Governing AI-Assisted Acute Stroke Imaging in Viet Nam: Large-Vessel Occlusion Detection, Specialist Review and Timely Reperfusion Pathways

Governing AI-Assisted Acute Stroke Imaging in Viet Nam: Large-Vessel Occlusion Detection, Specialist Review and Timely Reperfusion Pathways

IMPLEMENTATION INSIGHTS

Last reviewed: 25 October 2026

A patient arrives at a general hospital with sudden weakness on one side of the body and difficulty speaking. The emergency team suspects an acute stroke and activates the hospital's stroke pathway.

An urgent non-contrast brain CT is performed, followed by CT angiography (CTA) to examine the cerebral arteries. Within minutes, an artificial intelligence system analyzes the angiographic images and flags a suspected large-vessel occlusion.

The notification reaches the stroke team before the formal radiology report is completed. The finding may be important because the patient could be eligible for endovascular treatment at a specialized stroke center.

Yet several questions remain. Is the suspected occlusion real? Does it involve a vessel covered by the AI system's intended use? Can a qualified clinician review the images promptly? Has the patient been assessed for the relevant treatment options? If transfer is necessary, can the receiving center provide the required care?

In another case, the AI system produces no alert, but a clinician reviewing the CTA identifies an arterial occlusion requiring urgent attention.

These situations illustrate why AI-assisted stroke imaging must be governed as part of a time-critical clinical pathway rather than evaluated only as image-analysis software.

AI may help identify suspected vascular occlusion and notify clinicians earlier. Its value, however, depends on whether the information reaches the appropriate professionals, supports timely decisions and contributes to suitable treatment without introducing new delays or unnecessary interventions.

Faster detection of a suspected large-vessel occlusion is not the same as faster reperfusion or better stroke outcomes.

For hospitals in Viet Nam, responsible implementation requires a reliable connection between image acquisition, AI analysis, professional review, treatment decisions and coordination between healthcare facilities.

The objective is not to make an algorithm responsible for stroke treatment. It is to help clinical teams recognize important findings, communicate effectively and act appropriately within the limited time available.

Understanding AI's role in acute stroke imaging

Acute ischemic stroke occurs when blood flow to part of the brain is interrupted, usually because an artery becomes occluded. When a major intracranial artery is blocked, the resulting neurological injury can be severe.

For appropriately selected patients, timely intravenous thrombolysis or endovascular thrombectomy can improve outcomes. These treatments have different indications and eligibility requirements. Not every patient with ischemic stroke has a large-vessel occlusion, and not every patient with an occlusion requires an endovascular procedure.

Clinical assessment and emergency brain imaging remain central to the initial evaluation.

Non-contrast CT is commonly used to assess for intracranial hemorrhage and relevant early ischemic changes. CTA provides information about intracranial arteries and, when included in the examination, extracranial vessels. Additional imaging may be appropriate in selected circumstances to evaluate the extent of ischemic injury or potentially salvageable tissue.

These examinations answer different clinical questions and should not be treated as interchangeable.

The American Heart Association and American Stroke Association published updated guidelines for the early management of acute ischemic stroke on January 26, 2026 [1]. The guidance incorporates new evidence concerning reperfusion treatment, endovascular thrombectomy eligibility and the organization of stroke care.

It emphasizes timely assessment and treatment while recognizing the importance of local stroke systems, specialist availability and interhospital transfer arrangements.

In Viet Nam, Decision No. 3312/QĐ-BYT, issued by the Ministry of Health on November 5, 2024, promulgated the national professional guidance on diagnosis and treatment of stroke [2]. It replaced Decision No. 5331/QĐ-BYT of December 23, 2020.

The Vietnamese guidance addresses diagnostic imaging, intravenous thrombolysis and endovascular intervention. It provides an important professional foundation for considering where AI-assisted image analysis may fit within existing stroke services.

This article focuses on adults undergoing CTA as part of the assessment of suspected acute ischemic stroke, particularly the use of AI software to identify suspected intracranial large-vessel occlusion (LVO).

Applications that automatically evaluate non-contrast CT, infarct extent, CT perfusion or other imaging features have related but distinct purposes. Their validation requirements and clinical roles should be assessed separately.

The same applies to systems intended to predict treatment response, recommend thrombectomy eligibility or guide an interventional procedure.

The application considered here has a narrower role: helping clinicians recognize a suspected arterial occlusion and communicate that finding in time for appropriate clinical review.

Detecting an occlusion is not the same as determining treatment eligibility

Identifying a large-vessel occlusion may be an important step in assessing a patient with acute ischemic stroke. However, the presence of an occluded artery does not automatically establish that endovascular thrombectomy is appropriate.

Treatment selection depends on several factors, including the patient's neurological condition, the affected vessel, time from symptom onset or last known well, relevant imaging findings, previous functional status and other clinical circumstances.

The 2026 AHA/ASA guideline includes updated recommendations for thrombectomy in selected patients with anterior circulation occlusions, including some patients with larger areas of established ischemic injury than were covered by earlier recommendations [1]. It also addresses treatment in selected patients with basilar artery occlusion.

These developments reflect an evolving evidence base for reperfusion treatment. They do not mean that every arterial occlusion should be treated mechanically.

An AI system may identify a suspected obstruction in the internal carotid artery or middle cerebral artery without evaluating all the clinical information required for treatment selection.

