Governing AI-Assisted Emergency Department Triage in Viet Nam: Acuity Assessment, Ongoing Reassessment and Patient Safety
IMPLEMENTATION INSIGHTS
Governing AI-Assisted Emergency Department Triage in Viet Nam: Acuity Assessment, Ongoing Reassessment and Patient Safety
IMPLEMENTATION INSIGHTS
Last reviewed: 23 October 2026
A patient arrives at a busy emergency department complaining of dizziness and abdominal discomfort. The initial vital signs do not immediately suggest a life-threatening condition. An artificial intelligence system reviews the available information and recommends a relatively low-priority triage category.
The triage nurse, however, notices that the patient looks unusually pale and has difficulty explaining when the symptoms began. After further assessment, the nurse decides that an earlier clinical review is needed.
Elsewhere in the waiting area, another patient who was initially considered stable develops worsening breathlessness. The original triage decision may have been reasonable when the patient arrived, but the clinical situation has changed. Continuing to rely on the initial classification could delay necessary care.
These situations illustrate two challenges that emergency departments face every day: determining clinical urgency when information is incomplete and recognizing when an earlier assessment is no longer appropriate.
Artificial intelligence may help clinicians identify relevant information, recognize patterns and make more consistent initial assessments. However, even a technically accurate prediction cannot guarantee that a patient has been assigned the correct priority. It also cannot establish that the priority will remain appropriate throughout the waiting period.
For hospitals in Viet Nam, evaluating AI-assisted emergency triage therefore requires attention to more than the algorithm's performance. The technology needs to be considered within the actual clinical process, from the patient's arrival and initial assessment to reassessment, recognition of changing clinical needs and timely escalation.
Predictive accuracy does not necessarily establish safer emergency triage.
A model may predict hospital admission, intensive care use or mortality with considerable accuracy without reliably identifying which patient needs to be seen first. Similarly, an application that reduces the time needed to complete triage documentation does not necessarily reduce harmful delays in assessment or treatment.
The more important question is whether AI helps emergency teams assign appropriate priorities, maintain awareness of patients who are waiting and respond promptly when new clinical findings require greater attention.
Emergency triage is a decision about clinical priority
Emergency department triage is intended to determine how urgently a patient requires assessment or intervention. It allows emergency teams to identify those needing immediate attention while organizing care for others according to their clinical needs.
Triage becomes particularly important when several patients arrive at the same time or when demand exceeds the department's available capacity. Its primary purpose, however, is not to establish a final diagnosis or predict whether the patient will eventually be admitted to hospital.
A patient with an immediately life-threatening condition may require resuscitation without waiting for a complete triage interview or an AI-generated recommendation. Another patient may be clinically stable but need several investigations before a diagnosis or discharge decision can be made. The second patient's greater use of diagnostic resources does not necessarily mean that the patient should receive attention first.
Internationally, emergency departments use structured approaches such as the Emergency Severity Index (ESI), Manchester Triage System (MTS), Canadian Triage and Acuity Scale (CTAS) and Australasian Triage Scale (ATS). These systems differ in their methods, but each is intended to support decisions about clinical priority.
In their joint policy statement revised in January 2025, the American College of Emergency Physicians (ACEP) and Emergency Nurses Association (ENA) support the adoption of scientifically validated triage scales. They also emphasize that the triage process should not itself create avoidable delays in care [1].
The Australasian College for Emergency Medicine (ACEM) similarly describes triage as a process of prioritizing patients according to clinical urgency. Its guidance recognizes the need to balance rapid assessment with sufficient clinical information and to reassess patients whose condition may change while waiting [2].
These international documents provide useful professional references. They do not establish a single mandatory triage system for every healthcare organization in Viet Nam.
For AI implementation, the starting point should therefore be the specific clinical decision that the technology is intended to support. A system developed to recommend an initial acuity category should be evaluated for that purpose, rather than assumed to be clinically effective because it performs well on a related prediction task.
Three different questions that AI systems may attempt to answer
Technologies described as AI-assisted emergency triage systems do not necessarily perform the same function. Their intended purposes, inputs and outputs can differ substantially, and these differences matter when determining whether a product is suitable for clinical use.
Initial acuity assessment concerns how urgently a patient requires clinical attention. It considers presenting symptoms, observable clinical condition, relevant history, vital signs when available and other information needed to identify immediate or time-sensitive risks. Its purpose is to help determine an appropriate priority for assessment or intervention.
Prediction of clinical outcomes concerns the likelihood that a patient will experience an event such as hospital admission, ICU admission or short-term mortality. These predictions may provide useful information about illness severity, but they are not equivalent to determining how urgently the patient needs treatment at arrival.
Prediction of resource requirements estimates the investigations, procedures, treatments or healthcare services that a patient may need. This information can support operational planning, but expected resource use is not necessarily a direct measure of immediate clinical urgency.
The distinction requires particular care when considering the Emergency Severity Index. ESI incorporates anticipated resource use within parts of its established algorithm, after considering immediate life-saving needs and high-risk presentations. However, a separately developed machine learning model that predicts resource consumption cannot automatically be assumed capable of assigning clinically appropriate ESI categories.
Consider a patient with substantial chronic care needs who is likely to require hospitalization but is currently stable. An AI model may correctly predict a high probability of admission. Yet another patient with an evolving time-critical condition may need immediate assessment even if that patient is ultimately discharged after timely investigation and treatment.
