by Vietnamese law.
The response should reflect the nature of the discrepancy. Some technical inconsistencies may be resolved through an established automated process, while others require clinical, medical records or information technology review.
The organization should also determine who has the authority to resolve the mismatch and release the output for clinical use. This responsibility should be established before routine implementation rather than decided only after an error occurs.
An important qualification is that withholding an uncertain AI output must not unnecessarily delay urgent patient care.
Where a patient requires immediate assessment or treatment, healthcare professionals should be able to continue through an appropriate alternative process using verified clinical information rather than depend on the unresolved AI output.
The purpose is to prevent incorrectly attributed information from influencing care, not to interrupt essential healthcare services.
When an AI result has already been attached to the wrong record
Healthcare organizations also need a clear response when incorrect attribution is discovered after an AI output has entered the clinical workflow.
The immediate priority is to determine whether the error could affect patient care.
If an AI result has been linked to the wrong record, the organization should prevent further inappropriate use or distribution while maintaining access to the clinical information needed for safe care.
The relevant clinical team should assess whether the incorrect result has already been viewed, communicated or used to make a decision.
This may require identifying both the patient whose information was incorrectly attributed and the patient whose record received that information.
The assessment should consider whether the result influenced a diagnosis, investigation, medication decision, referral or other aspect of care. Where clinically important decisions may have been affected, the responsible professionals should determine whether reassessment or timely communication is needed.
The organization should also consider whether the patient whose information was misattributed has been deprived of a result that should have been available for their own care. A wrong-patient association may create risk for more than one person.
The organization must then determine where the incorrect association originated. It may have begun during registration, specimen or examination identification, system integration, AI processing or the return of the output to the receiving record.
The response should not automatically assume that the AI model caused the original error.
Correcting the clinical record requires appropriate control and traceability. Simply deleting or silently replacing incorrect information may make it difficult to establish what clinicians previously saw or how earlier decisions were made.
Depending on the information system and applicable requirements, correction may involve an authorized amendment, controlled record reconciliation or another documented process.
The organization should preserve sufficient information to establish what was originally recorded, what was corrected, who authorized the change and when it occurred.
Where the incorrect result was transmitted to another clinical system, healthcare provider or patient-facing portal, the organization should determine whether those recipients also require correction or notification.
Patient confidentiality and personal-data protection must be considered, since information incorrectly associated with another person's record may have been disclosed to someone without authorization.
The organization should assess whether the event meets the applicable criteria for a medical incident, personal-data breach or another reportable event. These determinations depend on the facts and legal requirements; a detected technical mismatch should not automatically be classified in the same way as a clinically harmful event or confirmed unauthorized disclosure.
The objective is broader than repairing one record. It is to protect affected patients, prevent further reliance on incorrect information and address the process failure that allowed the error to occur.
Test identity integrity before AI goes live
Patient identity integrity should be evaluated before an AI application is introduced into routine clinical use.
Testing should go beyond demonstrating that the system can return an output for a valid patient record.
The organization should examine how the workflow behaves when identity information is missing, inconsistent or unexpectedly changed.
Representative cases may include patients with similar names, duplicate medical records, several encounters for one patient and multiple diagnostic orders performed close together.
Other scenarios may involve an altered patient identifier, a returned result without the expected order reference, or a delayed AI response after the originating clinical record has changed.
For applications requiring manual uploads, testing should assess the possibility of selecting the wrong document or examination and whether the workflow offers an effective opportunity to identify the mistake.
For AI agents or other applications capable of updating records, testing should determine whether an incorrect or uncertain patient association prevents the relevant write action.
A successful evaluation should establish more than whether a warning message appears. It should determine whether the output is prevented from entering an incorrect record, whether the discrepancy reaches the appropriate personnel and whether there is an authorized way to resolve it.
Testing should also consider repeated messages, corrections, record merges and other realistic workflow conditions that may affect attribution.
The depth of evaluation should reflect the consequences of an incorrect output and the degree of automation involved.
A read-only AI application whose results undergo professional review may require different safeguards from an AI agent able to update clinical documentation or initiate downstream actions.
Vendor demonstrations can provide useful information about intended functionality, but they should not replace evaluation of the integrated workflow in the healthcare organization's own environment.
The hospital needs to know whether its combination of software, interfaces, records and clinical procedures can maintain correct patient attribution under realistic operating conditions.
Monitor identity-related safety after implementation
Patient identity integrity should remain part of ongoing AI safety monitoring after go-live.
A workflow that operates correctly during initial testing may encounter different conditions in routine care. Changes to registration procedures, information systems, interfaces, software versions or clinical operations may affect how patient information is associated with AI outputs.
Healthcare organizations should therefore examine whether identity-related discrepancies occur and whether existing controls detect them before they influence patient care.
Relevant measures may include confirmed incorrect patient or encounter associations, mismatches identified before output release, outputs withheld because attribution could not be verified and cases in which incorrectly attributed information reached clinical use.