A suspected occlusion may also represent severe arterial stenosis, chronic disease or another vascular condition requiring different management.

The distinction between detecting a vascular abnormality and determining appropriate treatment must therefore remain explicit.

An automated finding may justify urgent attention and preparation for further assessment. It should not independently initiate thrombectomy, determine that a patient is ineligible for intravenous thrombolysis or trigger an irreversible treatment decision.

Those decisions require assessment by appropriately qualified professionals within the hospital's authorized capabilities and applicable clinical guidance.

The intended detection target must be clearly defined

Not every AI system described as an LVO detector identifies the same arterial segments.

Some applications are primarily designed to detect proximal anterior circulation occlusions, including specified portions of the internal carotid artery and the M1 segment of the middle cerebral artery.

Others may include certain M2 occlusions, additional intracranial territories or posterior circulation vessels within their intended detection scope.

These differences matter because diagnostic performance can vary substantially between vessel locations.

An application developed mainly for proximal anterior circulation occlusions should not automatically be assumed to detect distal arterial occlusions or basilar artery occlusion reliably.

Nor should a system intended to identify intracranial LVO be treated as a comprehensive assessment of extracranial carotid disease, arterial dissection or every vascular abnormality visible on CTA.

Before adopting software, healthcare organizations should understand which arterial territories and occlusion types are included, which are excluded and what the output actually represents.

The distinction between large-vessel and medium-vessel occlusion is particularly important. Research may combine both categories in an overall performance measure, even though sensitivity differs between them.

A hospital should not rely on an overall accuracy figure without understanding which vessels and abnormalities contributed to that result.

The software's clinical function also needs to be clear.

Some systems primarily assist image interpretation. Others prioritize examinations for review or generate notifications to designated clinical teams. These functions may overlap, but they are not equivalent.

A system may perform adequately as an additional notification mechanism while remaining unsuitable as the sole basis for interpreting CTA.

The organization therefore needs to evaluate the actual clinical function being introduced rather than the general description of the product as an AI stroke solution.

Image quality and information flow affect detection

AI analysis depends on receiving appropriate and sufficiently complete imaging data.

CTA requires suitable acquisition and contrast timing to visualize the relevant vessels. Patient movement, inadequate contrast opacification, incomplete anatomical coverage and reconstruction problems can affect image interpretation.

These limitations may also influence AI performance.

An algorithm may be unable to process an examination because the required image series is incomplete, the files are incompatible or the acquisition falls outside the supported protocol.

In other circumstances, the software may generate an output despite image limitations that increase the likelihood of an incorrect finding.

However, failed AI processing is not always caused by poor-quality images. Problems can also arise from image selection, reconstruction settings, data transfer or the software's own quality-control mechanisms.

A multicenter study published by Andersson and colleagues in Radiology: Artificial Intelligence in April 2026 provides a useful example [3].

The researchers evaluated a commercial AI system using 3,031 emergency CTA examinations acquired across a regional network of ten hospitals in Sweden.

The system produced a valid result for 2,804 examinations, representing 92.5% of the sample.

The remaining 227 examinations, or 7.5%, did not generate valid AI results. A subsequent audit identified several causes, including missing image series, transfer problems and incorrect quality-control classification by the software.

Importantly, 173 of the 227 failures resulted from quality-control misclassification. Most of these cases were associated with a particular image window preset rather than a clinically unusable CTA examination.

Only five failures were attributed to the system correctly identifying unsuitable image data.

This finding matters because an AI application may fail to analyze an examination even when the images remain suitable for professional interpretation.

Hospitals should therefore distinguish unsuccessful image acquisition from unsuccessful AI processing.

A system's reported processing time also requires careful interpretation.

An algorithm may analyze images rapidly after receiving them, but delays can occur during reconstruction, transfer to the picture archiving and communication system (PACS), transfer to the AI platform or delivery of results to the clinical team.

A hospital may gain little benefit from rapid algorithm processing if the required images are consistently delayed before reaching the system.

Evaluation should cover the complete imaging and communication pathway rather than rely exclusively on vendor-reported processing times.

Most importantly, examinations that cannot be processed by AI must continue through the established clinical interpretation process.

A missing or invalid AI result must never be interpreted as evidence that no large-vessel occlusion is present.

False-positive and false-negative findings have different consequences

A false-positive AI alert may identify a suspected occlusion that is not confirmed during professional review.

Possible causes include severe arterial stenosis resembling occlusion, unusual vascular anatomy and limitations in image acquisition or analysis.

A study published in Neuroradiology in 2026 provides an example of these difficulties [4].

Ha and colleagues compared two commercial AI systems, RAPID CTA and JLK LVO, in a retrospective analysis of 176 consecutive patients with suspected stroke who underwent CTA and CT perfusion at a Korean stroke center.

Both systems demonstrated an area under the receiver operating characteristic curve of approximately 0.93.

However, the study identified important differences in the types of errors produced and the trade-offs between sensitivity and specificity.

High-grade intracranial stenosis was a common cause of false-positive findings, while well-developed collateral circulation in certain distal occlusions contributed to false-negative results.

These findings illustrate why AI outputs need to be interpreted alongside the underlying images and clinical circumstances.

A false-positive alert may lead to unnecessary specialist activation, additional review or consideration of an interhospital transfer that is not ultimately indicated.

A false-negative result presents a different concern.