The opposite situation is also possible. A patient may require urgent evaluation to exclude a serious condition but ultimately need few investigations or treatments. A model that relies heavily on subsequent resource use may underestimate the urgency of that presentation.
Hospital admission and ICU use are influenced by factors beyond the patient's condition at arrival, including treatment decisions, institutional practices and the availability of hospital beds.
These outcomes can help researchers understand associations between triage categories and subsequent events. They should not be treated as interchangeable measures of what the patient's priority should have been at first contact.
Before selecting an AI product, a hospital should therefore establish what the system is designed to predict, what information it requires and whether its outputs have been evaluated for the particular clinical decision it is expected to support.
Why strong predictive performance may not mean safer triage
AI studies commonly report accuracy, sensitivity, specificity and the area under the receiver operating characteristic curve (AUC). These measures describe different aspects of model performance, but none independently establishes that a technology improves emergency patient safety.
An AUC describes how well a model distinguishes between patients who do and do not experience a defined outcome across possible classification thresholds. It does not establish that a particular threshold is suitable for deciding who should receive emergency assessment first.
A model may demonstrate strong discrimination for ICU admission while still failing to recognize some patients who need immediate attention. It may also generate excessive high-priority recommendations when applied to a different clinical population.
The reference standard used to evaluate a model is equally important.
When AI-generated categories are compared with categories assigned by nurses or physicians, the resulting accuracy primarily measures agreement with those assessments. It does not independently establish that the reference category was clinically correct.
If the reference classifications contain errors or inconsistencies, reproducing them more closely is not necessarily an improvement in patient safety.
Retrospective studies also require attention to when information became available. A model evaluated using laboratory results, final diagnoses or other information recorded later in the emergency visit may perform well in research but may not have access to those inputs during initial triage.
A meaningful evaluation must therefore connect prediction to the actual decision being made. It should examine whether patients requiring urgent treatment are recognized, whether unnecessary escalation increases and whether the use of AI affects timely clinical assessment.
The consequences of incorrect recommendations matter at least as much as the overall prediction score.
Undertriage: when a patient receives too little priority
Undertriage occurs when a patient is assigned a lower level of urgency than their clinical condition warrants.
Its consequences can be serious because the patient may remain in the waiting area while others receive attention. The resulting delay may affect the timing of clinical examination, investigation, treatment or escalation.
Some time-sensitive conditions do not initially present with obvious signs of critical illness. Patients with evolving sepsis, acute coronary syndrome, stroke or other serious conditions may report nonspecific symptoms. Vital signs may be incomplete, only mildly abnormal or insufficient to convey the full level of concern.
AI may help identify some of these presentations, but it can also underestimate urgency when relevant information is missing, inaccurately recorded or inadequately represented in its development data.
An additional concern is how recommendations are displayed. A precise probability or clearly defined category may appear more authoritative than the available evidence justifies.
If an AI system recommends a lower priority, a clinician may become less inclined to investigate a concerning symptom or observation. This possibility deserves particular attention when software automatically populates a triage category or influences the patient's position in an electronic waiting queue.
Overall accuracy cannot adequately describe this risk. A system may classify most patients correctly while still missing an unacceptable proportion of those requiring immediate or time-sensitive intervention.
Hospitals should therefore examine clinically significant undertriage separately, including the circumstances of incorrect classifications and their potential consequences.
A small number of cases involving delayed recognition of serious illness may be more important for patient safety than a larger number of disagreements between clinically similar lower-priority categories.
The objective is not simply to improve agreement between AI and human assessments. It is to reduce the likelihood that a patient with urgent clinical needs is incorrectly left waiting.
Overtriage: when additional caution creates other pressures
Overtriage occurs when a patient is assigned a higher level of urgency than the clinical situation requires.
A cautious decision may be appropriate when information is incomplete and the consequences of missing a serious condition are substantial. Emergency triage often occurs before a diagnosis can be established.
However, excessive overtriage may create additional workload and consume limited high-acuity assessment capacity. In a crowded department, this can affect the timeliness of care for patients whose conditions are more urgent.
An AI model designed to minimize missed critical illness may generate many high-priority recommendations. Whether this represents an acceptable trade-off depends on how frequently unnecessary escalation occurs and how the department responds.
The issue is illustrated by research on conventional triage systems.
In 2025, Espejo and colleagues published a study in The Journal of Emergency Medicine simulating the application of ESI version 5 among 6,230 adult emergency department patients who had originally been triaged using ESI version 4 [10].
The newer version placed greater emphasis on identifying abnormal vital signs among patients initially considered lower acuity. In the simulation, 636 patients, or 10.2%, would have been moved to ESI level 2.
However, the investigators identified only 30 patients, or 0.5% of the study population, as having potential benefit according to the study's selected outcomes of ICU admission or death within 30 days.
This was a simulation rather than a prospective evaluation of ESI version 5 in routine practice. The definition of potential benefit was also limited to particular outcomes and did not capture every possible clinical benefit of earlier assessment.
Nevertheless, the findings illustrate an important trade-off. Increasing the sensitivity of a triage process can substantially expand the number of patients assigned high priority without necessarily demonstrating improved patient outcomes.
The same question applies when AI generates additional high-acuity recommendations.
Hospitals should examine whether increased escalation appropriately identifies patients in need or creates new delays and capacity problems elsewhere in the department.
