When Clinical AI Does Not Flag a Case in Viet Nam: Governing Unflagged Results, Worklist Prioritization and Delayed Review
PRACTICAL FRAMEWORKS & TOOLS
When Clinical AI Does Not Flag a Case in Viet Nam: Governing Unflagged Results, Worklist Prioritization and Delayed Review
PRACTICAL FRAMEWORKS & TOOLS
Last reviewed: 8 October 2026
A hospital introduces an AI application to help prioritize chest X-rays for radiologist review. Examinations identified by AI as potentially urgent move higher in the reporting worklist, allowing them to receive earlier attention. The hospital expects this approach to improve efficiency and reduce delays for patients who may need prompt clinical assessment.
But consider a patient whose chest X-ray contains an important abnormality that the AI does not flag. The examination remains in the routine worklist while newer, AI-flagged studies are repeatedly moved ahead of it. The patient's condition may be deteriorating, yet the image has not been reviewed by a radiologist.
The AI has not necessarily issued an incorrect diagnosis. It may simply have failed to identify a finding within its intended detection scope. However, when the absence of a flag affects reporting priority, it can also influence when the patient receives professional assessment and subsequent care.
This illustrates a patient-safety problem that may be overlooked when healthcare organizations focus mainly on how accurately AI identifies abnormalities or how quickly flagged cases are reviewed.
When clinical AI does not flag a case, the absence of a warning should not automatically be interpreted as evidence that the patient is safe, the examination is normal or professional review can be delayed without consequence.
AI-supported prioritization may offer meaningful operational benefits. The challenge is to ensure that earlier attention for selected cases does not create unmanaged delays for others.
For healthcare organizations in Viet Nam, this means looking beyond model performance to understand how AI changes reporting priorities, waiting times and the safeguards that remain in place for patients whose examinations are not flagged.
An unflagged examination is not necessarily a normal examination
Before introducing AI into a clinical worklist, healthcare organizations need to understand what it actually means when the system does not flag an examination.
A clinical AI application may be designed to detect selected abnormalities, estimate the likelihood of a particular condition or assign a priority category according to configured criteria. Its output is limited by its intended purpose, the information it analyzes and the findings it has been developed and evaluated to identify.
For example, an AI application used in chest radiography may identify certain pulmonary abnormalities or findings associated with urgent conditions. If it does not flag an image, this may mean only that the system did not identify a targeted finding meeting its configured criteria.
It does not necessarily mean the examination contains no abnormality. Nor does it mean that other clinically important conditions have been excluded.
This distinction becomes particularly important when AI outputs are displayed as “normal,” “low priority” or “no urgent finding.” Such labels may influence how clinicians understand an examination, even when the technology has been evaluated only to support prioritization.
A system designed to detect selected abnormalities should not be treated as though it can exclude every clinically significant finding.
There is also an important difference between an examination that AI has successfully analyzed without generating a flag and one for which AI analysis was unsuccessful or unavailable. The latter may occur because of an unsupported examination type, an image-quality problem, a processing error or a failure to transfer the relevant data.
These situations should remain distinguishable within the clinical workflow. An examination that was never successfully analyzed by AI cannot reasonably be treated as though the system assessed it and found nothing concerning.
The first governance question is therefore: When a clinician sees no AI flag, what exactly does that output establish, and what remains unknown?
Unless this distinction is clear, the absence of a flag may create reassurance or influence workflow decisions beyond what the system's intended use and supporting evidence justify.
How AI prioritization can change waiting times
Diagnostic departments regularly manage competing demands. Some examinations require immediate interpretation, while others can appropriately follow routine reporting processes.
Reporting worklists may already reflect examination time, clinical urgency, care setting, specialty and operational priorities. AI introduces another source of information by identifying studies that may warrant earlier attention.
This can be useful, particularly in high-volume services where an important abnormality may not be apparent from the original imaging request.
However, priority is a relative decision. Moving one examination toward the front of a worklist can affect when another examination is reviewed.