If software fails to identify an important occlusion and clinicians rely excessively on the absence of an alert, necessary assessment or treatment may be delayed.

The risk is greater when staff assume that the AI system has evaluated every intracranial vessel, including territories outside its validated scope.

The usual clinical interpretation of CTA should therefore continue regardless of whether AI identifies an abnormality.

A negative AI result cannot exclude all vascular pathology or override concerning neurological findings.

The objective is not simply to minimize false alerts. It is to understand the system's limitations and establish a workflow that manages clinically important errors safely.

What newer evidence tells us about diagnostic performance

Studies published in 2026 provide encouraging evidence that AI can assist in detecting intracranial arterial occlusions. They also demonstrate why diagnostic accuracy should not be confused with clinical benefit.

In the multicenter study by Andersson and colleagues, the AI system was compared with routine clinical radiology reporting, with expert review used to establish the reference standard for examinations flagged by either approach [3].

The investigators evaluated emergency head and neck CTA examinations from a broader clinical population, not exclusively patients ultimately diagnosed with acute ischemic stroke.

Among examinations with valid AI output, sensitivity for LVO was 92.8%, while sensitivity for medium-vessel occlusion was 76.1%.

For vessel occlusion detection within the system's intended scope, combining large- and medium-vessel occlusions, sensitivity was 81.7% and specificity was 99.6%.

The algorithm identified 42 examinations with vessel occlusions that had not been described in the original clinical radiology reports.

This represents a potentially useful contribution to detection, particularly when automated assessment draws attention to abnormalities that might otherwise be overlooked.

However, the study did not demonstrate a statistically significant difference in overall detection performance between the AI system and routine radiology reporting.

Some occlusions were identified by AI but not in the original reports, while others were identified through routine interpretation but missed by the algorithm.

The 42 additional findings should therefore not be interpreted as 42 patients who necessarily required thrombectomy or whose outcomes were improved by AI.

There is also a methodological limitation. Examinations considered negative by both the AI system and the clinical report were not all independently reread for the reference standard. This limits the ability to establish whether an occlusion was missed by both approaches.

The study assessed diagnostic accuracy rather than the effect of active AI deployment on disability, mortality or treatment time.

Another study published online in the American Journal of Neuroradiology in August 2026 evaluated an AI-based CTA application for detecting anterior circulation LVO [5].

Cendrero and colleagues analyzed 379 evaluable examinations from four external institutions in the United States and Europe.

The reported sensitivity was 95.8%, specificity was 88.6%, and mean time to notification was approximately 3.3 minutes.

Performance differed according to the type of occlusion, with sensitivity for M2 occlusions lower than for M1 and internal carotid artery occlusions.

These findings support the potential value of automated CTA analysis, but they describe a particular product and the conditions under which it was evaluated.

They do not establish that all LVO detection systems perform similarly or that faster notification necessarily improves treatment outcomes.

For hospitals, the relevant question is therefore not simply whether an algorithm identifies an occlusion accurately under study conditions.

It is whether the selected system provides sufficiently reliable and timely information in the patients, imaging protocols and clinical workflow where it will actually be used.

Faster notification does not necessarily mean faster reperfusion

One of the main reasons hospitals consider AI-assisted LVO detection is the possibility of reducing delays in acute stroke care.

In some settings, CTA images may become available before a formal radiology report is completed. AI can potentially identify a suspected occlusion and notify the stroke team earlier.

This may be useful when specialized image interpretation is not continuously available on site or when several professionals need to coordinate an urgent response.

However, treatment delays can occur at multiple points in the pathway.

A notification may be generated quickly but not acknowledged. A clinician may receive the alert but be unable to review the images immediately. The team may confirm the occlusion but face delays in specialist consultation, patient transfer or access to an endovascular service.

An alert therefore represents only one step in a longer process.

A systematic review and meta-analysis published by Dantas and colleagues in Neuroradiology in 2025 examined the effect of AI-assisted initial radiological assessment on workflow outcomes in patients with acute LVO stroke [6].

The review included nine studies involving 1,270 patients, of whom 671 underwent AI-assisted assessment.

Compared with standard workflows, AI assistance was associated with a mean reduction of approximately 20.6 minutes in door-to-groin puncture time and approximately 15 minutes in the interval from CTA to reperfusion.

These findings suggest potential operational benefits.

However, the pooled analysis did not demonstrate a statistically significant improvement in functional independence, defined as a modified Rankin Scale score of 0–2, or a statistically significant reduction in mortality.

A further systematic review and meta-analysis published in Cardiovascular and Interventional Radiology in 2026 examined 12 studies of AI-based LVO detection and stroke workflow management [7].

The review included one clinical trial and 11 observational studies. AI implementation was associated with shorter intervals across several stages of care, including an average reduction of approximately 17.1 minutes in door-to-groin puncture time.

The authors nevertheless emphasized that additional randomized evidence is needed.

These reviews provide useful information about potential workflow improvements, but their included studies may overlap. Their pooled findings should not be treated as entirely independent evidence.

The results also require attention to study design and possible confounding.

For example, installing AI software may occur alongside changes in staffing, imaging protocols, communication arrangements or interhospital coordination.

Observed improvements in treatment time may therefore reflect several changes rather than the algorithm alone.

Shorter workflow intervals are encouraging, but they are not sufficient proof of improved neurological outcomes.

A hospital needs to determine whether earlier notification actually leads to appropriate clinical action and whether the resulting changes improve care for the population entering its stroke pathway.