Evaluating overtriage does not mean lowering priorities simply to improve efficiency. It means recognizing that safe emergency care requires appropriate allocation of attention while protecting patients from avoidable delays.
The initial triage decision is not the end of assessment
An initial triage category reflects the clinical information available at a particular point in time. It should not be treated as a permanent assessment of the patient's condition.
A patient may develop new symptoms, worsening pain, abnormal vital signs or changes in consciousness while waiting. Additional information from relatives or accompanying persons may also alter the level of concern.
Even when the original priority was appropriate, it may no longer reflect the patient's current needs.
This is an important limitation of AI systems that assess patients only when they arrive. A recommendation generated at 9:00 a.m. does not necessarily provide meaningful information about the patient's condition an hour later.
During busy periods, the original category may remain visible in the hospital information system while the patient continues to wait. If the organization does not have a reliable reassessment process, important changes may go unrecognized.
International professional guidance addresses this concern.
ACEM's policy on the Australasian Triage Scale calls for suitable staffing and processes to allow regular reassessment of patients who remain waiting, with revision of priority when their clinical features change [2].
Its policy on emergency medicine consultation standards also identifies reassessment, recognition of deterioration and appropriate responses to concerns raised by patients or families as elements of care while patients await further assessment or treatment [3].
Ireland's National Quality Improvement Framework for Emergency Department Triage (2025–2027) provides another useful perspective [4].
The framework seeks to restore triage to its core function of rapid clinical prioritization while strengthening the care provided immediately afterward. It distinguishes the initial triage encounter from the subsequent clinical activities required to manage patients safely while they remain in the emergency department.
This distinction is particularly relevant when considering digital technologies. Adding too many tasks to the initial triage encounter may delay prioritization, while a poorly organized post-triage process may leave patients exposed to risks that a faster initial assessment cannot resolve.
These international approaches reflect their respective healthcare systems. Their operational arrangements and performance targets should not automatically be treated as mandatory requirements in Viet Nam.
For Vietnamese hospitals considering AI-assisted triage, the broader lesson is that the safety of the waiting period must be evaluated alongside the accuracy of the initial classification.
Reassessment needs a defined place in the waiting process
A patient who has been triaged but has not yet received the required clinical assessment remains within an active care process.
The emergency department needs a reliable way to identify which patients require reassessment, when it should occur and who is responsible for responding to new findings.
The appropriate reassessment frequency depends on the patient's clinical condition, the triage framework being used and the organization's established procedures. Some international systems specify category-based expectations, but maximum waiting-time targets for treatment should not be confused with universal reassessment intervals.
A scheduled reassessment time must not become a reason to postpone attention when new symptoms or clinical changes suggest greater urgency.
For a patient whose condition remains stable, reassessment may confirm that the original priority continues to be appropriate. For another patient, new symptoms or abnormal observations may require immediate reconsideration.
The process should involve more than confirming that the patient is still physically present.
Depending on the clinical circumstances, reassessment may include asking whether symptoms have changed, observing breathing and consciousness, reassessing pain, repeating relevant vital signs or reviewing information that was unavailable during the initial encounter.
When these findings indicate greater urgency, the patient should receive appropriate clinical attention without waiting for an automated score to be updated.
A practical challenge is assigning responsibility.
During busy periods, staff performing initial triage may be occupied with newly arriving patients. Those already waiting can become less visible, particularly when no individual or team has clear responsibility for ongoing review.
Hospitals should therefore establish how waiting-area observation, reassessment, communication and escalation are organized within their actual staffing arrangements.
Shift changes and handovers also matter. Responsibility for a patient requiring further review should not become unclear simply because staff members change or the department becomes busier.
Digital systems may support these arrangements by displaying waiting times, identifying patients whose reassessment is due or drawing attention to those who have not received the expected review.
However, a reminder is useful only if it reaches the responsible staff member and leads to appropriate action.
A system that repeatedly generates notifications without a reliable clinical response may create the appearance of monitoring without meaningfully protecting patients.
The important outcome is not how many reminders are issued, but whether reassessments occur when indicated and whether changes in condition lead to timely intervention.
New warning signs must take precedence over an earlier AI recommendation
A patient may develop new clinical findings after initial triage, or existing symptoms may become more severe.
Worsening breathlessness, altered consciousness, new neurological symptoms, persistent or worsening chest discomfort, significant bleeding, collapse and other concerning changes may require prompt reassessment and escalation.
Their significance depends on the patient's clinical circumstances and the applicable emergency care protocols. These examples are not intended as a comprehensive triage checklist or a substitute for professional assessment.
Suppose a patient initially receives a lower-priority AI recommendation because the available symptoms and observations appear reassuring. While waiting, the patient becomes confused and increasingly difficult to engage.
The previous recommendation should not delay evaluation of the new condition.
This principle remains important even when the AI system can update its predictions. A model can only respond to new information that it actually receives.
If vital signs have not been repeated or new symptoms are not recorded, the system may continue to display an assessment based on outdated information.
In many emergency department waiting areas, continuous physiological monitoring is not available for every patient. Clinical observation, communication with patients and attention to newly reported concerns therefore remain essential.
An AI system designed for initial acuity assessment should not automatically be assumed capable of monitoring deterioration or determining when reassessment is required. Those are different functions that require their own intended-use definition and supporting evaluation.