The time needed to resolve significant discrepancies may also be useful.
However, these indicators should be interpreted carefully.
An increase in detected mismatches does not necessarily mean the system has become less safe. It may indicate that controls are successfully identifying problems that previously went unnoticed.
Conversely, an absence of reported mismatches does not establish that every AI result has been correctly attributed.
Monitoring should distinguish confirmed wrong-patient events from technical alerts that did not represent an actual identity error. It should also consider whether safeguards functioned as intended, not merely whether an incident was ultimately recorded.
Where feasible, organizations should examine patterns by clinical service, AI application, interface or type of workflow. Repeated discrepancies at one integration point may require a different response from isolated errors involving manual document selection.
Quality and patient-safety personnel should work with clinicians, medical records staff and information technology teams to assess the significance of these findings and determine whether corrective action is effective.
The purpose is not simply to count identity-related incidents. It is to establish whether safeguards continue to maintain reliable patient attribution as clinical AI systems and workflows evolve.
What Viet Nam's legal framework means for patient identity integrity
Viet Nam's Law on Medical Examination and Treatment No. 15/2023/QH15, as amended, provides an important foundation for clinical documentation, professional accountability and patient-safety responsibilities.
Article 69 addresses medical records, including their creation, updating, storage and confidentiality. It also recognizes the equivalent legal value of paper-based and electronic medical records [8].
Article 71 establishes a framework for preventing medical incidents through identification, reporting, analysis of causes and measures intended to reduce recurrence. It also establishes responsibilities for medical incident prevention within healthcare establishments [8].
These provisions are relevant when incorrectly attributed AI information enters a medical record or contributes to a situation that may affect patient care.
Clause 10 of Article 7 prohibits erasing or altering medical records for the purpose of misrepresenting information about medical examination and treatment, as well as creating false medical records or other false documentation of examination and treatment results [8].
This prohibition must be distinguished from the legitimate correction of an error.
An accidental wrong-patient association is not automatically equivalent to deliberately falsifying a medical record. Likewise, correcting an error through an authorized, traceable process should not be confused with altering records to conceal or misrepresent clinical information.
For healthcare organizations, the practical priority is to ensure that record corrections preserve accuracy and comply with relevant legal and professional requirements.
Circular No. 13/2025/TT-BYT, issued on 6 June 2025 and effective from 21 July 2025, provides guidance on implementing electronic medical records.
Article 1 establishes principles concerning the creation, updating, management and use of electronic medical records. Clause 3 of Article 1 provides for connecting electronic medical record information with the personal identification numbers of Vietnamese citizens and foreign nationals who have been issued electronic identification accounts under the applicable identification legislation [9].
This provision is relevant to the development of digital medical records in Viet Nam.
However, linking an electronic medical record to a personal identification number does not automatically establish that a particular laboratory result, imaging examination, clinical encounter or AI output has been associated correctly with that person.
Healthcare organizations still need to maintain the integrity of relationships between the patient, encounter, order and clinical information.
Some circumstances may also require special handling, such as emergency care involving an unidentified patient or situations where identification information is incomplete or awaiting reconciliation.
The clinical workflow should accommodate such circumstances through appropriate authorized processes rather than assume that every patient record will always contain complete verified identifiers.
Where AI processing involves identifiable patient information, the Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP, both effective from 1 January 2026, are also relevant [10,11].
Their application depends on the actual processing activities, including how information is collected, shared, stored, accessed and disclosed.
Incorrect attribution may raise privacy concerns when one person's information becomes available through another patient's record. Whether an event triggers notification, reporting or other legal obligations must be assessed according to the applicable requirements and circumstances.
These legal instruments provide important foundations for patient safety, medical-record integrity and personal-data protection. They do not prescribe one universal AI-specific workflow requiring all healthcare organizations to implement the same three verification points described in this article.
The proposed Patient Identity and AI Output Attribution Review should therefore be understood as a practical organizational method intended to strengthen existing controls, not a replacement for legal or professional obligations.
Making the approach workable in Vietnamese healthcare organizations
Healthcare organizations do not necessarily need identical technology or additional manual confirmation for every AI result.
The appropriate safeguards should reflect the clinical workflow, the information systems involved and the potential consequences of incorrect attribution.
For a hospital integrating AI with its HIS, RIS/PACS or laboratory information system, the priority may be verifying how patient, encounter and order references are maintained across those systems.
For a smaller clinic using an AI application that requires manual uploads, the priority may be ensuring that the correct clinical document or examination is selected and that the resulting output is reviewed within the intended patient context.
Organizations should also define responsibility for resolving identity discrepancies before routine deployment.
Medical records and registration personnel may be needed to investigate duplicate or incorrectly linked records. Information technology staff may need to examine system interfaces and transaction logs. Clinical professionals should assess whether patient care has been affected. Quality and patient-safety teams can coordinate investigation, learning and corrective action.