AI must not delay other time-critical stroke treatment

AI-assisted CTA interpretation should be integrated into acute stroke care without delaying treatments that are already clinically indicated.

This is particularly important for intravenous thrombolysis.

The 2026 AHA/ASA guideline recommends that eligible adults with acute ischemic stroke presenting within the conventional 4.5-hour treatment window receive intravenous thrombolysis as quickly as possible, without avoidable delays associated with additional multimodal imaging, including CTA, magnetic resonance angiography or perfusion imaging [1].

In practical terms, a hospital should not require completion of AI-based CTA analysis before initiating intravenous thrombolysis when the patient has already been appropriately assessed and treatment is indicated.

Some hospitals can acquire vascular imaging efficiently without delaying thrombolysis. Others may need different arrangements depending on their resources, staffing and imaging workflow.

Additional imaging may also have a legitimate role in selected clinical situations, including assessment of certain patients outside the conventional early treatment window.

The important distinction is that an AI workflow should not create unnecessary delays in established emergency treatment.

Intravenous thrombolysis and endovascular thrombectomy also have different clinical indications.

A patient may be eligible for thrombolysis without having an LVO detectable by AI.

Conversely, an LVO alert does not automatically establish eligibility for thrombectomy or determine whether thrombolysis should be administered.

Where both treatments are indicated, the clinical team must coordinate them according to applicable evidence and professional guidance rather than allow the AI workflow to delay one while the other is being considered.

A workflow that accelerates LVO notification but delays another indicated treatment could introduce a new patient-safety problem.

For this reason, hospitals should evaluate the effect of AI on the wider reperfusion pathway rather than only the speed of automated detection.

Specialist review and professional accountability remain essential

AI-assisted LVO detection involves several professional groups whose responsibilities must be clearly understood.

The emergency team recognizes suspected stroke, initiates the clinical pathway and coordinates immediate assessment.

Radiologists or other appropriately qualified professionals interpret the relevant imaging within their authorized professional scope.

Stroke neurologists assess neurological findings, review relevant information and contribute to treatment decisions.

Neurointerventional specialists evaluate the appropriateness and feasibility of endovascular treatment when indicated.

Radiographers and imaging personnel remain responsible for relevant technical aspects of image acquisition and examination quality within their professional roles.

Depending on the hospital, some functions may involve remote consultation or collaboration between facilities.

AI can support communication of a suspected finding, but it does not replace professional interpretation or establish that the patient requires an invasive procedure.

The organization should determine who receives automated notifications, who reviews the images and who communicates clinically confirmed findings.

Professional review should be organized in a way that supports the urgency of stroke care. Where suitably qualified clinicians can assess the relevant images and initiate appropriate action, the workflow should not create unnecessary delay solely to obtain a completed formal report.

At the same time, an automated notification should not be represented as a definitive clinical interpretation.

The distinction becomes especially important when the AI result and a clinician's assessment differ.

A radiologist may conclude that an AI-positive finding represents severe stenosis rather than acute occlusion. In another case, a clinician may recognize a suspected arterial occlusion despite a negative AI result.

Neither situation should automatically be resolved in favor of the software.

The appropriate response is timely review of the images, the patient's clinical findings and the information available to the responsible professionals.

When disagreement has implications for urgent treatment or transfer, the team needs a reliable way to reach a clinical decision without creating avoidable delays.

The purpose is not to require lengthy documentation or multiple consultations for every discrepancy.

It is to ensure that clinically important uncertainty is recognized and managed by professionals responsible for the patient's care.

An AI alert must lead to an appropriate clinical response

The notification process can determine whether an AI application provides practical benefit.

A system may generate an alert within minutes, but that information is useful only when it reaches someone able to review and act on it.

Hospitals need to understand which professionals receive notifications during normal working hours, overnight and on weekends.

A notification sent to someone who is unavailable or no longer responsible for the patient may provide little value.

The same problem arises when several clinicians receive an alert but no one has a clearly assigned responsibility to respond.

The organization should therefore establish how notifications are acknowledged, how images are reviewed and how clinically important findings are communicated to the stroke team.

An automated alert should not be treated as a confirmed radiology report.

Where appropriate, the system should identify the result as a preliminary automated assessment requiring professional interpretation.

The notification may support preparation for possible specialist involvement while clinical review is taking place. This can be useful in time-critical care, provided that preliminary coordination is not confused with a definitive treatment or transfer decision.

Repeated alerts for the same examination, duplicate notifications and unclear escalation arrangements may increase workload without improving care.

Excessive notifications can also make it more difficult for clinicians to distinguish findings requiring immediate attention.

Conversely, suppressing too many notifications may reduce the system's ability to draw attention to important abnormalities.

These are operational and patient-safety concerns that should be evaluated during implementation.

Hospitals also need to know what happens when an alert cannot be delivered, the intended recipient is unavailable or the communication platform fails.

An examination may have been analyzed successfully without the result reaching the responsible clinician. This is not the same as successful clinical communication.

The established interpretation and communication pathway must remain available when AI is unavailable or unreliable.

An absent alert must not interrupt clinical assessment

One of the less visible safety concerns arises when an examination is not flagged by AI.

The system may fail to detect an occlusion, the affected vessel may be outside its validated scope or a technical problem may prevent analysis.