The emergency team must remain able to recognize and respond to clinically significant changes independently of AI output.
When the nurse or emergency physician disagrees with AI
Disagreement between a clinician and an AI system does not necessarily mean that either assessment is incorrect.
The clinician may have observed findings that were not included in the model's inputs. The system may identify a pattern that deserves further consideration. Alternatively, the recommendation may be inappropriate because of missing information, poor data quality or limitations in the model.
The appropriate response should depend on the patient's condition rather than an expectation that the clinician must agree with the software.
When AI recommends greater urgency than the initial clinical impression, the recommendation may justify closer assessment. The clinician should consider whether relevant information has been overlooked or whether the output reflects a known limitation.
When the clinician identifies findings suggesting greater urgency than the AI recommendation, the patient should not be maintained at a lower priority simply because the software appears reassuring.
An AI recommendation should not independently downgrade a patient's clinical priority without appropriate reassessment and a decision by authorized personnel.
This is particularly important when the software is integrated into electronic triage documentation. If an application automatically populates a category, staff should be able to review and, where clinically justified, change it without unnecessary delay.
The ENA's 2025 position statement on triage qualifications and competency recognizes that validated AI and machine learning programs may support triage but cannot replace the clinical judgment and real-time involvement of a competent triage clinician [5].
For nursing staff, this reinforces the importance of recognizing concerning findings, communicating changes and obtaining urgent medical review when required.
For emergency physicians, it reinforces the need to assess the patient's actual condition and respond to clinically significant concerns rather than treating algorithmic recommendations as definitive.
Hospitals should avoid creating workflows in which disagreement is technically permitted but practically discouraged. Requiring lengthy documentation before an urgent priority change, for example, could interfere with timely care.
Clinically important disagreements may warrant a brief record and subsequent quality review, particularly when they reveal a recurring system limitation or contribute to a change in clinical management.
Review should focus on understanding what happened and identifying opportunities to improve care. Agreement with AI should not automatically be treated as the correct outcome.
Patients and families can help identify changes during the wait
Reassessment should not depend entirely on whether staff notice a visible change in the patient's condition.
Some deterioration becomes apparent first through what the patient reports. A person who has been waiting may experience worsening symptoms that are not immediately obvious to others.
Relatives and accompanying persons may also recognize changes in behavior, consciousness or other features that differ from the patient's usual condition.
Emergency departments should therefore consider how patients and families can communicate new concerns while waiting and how those concerns will reach the appropriate clinical staff.
ACEM's policy on emergency medicine consultation standards specifically recognizes the need to respond appropriately to concerns raised by patients, family members and carers about deterioration while waiting for definitive care [3].
Providing clear information about the waiting process and how to request reassessment can support safer care. However, patients should not be expected to diagnose their own deterioration or determine the correct triage category.
Communication arrangements should also consider people who cannot readily describe their symptoms, including patients with communication difficulties, language barriers or altered cognition.
AI may process recorded information efficiently, but it cannot compensate for a clinical process in which important new concerns are not heard or acted upon.
What recent research tells us about AI-assisted triage
Research on AI-assisted emergency triage has expanded, but the available evidence does not yet establish that AI consistently improves triage safety in routine clinical practice.
A systematic review by Yi and colleagues, published in the Journal of Nursing Scholarship in 2025, identified seven prospective studies evaluating AI-related emergency triage outcomes [6].
The studies examined different technologies and outcomes, including classification performance, time-related measures and aspects of undertriage and overtriage. The review suggested that AI could serve as supportive technology for emergency nursing triage.
However, the included studies differed in their methods, populations and reference standards. Improved classification performance in one setting does not necessarily demonstrate equivalent clinical benefit elsewhere.
In June 2026, Cui and colleagues published a systematic review and meta-analysis examining large language models used for emergency department triage [7].
The review included 11 studies involving 3,088 real or simulated emergency cases. It did not demonstrate a statistically significant overall accuracy advantage for large language models compared with triage nurses.
For identifying patients assigned to the highest-acuity category by the included studies' reference standards, pooled sensitivity was 61%, while specificity was 97%.
This finding deserves attention because failure to recognize patients requiring the highest priority could have important safety consequences.
However, these estimates specifically describe the performance of the language models and evaluation methods included in the review. They should not be generalized to all AI triage technologies or interpreted as directly measured rates of harmful undertriage in routine clinical practice.
The results also reflect agreement with reference triage categories rather than direct measurement of every patient's underlying physiological urgency or risk of preventable harm.
A more recent systematic review and meta-analysis by Ashour and colleagues was published in the Journal of Medical Internet Research on October 19, 2026 [8].
The review included 27 studies, with 13 contributing to quantitative analysis and 14 included in a narrative synthesis. It examined AI, machine learning and other digitally supported approaches used at the first point of clinical contact.
Across four adult emergency department studies evaluating critical care or deterioration outcomes, the pooled AUC for AI-assisted models was 0.874. However, statistical heterogeneity was substantial, with an I² of 94.4%.
Studies predicting hospital admission or disposition reported AUC values ranging from 0.80 to 0.93, but differences between studies prevented meaningful pooling of those results.
These findings suggest that AI can distinguish between patients with different probabilities of certain outcomes. The pooled AUC, however, describes discrimination across the included models. It does not represent a pooled demonstration that AI produces safer triage decisions than conventional practice.