The consequences depend on how AI is integrated into the workflow. If the system highlights potentially urgent studies while preserving existing clinical priorities, its operational effects may be relatively limited. If it automatically rearranges the worklist, creates separate queues or repeatedly places newly flagged examinations ahead of unflagged studies, the impact can be more substantial.
An examination that is not flagged may continue to wait while newer studies receive attention.
In many situations, that may be clinically acceptable. But the unflagged examination could also contain an abnormality the AI missed, a finding outside its detection scope or a condition whose urgency depends on information unavailable to the model.
The concern is not simply that AI can produce false-negative results. It is that an AI error or limitation may have additional clinical consequences when it influences the timing of professional review.
A system may perform well in identifying urgent findings overall while creating an unfavorable distribution of waiting times for a smaller number of patients.
Healthcare organizations therefore need to examine the entire worklist rather than only the examinations that AI moves forward.
What international studies show about the trade-offs
Research on AI-supported worklist prioritization provides useful evidence that workflow benefits need to be assessed alongside possible delays.
Baltruschat and colleagues developed a clinical workflow simulation using data from a German university hospital to evaluate AI-supported chest X-ray prioritization. The study was published online in November 2020 and appeared in the 2021 volume of European Radiology [1].
The simulation found that AI prioritization could substantially reduce average reporting times for examinations containing critical findings. However, maximum reporting times increased in certain circumstances, particularly when the AI failed to identify an important abnormality and the examination was placed toward the end of the worklist.
The researchers also evaluated a safeguard that increased an examination's priority once it had waited beyond a defined limit. Within the simulated workflow, this reduced some of the excessive delays associated with false-negative AI predictions.
The findings illustrate an important principle: reducing average reporting time for priority cases is not sufficient if individual patients may experience unacceptably long delays.
However, the study was a simulation rather than a prospective clinical trial measuring patient outcomes. Its results should not be interpreted as establishing a universal reporting-time limit for clinical practice.
More recent evidence comes from the LungIMPACT randomized controlled trial, published in Nature Medicine in March 2026 [2].
This prospective, multicenter study evaluated whether AI-supported prioritization of chest X-rays requested from primary care could accelerate the lung cancer diagnostic pathway in the United Kingdom.
The analysis included 93,326 chest X-rays from 86,945 patients across five NHS hospital organizations. AI outputs were available in both study arms, while immediate AI-based worklist prioritization was enabled or disabled according to randomized allocation by day.
This distinction matters. The trial primarily evaluated the additional effect of activating AI-based prioritization, not the overall effect of introducing AI compared with an entirely AI-free reporting environment.
AI prioritization reduced the median time from image acquisition to reporting from approximately 47 hours to 34.1 hours. However, the trial did not find a statistically significant improvement in its two primary outcomes: time to CT or time to lung cancer diagnosis.
Median time to CT was 53 days in both groups. Median time to lung cancer diagnosis was 44 days with prioritization and 46 days without it, a difference that was not statistically significant.
These findings should be interpreted within the study's specific clinical context. The trial evaluated a defined AI product, patient population and lung cancer pathway across participating UK healthcare organizations. Its results do not establish that AI prioritization will be ineffective in every diagnostic service or healthcare system.
Nevertheless, the trial demonstrates that faster reporting does not necessarily result in faster diagnosis or subsequent treatment. The benefit of prioritizing an examination depends partly on what happens after the report becomes available.
A further study published in the Journal of Medical Internet Research in August 2026 provides a different perspective [3].
Researchers in Singapore evaluated 1,054 chest X-rays in a single-center, prospective paired study. Eight radiologists interpreted the same examinations during two separate reporting sessions, with a four-week interval between sessions. The study compared a conventional first-in, first-out worklist with an AI-assisted workflow combining worklist prioritization and report generation support.
The investigators reported substantial reductions in report generation time and session-based turnaround time. Median report generation time decreased from 2 minutes to approximately 0.53 minutes, while mean session-based turnaround time also decreased considerably.
However, the effects of AI prioritization could not be separated from those of AI-assisted report generation. The same examinations were interpreted on two occasions, and the turnaround-time measure covered the interval from the beginning of a reporting session to report finalization, rather than the complete time from image acquisition to delivery of the clinical report.