If clinicians assume that the absence of an alert means no important vascular abnormality is present, the software may unintentionally weaken clinical vigilance.

A safe workflow should preserve the established interpretation of CTA regardless of whether an automated notification is produced.

An absent alert should not prevent further assessment when neurological findings or professional image review remain concerning.

Hospitals should also distinguish a valid negative output from a failed analysis, an unsupported examination and an alert that was generated but not delivered.

These are different circumstances with different implications.

A system that cannot process an examination should not silently classify it as normal.

Likewise, a notification that was generated but never received should not be recorded as successful clinical communication.

Maintaining this distinction is particularly important in time-critical care, where an unnoticed technical failure could contribute to avoidable delay.

Interhospital coordination is part of the intervention

AI-assisted LVO detection may be particularly relevant in healthcare networks where patients are first assessed at hospitals without on-site endovascular thrombectomy services.

A patient may undergo CTA at a referring hospital and require specialist review to determine whether transfer is appropriate.

Earlier recognition of a suspected occlusion may help the referring team communicate with a receiving stroke center and prepare for subsequent care.

However, a positive AI alert should not independently determine a transfer decision.

The receiving team needs sufficient clinical and imaging information to assess whether specialized evaluation or treatment is appropriate.

Relevant information may include neurological findings, time of symptom onset or last known well, imaging results, treatment already administered and other circumstances affecting transfer and intervention.

The images themselves may need to be available for professional review rather than relying only on an automated notification.

Transfer decisions should also consider the capabilities of both the referring and receiving facilities.

An unnecessary transfer may expose a patient to additional travel and disruption while consuming limited specialist resources.

Conversely, delayed transfer of an eligible patient may reduce the opportunity for effective treatment.

The 2026 AHA/ASA guideline recognizes that prehospital routing and interhospital transfer arrangements should reflect the characteristics of the local stroke system [1].

Where rapid interhospital transfer systems are not functioning effectively, direct access to an endovascular-capable hospital may warrant consideration within the broader system of care.

This is relevant to Viet Nam, where diagnostic, neurological and neurointerventional capabilities vary among hospitals and geographic areas.

However, these differences should not automatically be attributed to a lack of AI.

Transport availability, specialist staffing, imaging access and coordination between facilities can all influence treatment times.

AI may support communication, but it cannot independently resolve these operational limitations.

Hospitals should therefore evaluate the complete referral and transfer pathway rather than rely on the number of AI alerts generated.

Local validation must reflect actual clinical conditions

Performance established in one healthcare network may not transfer directly to another.

Differences in CTA acquisition protocols, scanner models, contrast timing, patient populations and local patterns of vascular disease can influence AI performance.

Severe intracranial stenosis, for example, may be particularly relevant when evaluating a system in populations where intracranial atherosclerotic disease is common.

Performance may also vary between proximal and more distal arterial occlusions.

Hospitals should therefore assess the specific combination of imaging equipment, software and clinical workflow proposed for use.

International evidence provides an important starting point, but local verification should examine whether the system receives and processes images reliably under ordinary emergency conditions.

The assessment should include routine clinical examinations rather than only selected images of high technical quality.

It should also consider what happens when an examination falls outside the supported acquisition protocol or when the software cannot produce a valid result.

Local evaluation should distinguish diagnostic accuracy from operational reliability.

A system may correctly identify most LVOs among analyzable examinations but still provide limited benefit if it frequently fails to process studies, sends notifications to the wrong recipients or produces results too late to influence care.

The international FUTURE-AI consensus guideline, published in The BMJ in 2025, provides relevant principles for trustworthy healthcare AI, including fairness, universality, traceability, usability, robustness and explainability [8].

STARD-AI offers reporting guidance for diagnostic accuracy studies involving AI, while DECIDE-AI addresses reporting of early-stage clinical evaluations of AI-driven decision-support systems [9, 10].

These documents can help healthcare organizations examine the quality, completeness and applicability of published evidence.

They are not product approvals or guarantees that a particular application will operate safely in a Vietnamese hospital.

Monitoring must continue after implementation

Once AI becomes part of routine care, its performance and operational use need continued attention.

Changes in imaging equipment, reconstruction protocols, software versions or clinical workflow may affect results.

Hospitals should determine when such changes require additional review or verification.

Unexpected changes in alert rates, repeated processing failures, suspected missed occlusions and clinically significant disagreement between AI and professional interpretation may indicate that the system or workflow needs investigation.

Monitoring should also examine whether the software continues to operate within its intended use.

A system initially introduced to detect selected proximal occlusions may gradually be used for a broader range of vascular abnormalities if professional expectations are not maintained.

Likewise, an application introduced as a notification aid may eventually be treated as a substitute for formal image interpretation.

These changes in clinical behavior can introduce new risks even when the underlying algorithm has not changed.

Hospitals should provide a practical way for frontline personnel to report concerns and for responsible teams to determine whether further evaluation, workflow modification or temporary suspension is appropriate.

Not every disagreement between AI and a clinician represents a reportable adverse event. Its significance depends on the intended detection target, image quality, clinical circumstances and consequences for patient care.

The objective is to identify meaningful safety problems and respond proportionately rather than classify every discrepancy as a system failure.

What hospitals should measure after introducing AI

A hospital considering AI-assisted LVO detection should define the improvement it expects before introducing the technology.

One objective may be to reduce delays in recognizing important arterial occlusions. Another may be to improve communication between radiology and the stroke team or strengthen coordination between referring and receiving hospitals.