The review also considered digital workflow and tele-triage approaches, which perform different functions from predictive models used to recommend acuity categories. Their operational findings should not be interpreted as evidence that all AI triage systems reduce waiting times or improve patient safety.
The authors emphasized limitations involving external validation, calibration, performance at clinically relevant thresholds and insufficient evidence demonstrating improved patient outcomes.
Taken together, the available research supports further evaluation of AI as a clinical support tool. It does not justify assuming that better prediction, by itself, results in safer emergency care.
Conventional triage errors also need to be understood
Undertriage and overtriage are not problems introduced solely by AI. They also occur when emergency departments use conventional triage systems.
In 2023, Sax and colleagues published a large study in JAMA Network Open examining 5,315,176 adult emergency department encounters across 21 US hospitals using version 4 of the Emergency Severity Index [9].
Using the study's outcome-based definitions, investigators estimated that 32.2% of encounters involved mistriage, including 3.3% undertriage and 28.9% overtriage.
The sensitivity of assigned ESI categories for identifying high-acuity patients who subsequently received life-stabilizing interventions was approximately 65.9%.
These findings highlight the difficulty of determining clinical urgency from limited information at first contact. However, they require careful interpretation.
The estimates relate to a defined US healthcare population, a particular version of ESI and the researchers' operational definitions. They should not be presented as measured triage error rates in Vietnamese hospitals or as evidence about the performance of newer ESI versions.
The study also illustrates why determining whether an initial classification was correct can be difficult.
Some information becomes available only after further examination or investigation. A patient may also deteriorate after an initially reasonable assessment.
A retrospective review should therefore consider what clinicians could reasonably have recognized at the time of triage rather than assume that every later adverse outcome proves the original decision was incorrect.
For AI evaluation, it is useful to distinguish an inappropriate initial recommendation from a failure to recognize a subsequent change in condition.
Both may create patient-safety risks, but they point to different opportunities for improvement.
Local evaluation should reflect Vietnamese emergency care
Before introducing AI-assisted triage, a hospital should understand the evidence supporting the specific product and the conditions under which it was evaluated.
The first consideration is whether the system uses information that would genuinely be available at the intended decision point.
A model developed using laboratory results, final diagnoses or other information recorded later in the emergency visit may appear accurate retrospectively. That performance cannot automatically be expected when the system is used at arrival, before those results exist.
The organization should also examine how the model handles incomplete information.
Vital signs may not be available immediately for every patient. Clinical histories can be incomplete, and presenting complaints may be recorded differently by different staff members. Missing information should not automatically be interpreted as a normal or reassuring finding.
Language is another practical consideration. A system developed mainly using English-language emergency records may not perform identically with Vietnamese documentation, local clinical terminology or commonly used abbreviations.
Patient populations and clinical services also vary.
A model evaluated primarily in adults should not automatically be considered suitable for pediatric triage. Performance may also differ according to age, presenting complaint, clinical complexity and other relevant characteristics.
Differences in staffing, available equipment, waiting-area arrangements and documentation practices can influence how AI recommendations are used.
These issues are particularly relevant in Viet Nam, where emergency services operate in institutions with different clinical capabilities, patient populations and organizational arrangements.
Local evaluation should therefore address more than technical compatibility with the hospital information system. It should examine whether recommendations are clinically appropriate, understandable to staff and compatible with the department's actual triage process.
The evidence should also be assessed for potential bias. A model may perform differently across clinically relevant patient groups, and an overall accuracy measure can conceal important variation.
Where sufficient data are available, hospitals should consider whether performance remains acceptable across the populations and operating conditions in which the system will be used.
International research reporting guidance can help organizations assess supporting evidence.
TRIPOD+AI provides recommendations for transparent reporting of clinical prediction model studies, while DECIDE-AI addresses early-stage live clinical evaluation of AI decision-support systems [11, 12].
These are scientific reporting guidelines, not Vietnamese regulatory approvals or guarantees of clinical safety.
Their value lies in encouraging clearer reporting of intended use, development and evaluation methods, model limitations, human interaction and clinically meaningful outcomes.
What should hospitals measure after introducing AI?
The effectiveness of AI-assisted triage should be evaluated using measures that reflect the purpose of emergency care: appropriate prioritization and timely clinical attention.
The appropriateness of the initial acuity classification is an important starting point. Hospitals should examine clinically significant undertriage and overtriage using consistent definitions and an appropriate clinical review process.
Overall agreement with an existing triage category may provide useful information, but errors involving patients who require immediate or time-sensitive intervention deserve separate attention.
Time from patient arrival to initial triage is another relevant measure. However, shorter triage encounters are not necessarily beneficial if essential symptoms, observations or risk factors are missed.
Time from arrival or triage to appropriate clinical assessment provides additional information about whether patients receive timely attention.
These intervals should be examined according to clinical priority and relevant patient groups rather than relying only on an overall average.
An improvement in average waiting time may conceal persistent delays among patients with more urgent needs.
Reassessment performance requires its own evaluation.
Hospitals should understand the proportion of waiting patients who receive reassessment when indicated and whether those reassessments occur within the applicable clinical expectations.
The denominator is important. A meaningful measure should account for patients whose reassessment became due while they were still waiting, rather than simply counting the number of reassessments recorded.
The organization should also examine whether reassessment identifies changes that require a different priority or earlier clinical attention.