The study therefore provides useful evidence about reporting efficiency under the evaluated conditions. It does not establish improvements in actual patient waiting times across the entire diagnostic pathway or in downstream clinical outcomes.
Taken together, these studies support a balanced conclusion. AI-supported prioritization can improve selected workflow measures, but its clinical value and safety depend on how it is implemented, how waiting times are distributed and whether the intended improvements translate into better patient care.
For healthcare organizations, the question is not whether AI prioritization is inherently beneficial or harmful. It is whether the chosen implementation improves the clinical process without introducing unacceptable risks for patients who are not flagged.
What Viet Nam's current legal framework means
Viet Nam's Law No. 134/2025/QH15 on Artificial Intelligence has been effective since 1 March 2026 [4].
Article 4 establishes human-centred principles, including that AI should serve people rather than replace human authority and responsibility. It also emphasizes human control, accountability and the ability to intervene in AI-supported activities.
Article 6 requires stricter risk management appropriate to essential sectors directly affecting life, health and other protected interests. For healthcare, the Law specifically refers to patient safety, reliability under real-world conditions and protection of health data [4].
These principles are relevant when AI influences the sequence and timing of clinical review. Although an AI system may primarily support prioritization rather than make diagnoses, its operational effects can still influence patient safety.
Decree No. 142/2026/NĐ-CP, effective since 1 May 2026, provides further requirements concerning AI risk classification, conformity assessment and the responsibilities of relevant participants [5].
Clause 3 of Article 5 requires deployers to coordinate with providers in reviewing and reclassifying an AI system when deployment, modification, integration or changes to its functions or intended purpose create new or higher risks compared with the existing classification.
Article 11 further specifies circumstances requiring risk-classification review, including significant changes affecting the original classification criteria, serious incidents demonstrating higher actual risk and changes to the legally established high-risk list. It also distinguishes these circumstances from routine technical updates that do not change the nature of the system's risks [5].
These provisions may be relevant when a healthcare organization materially changes how AI outputs influence clinical worklists. However, they should not be interpreted as meaning that every routine adjustment to a reporting queue automatically requires statutory reclassification. The legal implications depend on the actual change and whether the relevant regulatory conditions are met.
Decision No. 33/2026/QĐ-TTg, effective since 15 August 2026, establishes Viet Nam's current list of high-risk AI systems [6].
Its healthcare-specific entries concern defined AI-assisted surgical and robotic-surgery systems meeting particular conditions. AI applications used for diagnostic worklist prioritization should therefore not automatically be described as legally classified high-risk merely because they influence clinical workflows.
However, the absence of a particular application from the statutory high-risk list does not establish that it presents little clinical risk.
Healthcare organizations should distinguish legal risk classification from the consequences an AI system may have when used in actual patient care.
The Law on Medical Examination and Treatment No. 15/2023/QH15, as amended, also provides an important foundation for professional responsibilities and patient safety [7].
Article 71 addresses medical incident prevention through identification, reporting, analysis of causes and preventive measures. Circular No. 19/2013/TT-BYT continues to provide a framework for quality management in hospitals [8].
Where AI-enabled software or equipment is regulated as a medical device, the applicable medical-device framework must also be considered separately, including Decree No. 98/2021/NĐ-CP, as amended, and Circular No. 24/2026/TT-BYT where relevant [9, 10].
These legal instruments support appropriate governance and professional accountability. They do not establish a universal requirement to manually double-read every AI-unflagged examination, nor do they prescribe one maximum waiting time for all AI-prioritized diagnostic worklists.
The necessary safeguards depend on the clinical service, applicable specialized requirements, professional responsibilities, the AI system's intended purpose and the potential consequences of delayed review.
Worklist priority should not be mistaken for clinical interpretation
An important distinction is the difference between using AI to influence the order of professional review and using AI to determine the clinical meaning of an examination.