These objectives require different evaluation measures.

Diagnostic performance remains important. Hospitals should understand sensitivity and specificity for the intended arterial territories, as well as the characteristics of examinations the software cannot process reliably.

However, accuracy measured only among successfully analyzed studies provides an incomplete picture.

The organization should account for all relevant CTA examinations entering the intended workflow, including those with invalid outputs, unsupported acquisitions and processing failures.

It should also distinguish whether performance is measured per examination, per patient or per vascular lesion.

The time required to generate an automated notification is another measure, but it should not be considered in isolation.

Hospitals may examine the intervals from CTA completion to image availability, AI analysis, alert delivery, professional review and confirmation of findings.

They should then determine whether earlier notification influences clinical decisions.

For patients requiring specialist consultation, the time between imaging and meaningful communication with the responsible team may be important.

For patients transferred between facilities, evaluation may consider the interval from imaging to acceptance by the receiving center, door-in-door-out time at the referring hospital and time to arrival at the receiving center.

For patients undergoing thrombectomy, relevant measures may include door-to-groin puncture, puncture-to-reperfusion and other treatment intervals defined according to the local stroke pathway.

The interpretation of these measures requires care.

If AI contributes to earlier transfer of patients with suspected LVO, the characteristics of the transferred population may change.

Measuring treatment times only among patients who ultimately undergo thrombectomy may overlook those whose assessment was delayed, whose transfer proved unnecessary or whose treatment did not proceed.

The organization should therefore consider the broader population entering the relevant stroke pathway.

It is also important to determine whether the AI workflow affects other time-critical treatments.

A shorter interval from CTA to notification would not represent an unqualified improvement if it were accompanied by delays in intravenous thrombolysis for eligible patients.

Patient outcomes remain important.

Where appropriate data are available, evaluation may include functional status at a defined follow-up point, such as the modified Rankin Scale at 90 days, mortality and clinically significant treatment complications.

However, changes in these outcomes cannot automatically be attributed to AI.

Baseline stroke severity, patient selection, treatment availability, transport arrangements and other quality-improvement activities may influence results.

Evaluation should also examine unintended consequences, including missed occlusions, inappropriate specialist activation, unnecessary transfers, delayed intravenous thrombolysis, excessive notification burden and failure to act on important findings.

Performance during nights and weekends may differ from daytime operations, particularly when specialist availability changes.

Hospitals should also consider the staff time and resources required to operate the system, respond to notifications and manage unexpected findings.

A technology that improves one step while creating excessive work elsewhere may provide less overall benefit than its diagnostic performance suggests.

These proposed measures are organizational evaluation considerations, not a mandatory Vietnamese AI stroke scorecard or an independently validated DMVN assessment instrument.

The success of AI-assisted LVO detection should be judged by whether it contributes to appropriate clinical action and safer, more timely care, not simply by the speed or accuracy of automated notifications.

What the Vietnamese stroke care context tells us

Stroke services in Viet Nam operate across hospitals with different levels of diagnostic, neurological and neurointerventional capability.

Some facilities provide comprehensive assessment and endovascular treatment, while others rely on consultation and transfer arrangements with specialized centers.

A hospital may have access to CTA without having an endovascular service available at all times.

For these facilities, AI may contribute by supporting earlier recognition and more effective communication with specialist teams.

However, the technology should be considered only after examining the actual care pathway.

If delays arise primarily because of limited transport capacity or difficulty identifying an accepting hospital, faster AI notification alone may have little effect.

If imaging is performed efficiently but specialist review is delayed overnight, improving communication and staffing arrangements may be as important as introducing software.

Evidence from a Vietnamese stroke center illustrates why the broader pathway matters.

A conference abstract published in the European Stroke Journal in 2026 reported findings from the Registry of Stroke Care Quality (RES-Q) at Bach Mai Hospital [11].

The retrospective analysis included 7,909 patients with acute ischemic stroke recorded between 2017 and 2023. The assessment of care timeliness focused on data from 2022 to 2023.

The authors reported that only 39.5% of patients arrived within 4.5 hours of symptom onset, according to the study's definition of timely arrival.

Among patients receiving the relevant treatments, the reported median door-to-needle and door-to-groin times were 40 and 87 minutes, respectively. Nevertheless, delayed presentation remained an important challenge.

These findings come from a single-center observational registry analysis reported in a conference abstract.

They should not be treated as representative of all Vietnamese stroke services or interpreted as evidence about the effectiveness of AI.

They highlight an important health-system consideration: improvements after arrival at the hospital cannot fully compensate for delays occurring before patients reach appropriate care.

The relevant question is therefore not whether every stroke center should introduce an AI platform.

It is whether a particular application addresses an identified weakness in the local pathway and whether the organization can demonstrate that it contributes to better care.

Vietnamese professional and legal responsibilities

AI-assisted acute stroke imaging must be considered within Viet Nam's legal framework governing medical examination and treatment, stroke care, medical devices, artificial intelligence and personal data protection.

These frameworks address related but distinct responsibilities.

The technical availability of AI software or its authorization in another jurisdiction does not automatically establish that it may be used for every proposed clinical purpose in Viet Nam.

Medical examination, treatment and stroke care

The Law on Medical Examination and Treatment No. 15/2023/QH15, effective January 1, 2024, as subsequently amended, provides an important legal foundation for professional practice, healthcare services and patient rights [12].