When a new concerning finding is recognized, the time to further clinical assessment or action may reveal delays in the escalation process.
Serious events involving patients awaiting assessment deserve particular review. These may include unexpected resuscitation, clinically significant deterioration, urgent transfer for higher-acuity care or harm associated with delayed intervention during the waiting period.
Such events should be interpreted in context. Their occurrence does not automatically prove that the initial triage decision was incorrect or that AI caused the outcome.
A meaningful review should consider the patient's condition at arrival, information available to staff, subsequent changes, reassessment arrangements and timeliness of the clinical response.
Where a delay contributed to harm, the review should distinguish the role of initial classification, missed reassessment, communication problems, limited clinical capacity and other relevant factors.
The organization should also monitor clinically important disagreements between staff and AI recommendations, technical failures and unexpected changes in workload.
A technology that increases the number of high-priority recommendations may affect other patients even when no individual AI recommendation directly causes an adverse event. This broader effect should not be overlooked.
Similarly, frequent overrides may indicate limitations in the model, problems with the workflow or differences between recorded information and clinical observations. The reasons should be understood rather than assuming that every override represents human error.
Where practical, hospitals should compare performance before and after implementation while considering changes in patient mix, staffing, triage procedures and emergency department demand.
A simple improvement after installation should not automatically be attributed to AI.
These measures are proposed areas for organizational evaluation, not a mandatory Vietnamese national AI triage scorecard or an independently validated instrument.
Reducing triage data-entry time is not sufficient evidence that AI has improved emergency patient safety.
A meaningful evaluation connects the technology to the appropriateness of clinical priority, timeliness of assessment, reliability of reassessment and consequences of delayed care.
Vietnamese professional and legal responsibilities
AI-assisted emergency triage operates within Viet Nam's legal framework for medical examination and treatment, emergency care, artificial intelligence, medical devices and personal data protection.
These areas involve related but distinct responsibilities. Compliance with product-related regulations does not automatically establish that a technology is clinically appropriate or safe for a particular emergency department.
Emergency care and professional practice
The Law on Medical Examination and Treatment No. 15/2023/QH15, effective January 1, 2024, as subsequently amended and reflected in Consolidated Document No. 26/VBHN-VPQH dated February 26, 2026, provides an important legal foundation for healthcare services and professional practice in Viet Nam [13].
Decision No. 01/2008/QĐ-BYT, issued by the Ministry of Health on January 21, 2008, promulgated regulations concerning emergency care, intensive care and poisoning management [14].
These regulations address the reception, assessment, classification and emergency management of patients according to clinical priority.
Article 9 is particularly relevant to the responsibilities of emergency department personnel.
It provides for nurses in the emergency department to participate in receiving patients and performing initial classification. Where a patient's condition is serious, nurses are required to undertake appropriate initial emergency measures and promptly notify a physician.
The provision also addresses nurses' responsibilities for observing patients, identifying abnormal developments, responding appropriately and reporting such changes to physicians. It includes responsibilities concerning the handover of patient care between shifts.
Emergency physicians have corresponding professional duties concerning the reception, clinical examination and treatment of emergency patients.
These provisions are directly relevant when determining how AI recommendations should interact with clinical responsibilities during initial triage and subsequent care.
However, they should not be interpreted as establishing a specific national AI triage protocol or a universal schedule of waiting-area reassessment intervals applicable to every hospital.
The organization of reassessment for patients waiting to be examined should therefore be developed in accordance with applicable legal requirements, professional responsibilities, the clinical population and the hospital's operating arrangements.
AI assistance does not independently expand a practitioner's permitted scope of practice or remove the responsibilities of healthcare facilities and their clinical personnel.
Artificial intelligence regulation
Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on March 1, 2026 [15].
Decree No. 142/2026/NĐ-CP, effective May 1, 2026, provides implementing provisions, including requirements concerning AI risk classification and related responsibilities [16].
Decision No. 33/2026/QĐ-TTg, effective August 15, 2026, promulgates the list of high-risk AI systems [17].
The healthcare-specific entries expressly identified in that list concern defined AI-assisted surgical and robotic surgical applications. General software intended to support emergency department triage is not separately identified among those healthcare entries.
However, this should not be interpreted as a definitive determination that every AI-assisted triage system is legally low-risk.
The applicable classification and obligations depend on the individual system's actual functions, intended purpose, deployment arrangements and relevant legal criteria.
The distinction between statutory AI risk classification and clinical risk assessment is important.
Even where a particular application does not fall within a specified high-risk category, an incorrect recommendation could still delay necessary emergency care and expose patients to significant harm.
Healthcare organizations should therefore assess both the applicable legal obligations and the clinical consequences of using the technology.
The use of AI does not, by itself, change or transfer the legally established authority and responsibilities of healthcare professionals or organizations.
Human oversight must be meaningful in practice. Staff should be able to evaluate recommendations, intervene when necessary and continue care when the technology is unavailable or unsuitable.
Medical device regulation
Medical device requirements may apply depending on the software's intended medical purpose and functions.
Decree No. 98/2021/NĐ-CP on medical device management, as amended by subsequent regulations including Decrees No. 07/2023/NĐ-CP and No. 04/2025/NĐ-CP, provides part of the relevant regulatory framework [18].
Circular No. 24/2026/TT-BYT, effective July 1, 2026, further addresses risk determination and management measures for products classified as medical devices [19].