An AI application may be intended and evaluated to identify studies that potentially warrant earlier attention. In such an arrangement, the AI supports prioritization while the examination continues through the applicable professional interpretation process.
The absence of an AI flag does not mean professional interpretation has been completed.
For example, a chest X-ray not prioritized by AI may contain an abnormality outside the model's detection scope. It may also contain a finding that becomes clinically significant when compared with earlier images or interpreted alongside the patient's medical history.
Professional assessment may therefore remain necessary even when no AI flag appears.
The appropriate review arrangement depends on the system's authorized function and the clinical service involved. An AI application specifically authorized for a defined autonomous function may require governance arrangements different from those used for an AI tool intended only to prioritize examinations for radiologist reporting.
The essential point is that the organization should not expand the meaning of an AI output beyond its intended and appropriately evaluated purpose.
A decision about when an examination should be reviewed is not, by itself, a decision about what the examination means clinically.
This distinction should be reflected in reporting procedures and understood by the healthcare professionals using the system.
Preserve clinical urgency independently of AI
Healthcare organizations need to determine how AI-generated priorities interact with existing clinical urgency categories.
Consider a patient whose treating physician suspects an urgent condition based on symptoms and examination findings. The imaging request may already identify the examination as requiring prompt interpretation.
If AI does not detect a finding within its target scope, that result should not automatically downgrade the urgency assigned through the clinical process.
The same concern arises when new information becomes available after imaging. A patient's condition may deteriorate, or a clinician may communicate an additional concern requiring earlier review.
The worklist should allow authorized professional intervention when these circumstances arise.
Healthcare organizations should establish whether AI can only increase priority, whether it can also reduce an existing priority and under what circumstances any automated change is permitted.
For a system intended to identify additional potentially urgent examinations, preserving clinically assigned urgency while allowing AI to bring forward other concerning cases may be a reasonable approach.
This is a practical safeguard rather than a universal legal requirement. The appropriate configuration depends on the system's intended purpose, supporting evidence and locally approved workflow.
The organization should also identify who can override an AI-based priority and how that change becomes visible to the reporting team.
The objective is not to create competing sources of authority. It is to ensure that information from AI complements rather than unnecessarily weakens established clinical prioritization.
Protect unflagged examinations from excessive waiting
AI prioritization should not create a reporting queue in which certain examinations repeatedly lose their place without a reliable mechanism for timely review.
This is particularly relevant when new examinations arrive continuously.
A radiology department may begin the day with a manageable routine worklist. As more examinations are performed, AI may identify additional potentially urgent cases and repeatedly move them ahead of earlier studies. If this continues without appropriate safeguards, some unflagged examinations may wait substantially longer than expected.
Healthcare organizations should therefore consider how they will protect these cases from excessive delays.
One possible approach is to establish clinically justified waiting-time limits for relevant examination categories. When an examination approaches or exceeds its expected reporting timeframe, the system may require additional attention or an increase in priority.
Another approach is to manage reporting capacity so that routine examinations continue to receive appropriate review even when AI-flagged studies are being prioritized. These measures may be supported by professional review of unusually old worklist items.
The thresholds should reflect clinical urgency, examination type, available resources and any applicable professional or regulatory requirements.
A single numerical limit should not be applied across every specialty or service without justification. An examination needed to guide emergency treatment has different timing requirements from a routine outpatient investigation.
The simulation by Baltruschat and colleagues illustrates how limiting excessive waiting may reduce certain risks associated with false-negative AI outputs [1]. However, its specific thresholds should not be adopted as reporting-time standards for Vietnamese hospitals without local evaluation.
It is also important to recognize that a reporting target does not eliminate every possibility of delay-related harm. An examination may be reported within the nominal timeframe yet still be too late for the patient's clinical circumstances.
The purpose of these safeguards is to prevent AI prioritization from making important examinations less visible or allowing delays to accumulate without appropriate oversight.
The worklist should communicate AI status and uncertainty clearly
The worklist interface influences how healthcare professionals interpret AI information.
A simple distinction between “flagged” and “not flagged” may appear practical, but it can conceal clinically important differences.