The Law establishes responsibilities for healthcare practitioners and organizations relevant to professional decision-making, patient information and appropriate care.

For acute stroke, Decision No. 3312/QĐ-BYT of November 5, 2024, provides directly relevant national professional guidance [2].

The decision promulgated the guidance on diagnosis and treatment of stroke for application in healthcare facilities throughout Viet Nam.

Its clinical content includes guidance concerning acute ischemic stroke imaging, intravenous thrombolysis and endovascular intervention.

The document distinguishes the roles of non-contrast CT, vascular imaging and other investigations in supporting appropriate clinical decisions.

These provisions are relevant when determining how AI-assisted CTA analysis may be incorporated into existing stroke services.

An AI system may support identification of suspected LVO, but it does not replace professional assessment or independently authorize treatment.

Hospitals should evaluate the technology's role in relation to authorized services, professional scope, clinical responsibilities and established emergency procedures.

International recommendations, including those issued by AHA/ASA in 2026, can inform evidence review and quality improvement. They should not automatically be interpreted as superseding applicable Vietnamese requirements.

Artificial intelligence regulation

Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on March 1, 2026 [13].

Decree No. 142/2026/NĐ-CP, effective May 1, 2026, provides implementing provisions concerning AI activities, including matters related to risk classification and management [14].

Decision No. 33/2026/QĐ-TTg, effective August 15, 2026, promulgates the list of high-risk AI systems [15].

The legal classification of a particular stroke imaging application depends on its functions, intended purpose, deployment arrangements and applicable regulatory criteria.

A system that provides preliminary notifications for professional review may have a different regulatory profile from one designed to automate clinical decisions or control medical procedures.

Healthcare organizations should not determine legal classification solely from a product's commercial description as AI-assisted stroke triage.

Nor should the absence of an explicitly named product category be interpreted as definitive confirmation that an application is legally low-risk.

Statutory AI classification must also be distinguished from clinical risk.

A system may create substantial patient-safety concerns if its outputs lead to missed occlusions, inappropriate transfers or treatment delays, regardless of its formal classification.

The use of AI does not independently transfer or remove professional authority and responsibilities established under applicable law.

The Artificial Intelligence Law also contains transitional provisions for systems already operating before its effective date, including a specific compliance transition period for healthcare AI systems [13]. These provisions should be assessed for the individual system rather than treated as a general exemption from other applicable healthcare or medical device requirements.

Hospitals should therefore consider both the legal obligations associated with the system and the potential consequences of using it in clinical care.

Medical device regulation

AI software used for medical image analysis may fall within medical device regulation depending on its 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 framework [16].

Circular No. 24/2026/TT-BYT, effective July 1, 2026, further addresses risk determination and management measures for products classified as medical devices [17].

Before deployment, a hospital should determine the regulatory status of the specific software and the requirements applicable to its proposed clinical use.

The legal status of a CT scanner or PACS does not automatically establish the permitted use of an AI application connected to those systems.

Likewise, foreign regulatory authorization or published clinical validation does not independently demonstrate compliance with all Vietnamese requirements.

The assessment must relate to the actual product, its functions and the proposed clinical deployment.

Personal data protection and interhospital information sharing

AI-assisted stroke imaging may involve processing identifiable medical information, including imaging data, examination details, clinical records and patient identifiers.

The Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP took effect on January 1, 2026 [18, 19].

Healthcare organizations should assess the applicable lawful basis and purposes of processing, consent requirements or exceptions, access controls, information security, retention and arrangements with technology providers.

These responsibilities become particularly important when CTA images are processed through external platforms or shared between hospitals.

The organization should understand who can access the information, where it is processed and how patient data are protected during clinical communication.

If notifications are delivered through mobile applications or external communication services, their use should be assessed against applicable privacy, information security and professional requirements.

The urgency of stroke care does not automatically remove every legal obligation concerning personal data.

At the same time, privacy requirements should be interpreted according to the lawful bases and exceptions that actually apply, rather than assuming that all urgent clinical information sharing is prohibited or requires identical consent arrangements.

Secondary use of imaging data for algorithm development, model training or commercial product improvement requires separate assessment of applicable legal and institutional requirements.

Where software is being evaluated through clinical research, the relevant research ethics and participant protection requirements should also be considered.

These assessments should support lawful and safe implementation without creating avoidable interference with clinically necessary emergency care.

A practical direction for hospitals and stroke networks in Viet Nam

Responsible implementation should begin with a clearly defined operational problem.

A hospital may wish to reduce delays in specialist CTA interpretation, improve communication of suspected LVO or shorten the time needed to coordinate transfer to an endovascular-capable center.

These are reasonable objectives, but each requires a different approach.

The organization should first review its existing stroke imaging and treatment pathway.

It needs to understand when CTA is requested and performed, how images are interpreted, who communicates significant findings and how decisions are coordinated with specialist services.

Clinical leaders from emergency medicine, neurology, radiology and neurointervention should participate in defining the intended use.

Where interhospital transfer is involved, the receiving center should also have an appropriate role in establishing communication and escalation arrangements.

The selected AI system should be assessed against the intended imaging protocols, patient population and clinical workflow.

Hospitals should establish how positive findings are verified, how negative results are handled and what happens when the application cannot analyze the images.

Responsibility for responding to notifications should be clear, including when the intended recipient is unavailable.