Where an external AI vendor is involved, the organization should understand how the application receives and preserves references to patients and clinical activities, how outputs are returned and what happens when essential information is missing or inconsistent.
This is especially important when the application can automatically update records or initiate actions.
Controls should remain proportionate to risk. Requiring staff to repeat unnecessary confirmation steps for every low-consequence activity may increase workload without meaningfully improving safety.
However, an AI result capable of influencing a significant clinical decision should not depend solely on the assumption that information appearing in the electronic medical record must belong to the correct patient.
A workable approach combines reliable system design, appropriate professional checks and a clear process for withholding, correcting or safely bypassing uncertain AI outputs when necessary.
From accurate AI outputs to reliable patient care
Clinical AI can generate accurate findings, predictions and recommendations while still contributing to unsafe care if those outputs are associated with the wrong patient or clinical activity.
The problem is not limited to the model's ability to interpret medical information. It also involves how information reaches the model, how the output returns to the clinical system and how that information is subsequently used.
For healthcare organizations in Viet Nam, maintaining patient identity integrity should therefore be part of the entire AI-enabled workflow.
Before AI processing, the organization should establish that the selected information belongs to the intended patient and clinical activity. When the output returns, its relationship with the originating information and intended destination should be verified. Before clinical use, the result should remain appropriately traceable to the relevant patient, encounter, order or examination.
When that relationship cannot be reliably established, the workflow should support clarification, controlled withholding and safe alternative care.
Where incorrect attribution has already occurred, the organization must address both the clinical consequences and the integrity of the affected records.
These safeguards can be incorporated into existing clinical quality, medical records, information technology and patient-safety activities rather than developed as an entirely separate governance system.
A clinically accurate AI output is useful only when the healthcare organization can also establish that it belongs to the right patient, the right encounter and the right clinical context.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, medical-record, cybersecurity, personal-data-protection, accreditation or quality-management advice.
The Patient Identity and AI Output Attribution Review described here is a proposed practical organizational method. It is not a defined statutory category, prescribed Vietnamese Ministry of Health procedure, WHO certification requirement or internationally validated clinical AI assessment standard.
The three verification points, conditions for withholding AI outputs, monitoring measures and correction approaches represent practical patient-safety recommendations. They should not be interpreted as identical mandatory requirements for every AI application or healthcare organization under Vietnamese law.
Applicable obligations depend on the healthcare service, the AI system's intended purpose and functionality, the degree of automation, the clinical information systems involved and relevant professional and legal requirements.
The international studies cited involve different technologies, clinical settings and evaluation methods. In particular, the 2026 clinical agent study used a simulated electronic health record environment. Its results should not be interpreted as estimates of actual wrong-patient incidents or patient harm in Vietnamese healthcare organizations.
International patient identification guidance and health information technology standards provide useful reference points, but their application should reflect the organization's clinical and regulatory context.
The proposed approach is intended to strengthen existing patient-safety and clinical governance responsibilities, not replace professional judgment or applicable requirements for managing medical records and clinical incidents.
References
Klang E, Glicksberg BS, Gorenshtein A, et al. Clinical Agents Fail Silently on Patient Identity. International Journal of Medical Informatics. 2026;218:106514. DOI: 10.1016/j.ijmedinf.2026.106514.
Adelman JS, Kalkut GE, Schechter CB, et al. Understanding and Preventing Wrong-Patient Electronic Orders: A Randomized Controlled Trial. Journal of the American Medical Informatics Association. 2013;20(2):305–310. DOI: 10.1136/amiajnl-2012-001055.
Adelman JS, Applebaum JR, Schechter CB, et al. Effect of Restriction of the Number of Concurrently Open Records in an Electronic Health Record on Wrong-Patient Order Errors: A Randomized Clinical Trial. JAMA. 2019;321(18):1780–1787. DOI: 10.1001/jama.2019.3698.
National Electrical Manufacturers Association. Digital Imaging and Communications in Medicine (DICOM) Standard. PS3.3: Information Object Definitions. Patient and Study Identification Attributes.
Health Level Seven International. HL7 FHIR Specification, Release 5. Patient, Encounter, ServiceRequest and DiagnosticReport Resources. 2023.
World Health Organization Collaborating Centre for Patient Safety Solutions. Patient Identification. Patient Safety Solutions. Volume 1, Solution 2. Geneva: World Health Organization; May 2007.
Office of the National Coordinator for Health Information Technology. 2025 SAFER Guide: Patient Identification. United States Department of Health and Human Services; 2025.
National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment, as amended. Adopted 9 January 2023; effective 1 January 2024.
Ministry of Health of Viet Nam. Circular No. 13/2025/TT-BYT Providing Guidance on the 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 a Number of Articles and Measures for Implementation of the Law on Personal Data Protection. Issued 31 December 2025; effective 1 January 2026.