An examination may have been successfully analyzed without triggering a priority flag. Another may have no AI output because the examination was unsupported, the image was unsuitable or processing failed.
These situations should remain distinguishable.
The interface should allow authorized users to understand whether AI analysis was completed, what the output means and whether any limitation affects its interpretation. The original clinical urgency and the amount of time an examination has been waiting should also remain sufficiently visible for the intended workflow.
Healthcare organizations should be cautious about labels or colors that give unflagged cases a misleading appearance of safety.
For example, a green indicator may suggest that an examination has been confirmed as normal, even when the AI has merely determined that no targeted finding crossed its prioritization threshold.
Depending on the system's intended purpose, a label such as “No AI priority flag” may communicate a more accurate meaning than “Normal.”
The organization should confirm that radiologists, referring clinicians and other users understand the distinction.
This is particularly important when AI outputs are integrated into existing radiology information systems or picture archiving and communication systems. An interface may be technically correct while still creating patient-safety risks if staff interpret its output more broadly than intended.
The design should support appropriate professional judgment rather than encourage assumptions that AI has already excluded clinically important abnormalities.
Clinical information outside the AI model must still influence priority
Clinical urgency may depend on information that the AI system does not analyze.
A radiology prioritization model may assess images but have limited access to the patient's symptoms, examination findings, recent deterioration or relevant medical history.
The treating team may therefore have information that changes how quickly an examination should be interpreted.
For example, an image may not contain a targeted abnormality identified by AI, while the referring clinician remains concerned about a serious clinical condition.
The absence of an AI flag should not automatically override that concern.
This principle extends beyond diagnostic imaging to other services where AI-supported categorization influences the order of professional review.
An application may assign priorities to selected laboratory findings or clinical information, but patient-specific concerns may require attention even when the AI does not identify a risk according to its configured criteria.
Healthcare organizations should understand which inputs the AI uses and which clinically relevant information remains outside its assessment.
When different sources of information suggest different levels of urgency, the workflow should allow appropriate professional judgment and escalation.
The governance issue is not whether clinicians always agree with AI. It is whether clinically important information beyond the AI model can still influence when and how a patient receives care.
Reporting capacity and operating arrangements matter
The safety of worklist prioritization depends partly on whether qualified professionals are available to respond to the order it creates.
An AI system may identify a potentially urgent examination within seconds, but the benefit of earlier prioritization depends on whether someone is available to review it. At the same time, unflagged examinations still require appropriate reporting capacity.
This is relevant to healthcare organizations in Viet Nam, where staffing arrangements, examination volumes, information systems and access to specialist interpretation may differ considerably between facilities.
Some hospitals may have continuous specialist coverage and integrated reporting worklists. Others may rely on scheduled reporting sessions, external diagnostic services or workflows involving several information systems.
An arrangement that works well during normal daytime operations may behave differently at night, on weekends or during periods of unusually high demand.
A department may process AI-flagged examinations promptly during the day while routine studies accumulate after hours. Alternatively, urgent examinations may be identified quickly without sufficient capacity for immediate professional review.
These problems cannot always be resolved by changing an AI threshold. The organization may need to examine staffing, reporting hours, workload distribution or arrangements for obtaining urgent specialist interpretation.
The practical question is whether the service can sustain timely review across the entire worklist once AI changes the pattern of demand.
AI should be introduced into a process that the organization can support operationally. Otherwise, the technology may make selected cases more visible without ensuring that patients receive appropriate care.
Review unflagged cases without treating them all as AI errors
A clinical safety review should include examinations that AI did not prioritize, but the absence of a flag should not automatically be classified as a false-negative result.
A false negative generally refers to a case where a target condition or finding is present according to an appropriate reference standard, but the AI does not identify it.
An unflagged examination may genuinely lack the targeted finding. It may contain another abnormality outside the AI system's scope or may have no valid AI output at all.
These categories should remain distinct.
For systems supporting professional reporting, healthcare organizations may compare AI prioritization outputs with finalized diagnostic reports to identify cases where clinically important target findings were not flagged.