Before deployment, healthcare organizations should establish whether a proposed AI application falls within medical device regulation and determine which classification, product documentation and other requirements apply.
Authorization or certification obtained in another jurisdiction may provide useful supporting information, but it should not automatically be treated as evidence of compliance with all Vietnamese requirements.
The assessment should be specific to the product rather than based solely on its commercial description as an AI triage system or clinical decision-support application.
Personal data protection
AI-assisted emergency triage may involve processing identifiable patient information, including presenting complaints, medical history, clinical observations and vital signs.
The Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP took effect on January 1, 2026 [20, 21].
Healthcare organizations should assess the relevant legal basis and purposes of data processing, applicable consent requirements or exceptions, access controls, information security, retention and arrangements with technology providers.
Particular attention is needed where identifiable patient information is processed outside the hospital's own systems or proposed for secondary purposes, including AI development, model improvement or training.
The applicable obligations depend on the actual data, processing purposes, technical architecture and respective responsibilities of the parties involved.
Where AI is being evaluated as part of clinical research rather than solely introduced for ordinary clinical service, the relevant research ethics and participant protection requirements should be considered separately.
These legal assessments should inform technology selection and implementation. They do not replace the need to evaluate clinical performance and patient safety.
Implementation should begin with the existing triage process
For Vietnamese hospitals considering AI-assisted emergency triage, a useful starting point is to examine how patients currently enter the emergency department and what happens while they wait.
The hospital should understand whether patients are identified and assessed promptly, how clinical priority is assigned and whether staff can recognize those requiring immediate attention.
It should then review the period following initial triage.
Are waiting patients visible to the responsible team? Are reassessments performed when indicated? Can patients or accompanying persons report new concerns? Is there a reliable pathway for urgent medical review when a patient's condition changes?
These questions matter whether or not AI is used.
If the existing process has weaknesses, improvements in staffing responsibility, communication, waiting-area observation and clinical documentation may provide important benefits independently of technology.
Where AI is being considered, clinical leaders and frontline staff should participate in defining its intended function and expected contribution.
An organization seeking to reduce clinically significant undertriage will require different supporting evidence from one evaluating whether AI can improve documentation efficiency.
A system intended to recommend initial urgency should not automatically be used to determine reassessment intervals or revise priorities while patients wait, unless those additional functions are within its intended use and have been appropriately evaluated.
The hospital should establish how recommendations will be reviewed, who can change the assigned priority and what happens when a clinician disagrees with the system.
Staff training should involve realistic situations, including incomplete information, incorrect recommendations, changing patient conditions and clinically important disagreement between human and AI assessments.
Training should also reinforce the distinction between an initial triage category and the patient's current clinical condition.
A fallback process is essential. If the system becomes unavailable or produces unreliable output, emergency staff must remain able to assess and prioritize patients using established professional methods.
A supervised, staged introduction may be appropriate where permitted by applicable requirements and supported by sufficient clinical oversight. Such an approach can help the organization evaluate the technology's behavior under actual operating conditions before considering wider deployment.
Software updates, changes to data inputs and modifications to clinical workflow may affect performance. Significant changes should therefore be considered within the hospital's technology management and clinical safety arrangements.
When unexpected problems arise, there should be a process for investigating the concern, obtaining relevant technical information and modifying or suspending reliance on the affected AI function when appropriate.
The hospital should also consider whether the technology introduces additional work that was not anticipated. More alerts, repeated manual checks or difficult override procedures may reduce the practical benefit of an otherwise promising system.
Successful implementation should not be defined by how many patients receive AI-generated recommendations or how quickly those recommendations appear.
It should be judged by whether the technology improves the existing clinical process without weakening the safeguards that patients depend on while waiting.
Safer emergency triage requires more than a correct first decision
Artificial intelligence may become a useful source of support for emergency department triage. It can help clinicians process information, recognize patterns and potentially make more consistent initial assessments.
However, emergency triage is not simply a prediction problem.
It is a clinical process in which staff must recognize urgency, assign appropriate priority and remain responsive as new information becomes available.
A model that predicts hospital admission accurately is not necessarily capable of prioritizing treatment safely. A system that identifies patients with a greater probability of ICU admission may still fail to recognize someone who needs immediate clinical attention.
The same distinction applies to operational efficiency. Faster documentation or more rapid generation of a category does not necessarily mean that patients receive earlier assessment or experience fewer harmful delays.
For hospitals in Viet Nam, responsible AI-assisted triage should remain grounded in the purpose of emergency care: recognizing clinical urgency, providing appropriate priority and responding when the patient's condition changes.
The initial decision matters, but so does what happens afterward.
Nurses and emergency physicians need to remain able to identify concerning findings, question AI recommendations and arrange timely escalation when necessary.
Healthcare organizations, in turn, need processes that make reassessment reliable and ensure that clinically important changes reach the responsible team without avoidable delay.
Evaluation should focus on meaningful undertriage and overtriage, time to appropriate clinical assessment, reassessment performance and the consequences of delayed care.
Ultimately, one of the most important patient-safety questions arises after the initial triage decision has already been made: what happens if the patient becomes more unwell while waiting?
AI-assisted emergency triage should not be judged only by how accurately it predicts risk at arrival. Its value depends on whether it helps emergency teams assign appropriate priorities, recognize changing clinical needs and deliver timely care without avoidable harm.