Such comparisons can help identify possible missed findings, but they are not necessarily a complete measure of AI accuracy. Radiology reports have their own limitations, and not every finding is recorded in a structured form that can be reliably extracted for automated comparison.
Selected discordant cases may therefore require professional review, depending on their clinical significance and the organization's available resources.
Particular attention should be given to situations where the absence of an AI flag appears to have affected the timing of assessment.
A missed finding that did not materially change reporting priority may have different implications from one associated with a substantial delay.
The review should therefore examine both the AI output and what happened in the clinical workflow afterward.
The objective is not to retrospectively re-examine every routine study without justification. It is to use proportionate case review and sampling to understand whether the organization is managing unflagged examinations safely.
Measure the entire worklist, not only the cases AI prioritizes
A healthcare organization may evaluate AI prioritization by measuring how much faster flagged examinations receive reports.
That is useful information, but it is incomplete. The organization also needs to understand what happens to examinations that are not prioritized.
Relevant measures may include reporting times for flagged and unflagged examinations, waiting-time distributions within different clinical urgency categories, cases exceeding locally approved reporting limits and the frequency of professional overrides.
Average turnaround times should be considered alongside measures that reveal unusually long delays. A system may improve average reporting time while allowing a smaller number of patients to wait substantially longer.
In some services, the upper end of the waiting-time distribution may therefore be particularly important.
Worklist age should also remain visible. Healthcare organizations need to identify examinations that are still unread or unreported beyond their expected timeframe, regardless of whether AI flagged them.
Where data and professional review arrangements permit, organizations may examine cases in which AI did not flag an examination but subsequent interpretation identified an important targeted finding.
These reviews should consider whether the absence of a flag changed reporting order and whether any delay had potential clinical consequences.
Measures must also distinguish different time intervals. The time from image acquisition to report completion is not the same as the time a radiologist spends generating the report or the time from the beginning of a reporting session to finalization.
These measures answer different operational questions and should not be presented as interchangeable.
The organization may also compare worklist performance before and after AI introduction, where suitable data are available. Such comparisons can identify changes in waiting times and backlogs, although concurrent changes in staffing, examination volume and reporting procedures may complicate interpretation.
Where possible, evaluation should consider both the overall service and clinically relevant examination groups. Improvements for selected patients should not obscure substantial delays affecting others.
The objective is to establish whether AI-supported prioritization improves the intended service without causing clinically unacceptable deterioration elsewhere.
A reporting worklist should be evaluated as a whole, not only through the cases AI successfully brings forward.
A practical Clinical AI Non-Flag and Worklist Safety Review
Healthcare organizations in Viet Nam can address these issues through a Clinical AI Non-Flag and Worklist Safety Review.
This is proposed as a practical organizational approach, not a statutory Vietnamese framework, Ministry of Health form or international certification requirement.
The review should begin with one clearly defined AI-supported prioritization workflow. The organization should establish which findings the AI is intended to detect, what lies outside its scope and how its outputs influence worklist order.
An important question is whether AI can only increase priority or whether it can also reduce urgency assigned through other clinical processes.
The organization should then clarify what an unflagged result means. Healthcare professionals need to distinguish a completed analysis without a target finding from an output below the prioritization threshold, an invalid analysis or a missing AI result.
Next, the organization should examine what happens to unflagged examinations. This includes where they appear in the worklist, how long they are expected to wait, whether newly flagged examinations repeatedly move ahead of them and how excessive waiting is identified.
Existing clinical urgency should also be considered. The review should establish whether urgency assigned through professional processes remains effective when AI does not detect a target finding and whether authorized clinicians can request earlier review.
The organization should assess whether reporting capacity is sufficient to manage both flagged and unflagged examinations, including during busy periods and outside normal operating hours.
Foreseeable failures may include false-negative outputs, unsuitable examinations, failed processing, misleading interface labels, excessive accumulation of routine studies and urgent clinical information that is not reflected in worklist priority.
Safeguards should be proportionate to these risks. Depending on the service, they may include preserving independently assigned clinical urgency, maintaining visibility of worklist age, establishing locally justified reporting limits, allowing professional overrides and defining responses when backlogs become unacceptable.
Finally, the organization should determine how the workflow will be evaluated after deployment. The review should examine reporting-time distributions, selected discordant cases, excessive delays and situations where the absence of an AI flag may have affected clinical care.
These activities can be integrated into existing clinical governance, diagnostic quality and patient-safety arrangements.
A smaller diagnostic service may be able to conduct a relatively straightforward review involving its clinical lead, reporting team and information technology personnel. A high-volume service that allows AI to automatically reorder several clinical worklists may require a more structured multidisciplinary assessment.
The depth of governance should reflect the potential consequences of delay and the extent to which AI controls the workflow.
International guidance reinforces the need for clinical accountability
International professional guidance increasingly treats AI implementation in diagnostic imaging as an ongoing clinical governance responsibility rather than a one-time technology installation.
In May 2026, the American College of Radiology announced approval of the ACR–SIIM Practice Parameter for Imaging Artificial Intelligence [11].
The initiative addresses issues relevant to imaging AI implementation, including governance, selection, evaluation before deployment, monitoring and continuous quality improvement.
The Royal College of Radiologists published the third edition of its Standards for Interpretation and Reporting of Imaging Investigations in October 2025 [12]. The guidance emphasizes quality and consistency in professional imaging interpretation and reporting.
WHO's Regulatory Considerations on Artificial Intelligence for Health, published in 2023, provides a high-level international resource addressing safety, performance and regulatory considerations associated with healthcare AI [13].
These sources reinforce the importance of evaluating AI within the clinical environment in which it operates.
However, they do not establish the specific Clinical AI Non-Flag and Worklist Safety Review proposed in this article, nor do they create binding reporting-time requirements for healthcare organizations in Viet Nam.
The relevant legal requirements and professional responsibilities must be determined within Viet Nam's regulatory and clinical context.
What healthcare organizations in Viet Nam should prioritize
Healthcare organizations do not necessarily need a new committee or extensive documentation to begin addressing this issue.
A practical first step is to identify where AI currently influences, or is expected to influence, the order in which diagnostic examinations are reviewed.
The organization should then determine whether the absence of an AI flag changes the reporting workflow.
If all examinations continue to receive appropriate professional review within established clinical timeframes, and AI simply helps bring additional potentially urgent cases forward, some of the relevant risks may be easier to manage.
If AI actively rearranges the worklist, creates separate queues or causes unflagged examinations to wait longer, the organization needs stronger assurance that these effects remain clinically acceptable.
Healthcare leaders should ask whether established clinical urgency can override AI-based priority, how unflagged examinations are protected from excessive delay and whether the reporting team can identify studies approaching or exceeding their expected reporting times.
They should also examine how users interpret AI information. A tool intended to detect selected abnormalities should not lead staff to assume that every unflagged examination is normal.
Clinical leaders, reporting professionals, healthcare quality teams and information technology personnel should work together on these questions. The technology provider may also need to explain how outputs are generated, displayed and integrated into the reporting system.
Where staffing or infrastructure is constrained, the organization should carefully consider whether the proposed prioritization model is suitable for its actual service capacity.
Depending on the evidence and intended use, introducing AI initially as additional information for the reporting team may be a reasonable option before allowing it to influence worklist order more extensively.
The objective is not to discourage AI use in diagnostic services. It is to ensure that earlier attention for some patients does not create unmanaged delays for others.
Protecting patients who are not prioritized
Clinical AI is often judged by what it identifies successfully. Healthcare organizations examine how many abnormalities it detects, how quickly it raises warnings and whether flagged examinations receive earlier attention.
Those questions matter, but they are not sufficient when AI influences clinical worklists.
Patients whose examinations are not flagged remain part of the healthcare organization's responsibility.
Some will appropriately receive routine review. Others may have important findings that AI did not detect, conditions outside its intended scope or circumstances requiring earlier attention based on information unavailable to the model.