Professional Education Notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized medical, nursing, emergency, legal, regulatory or patient-safety advice.
The opening clinical examples are illustrative and do not represent documented patient cases.
This article focuses on initial emergency department triage and the period during which patients await appropriate clinical assessment or intervention. It does not provide a comprehensive assessment of AI systems for inpatient deterioration monitoring, intensive care prognostication, prehospital mass-casualty triage or autonomous emergency treatment decisions.
The organizational approaches discussed, including AI recommendation review, waiting-area reassessment, escalation, documentation and quality monitoring, are proposed considerations for responsible AI implementation and healthcare quality improvement. They are not mandatory Vietnamese Ministry of Health AI triage protocols or independently validated clinical assessment instruments.
International triage frameworks and professional guidance reflect specific healthcare systems, populations and intended uses. Their performance thresholds and reassessment arrangements should not automatically be interpreted as binding requirements for healthcare facilities in Viet Nam.
Evidence showing improved prediction of hospital admission, intensive care use, mortality or resource requirements does not automatically establish improved initial acuity classification, safer patient prioritization or reduced harm.
Findings from retrospective datasets, simulated clinical cases and research conducted outside Viet Nam should not be generalized to routine Vietnamese emergency care without appropriate consideration of local clinical populations, workflow, validation and implementation conditions.
The regulatory status and obligations associated with individual AI applications, medical device software, professional practice, personal data processing and clinical research require separate assessment under the applicable Vietnamese legal framework.
References
American College of Emergency Physicians and Emergency Nurses Association. Emergency Department Triage. Joint Policy Statement. Revised January 2025.
Australasian College for Emergency Medicine. Policy on the Australasian Triage Scale. P06, Version 5. November 2023.
Australasian College for Emergency Medicine. Policy on Emergency Medicine Consultation Standards of Care. P55, Version 4. November 2023.
Health Service Executive, Ireland. National Quality Improvement Framework for Emergency Department Triage (2025–2027). Emergency Medicine Programme. September 2025.
Proehl JA. Triage Qualifications and Competency. Emergency Nurses Association Position Statement. 2025.
Yi N, Baik D, Baek G. The Effects of Applying Artificial Intelligence to Triage in the Emergency Department: A Systematic Review of Prospective Studies. Journal of Nursing Scholarship. 2025;57(1):105–118. DOI: 10.1111/jnu.13024.
Cui L, Wang M, Xu Y, et al. Diagnostic Accuracy of Large Language Models for Emergency Department Triage: A Systematic Review and Meta-Analysis. BMC Emergency Medicine. 2026;26:228. DOI: 10.1186/s12873-026-01639-z.
Ashour FOM, Sami W, Alrimawi I, Rizvi MR, El-Banna MM. Technological Integration in Triage With AI and Machine Learning Compared With Conventional Triage and Narrative Synthesis of Telehealth for Patient Prioritization in Emergency and Acute Care Settings: Systematic Review and Meta-Analysis. Journal of Medical Internet Research. 2026;28. Published October 19, 2026. DOI: 10.2196/93931.
Sax DR, Warton EM, Mark DG, et al. Evaluation of Version 4 of the Emergency Severity Index in US Emergency Departments for the Rate of Mistriage. JAMA Network Open. 2023;6(3). DOI: 10.1001/jamanetworkopen.2023.3404.
Espejo T, Grossmann FF, Riedel HB, Bingisser R, Nickel CH. The Emergency Severity Index (ESI) Version 5: Simulation of Predictive Validity and Triage Level Distribution. The Journal of Emergency Medicine. 2025;78:57–70. DOI: 10.1016/j.jemermed.2025.07.035.
Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI Statement: Updated Guidance for Reporting Clinical Prediction Models That Use Regression or Machine Learning Methods. BMJ. 2024;385. DOI: 10.1136/bmj-2023-078378.
Vasey B, Nagendran M, Campbell B, et al. Reporting Guideline for the Early-Stage Clinical Evaluation of Decision Support Systems Driven by Artificial Intelligence: DECIDE-AI. Nature Medicine. 2022;28:924–933. DOI: 10.1038/s41591-022-01772-9.
National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment. Adopted January 9, 2023; effective January 1, 2024, as subsequently amended. Consolidated Document No. 26/VBHN-VPQH, issued February 26, 2026.
Ministry of Health of Viet Nam. Decision No. 01/2008/QĐ-BYT Promulgating the Regulations on Emergency Care, Intensive Care and Poisoning Management. Issued January 21, 2008.
National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted December 10, 2025; effective March 1, 2026.
Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing Certain Provisions and Measures for Implementing the Law on Artificial Intelligence. Issued April 30, 2026; effective May 1, 2026.
Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. Issued June 30, 2026; effective August 15, 2026.
Government of Viet Nam. Decree No. 98/2021/NĐ-CP on Medical Device Management, as amended, including by Decree No. 07/2023/NĐ-CP and Decree No. 04/2025/NĐ-CP.
Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Risk Determination and Management Measures for Products Classified as Medical Devices and Amendments to Circular No. 05/2022/TT-BYT. Issued June 30, 2026; effective July 1, 2026.
National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted June 26, 2025; effective January 1, 2026.
Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing Certain Provisions and Measures for Implementing the Law on Personal Data Protection. Issued December 31, 2025; effective January 1, 2026.