A well-governed prioritization process should make these distinctions understandable and protect patients from delays arising simply because their examinations did not receive an AI flag.
For healthcare leaders, the essential question is: When AI does not flag an examination, can we demonstrate that the patient still receives appropriate professional assessment within a clinically acceptable timeframe?
Answering this requires more than confidence in the model's performance. It requires understanding the reporting queue, clinical responsibilities, waiting times, operational capacity and safeguards for examinations that AI does not prioritize.
Responsible implementation should therefore evaluate both sides of the worklist: the cases AI brings forward and those that remain behind.
Ultimately, AI should help healthcare organizations direct clinical attention more effectively without allowing the absence of an AI warning to become a source of avoidable delay or false reassurance.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, patient-safety, medical-device, cybersecurity, accreditation or quality-management advice.
The terms clinical AI non-flag safety and Clinical AI Non-Flag and Worklist Safety Review are used as practical organizational concepts. They are not defined statutory categories or prescribed documentation requirements under Vietnamese law.
The proposed review is not a Ministry of Health framework, mandatory AI worklist approval process, ACR or RCR certification requirement, or WHO standard.
Applicable responsibilities depend on the AI system's intended purpose, regulatory status, actual workflow, professional activities, clinical service and relevant legal requirements. Systems authorized for defined autonomous functions may require governance arrangements different from those described for AI-assisted professional reporting.
References to waiting-time limits, professional overrides, sampling of unflagged cases and worklist safeguards are practical recommendations. They should not be interpreted as universal legal obligations or fixed numerical standards prescribed for every healthcare organization in Viet Nam.
The international studies discussed have different designs, populations, AI functions and outcome measures. Their findings should not be generalized automatically to other technologies or healthcare settings.
The proposed approach does not replace applicable regulatory authorization, clinical evaluation, professional judgment or existing healthcare quality and patient-safety responsibilities.
References
Baltruschat I, Steinmeister L, Nickisch H, et al. Smart Chest X-Ray Worklist Prioritization Using Artificial Intelligence: A Clinical Workflow Simulation. European Radiology. 2021;31:3837–3845. Published online 21 November 2020. DOI: 10.1007/s00330-020-07480-7.
Woznitza N, Smith L, Rawlinson J, et al. AI-Based Chest X-Ray Prioritization in the Lung Cancer Diagnostic Pathway: The LungIMPACT Randomized Controlled Trial. Nature Medicine. 2026;32:1737–1744. Published online 24 March 2026. DOI: 10.1038/s41591-026-04253-5.
Sridharan S, Png N, Seah AXH, et al. Impact of AI-Triaged Worklists and AI-Assisted Report Generation on Radiology Turnaround Times: Prospective Real-World Study. Journal of Medical Internet Research. 2026;28. Published 31 August 2026. DOI: 10.2196/92181.
National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Artificial Intelligence. Issued 30 April 2026; effective 1 May 2026.
Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. Issued 30 June 2026; effective 15 August 2026.
National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment, as amended. Effective 1 January 2024.
Ministry of Health of Viet Nam. Circular No. 19/2013/TT-BYT Providing Guidance on Quality Management of Medical Examination and Treatment Services in Hospitals. Issued 12 July 2013; effective 15 September 2013.
Government of Viet Nam. Decree No. 98/2021/NĐ-CP on Medical Device Management, as amended.
Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Determining Risk Levels and Management Measures for Medical-Device Products and Amending Certain Provisions of Circular No. 05/2022/TT-BYT. Issued 30 June 2026; effective 1 July 2026.
American College of Radiology. American College of Radiology Approves First Ever Practice Parameter for Imaging Artificial Intelligence. Published 5 May 2026. Announcement of the approval of the ACR–SIIM Practice Parameter for Imaging Artificial Intelligence.
The Royal College of Radiologists. Standards for Interpretation and Reporting of Imaging Investigations. Third edition. London: The Royal College of Radiologists; October 2025.
World Health Organization. Regulatory Considerations on Artificial Intelligence for Health. Geneva: World Health Organization; 2023. ISBN: 978-92-4-007887-1.