Where permitted by applicable requirements, a supervised evaluation may help determine whether AI performs as expected and changes clinical actions appropriately.

The hospital should monitor whether the technology shortens relevant workflow intervals without delaying other indicated care.

It should also examine false alerts, missed findings, unnecessary transfers and technical failures.

A reliable fallback remains essential. Emergency stroke assessment and imaging interpretation must continue when AI is unavailable or unreliable.

The purpose is not to make the stroke pathway dependent on one software platform.

It is to determine whether AI can strengthen the existing service while preserving professional judgment, effective communication and patient safety.

From faster image detection to safer reperfusion pathways

AI-assisted detection of large-vessel occlusion on CTA offers a potentially useful way to support acute stroke care.

Recent studies have demonstrated encouraging diagnostic performance, and systematic reviews suggest that AI-supported workflows may shorten selected treatment intervals.

However, faster detection does not automatically mean that patients receive appropriate treatment earlier or achieve better neurological outcomes.

Clinical value depends on the complete pathway.

Images must be acquired and processed correctly. Important findings need professional review. Notifications must reach the responsible team, and treatment or transfer decisions must reflect the patient's condition and available evidence.

A positive AI result should not automatically initiate thrombectomy or interhospital transfer. A negative or missing result should not interrupt appropriate clinical assessment.

AI must also be introduced without delaying other time-critical treatments, including intravenous thrombolysis when indicated.

For hospitals in Viet Nam, responsible implementation requires attention to specialist availability, imaging workflow, communication between facilities and the practical capabilities of local stroke services.

Evaluation should extend beyond diagnostic accuracy and notification speed to examine clinical decisions, treatment intervals, unintended consequences and patient outcomes.

The central question is not whether an algorithm can identify a suspected vascular occlusion within minutes.

It is whether the healthcare organization can use that information to support timely, appropriate and safer care.

AI-assisted stroke imaging creates clinical value when reliable detection is connected to specialist review, coordinated decision-making and timely reperfusion for patients who are appropriately selected for treatment.

Professional Education Notice

This article is intended for professional education and knowledge exchange. It does not constitute individualized medical, neurological, radiological, emergency, neurointerventional, legal, regulatory or patient-safety advice.

The opening clinical scenarios are illustrative and do not represent documented patient cases.

This article focuses on AI-assisted detection of suspected intracranial large-vessel occlusion using CT angiography in adults undergoing assessment for acute ischemic stroke. It does not establish the clinical safety or effectiveness of AI systems for autonomous treatment selection, intravenous thrombolysis decisions, endovascular intervention, automated thrombectomy or other applications outside the defined scope.

Different AI applications may support different arterial territories, imaging protocols, patient populations and clinical functions. Evidence supporting one product should not automatically be generalized to another.

International clinical recommendations, including the 2026 AHA/ASA guideline, reflect specific evidence assessments and healthcare contexts. They should not automatically be interpreted as binding requirements that supersede applicable Vietnamese professional guidance.

The implementation approaches discussed, including specialist review, notification management, interhospital coordination, local validation and performance monitoring, are proposed considerations for responsible AI implementation and healthcare quality improvement. They are not mandatory Vietnamese Ministry of Health AI stroke protocols or independently validated organizational assessment instruments.

Evidence showing improved diagnostic accuracy, notification speed or treatment intervals does not automatically establish that an AI system improves functional recovery, reduces mortality or produces better patient outcomes.

Results from selected studies and healthcare systems outside Viet Nam should not be generalized to local clinical practice without appropriate assessment of intended use, imaging protocols, professional responsibilities, service capabilities and implementation conditions.

The legal status and obligations associated with individual AI systems, medical device software, professional practice, interhospital information sharing, personal data processing and clinical research require separate assessment under the applicable Vietnamese legal framework.

References

1. Prabhakaran S, Gonzalez NR, Zachrison KS, et al. 2026 Guideline for the Early Management of Patients With Acute Ischemic Stroke: A Guideline From the American Heart Association/American Stroke Association. Stroke. 2026;57(8)–e436. Published online January 26, 2026. DOI: 10.1161/STR.0000000000000513.

2. Ministry of Health of Viet Nam. Decision No. 3312/QĐ-BYT Promulgating the Professional Guidance on Diagnosis and Treatment of Stroke. Issued November 5, 2024. Replaces Decision No. 5331/QĐ-BYT of December 23, 2020.

3. Andersson H, Hansen B, Wassélius J. Commercial AI Model Diagnostic Accuracy for Intracranial Large- and Medium-Vessel Occlusion at Emergency CT Angiography. Radiology: Artificial Intelligence. 2026;8(3). DOI: 10.1148/ryai.250749.

4. Ha HJ, Ryu WS, Kim BJ, et al. RAPID CTA Versus JLK LVO for Large Vessel Occlusion Detection: A Pragmatic Comparison of Performance and Common Pitfalls. Neuroradiology. 2026;68:877–886. DOI: 10.1007/s00234-026-03921-0.

5. Cendrero J, Tanzi L, Hernández-Pérez M, et al. Multicenter Validation of an AI-Based CTA Tool for Anterior Circulation Large Vessel Occlusion Detection. American Journal of Neuroradiology. Published online August 11, 2026. DOI: 10.3174/ajnr.A9575.

6. Dantas J, Barros G, Mutarelli A, et al. Detection of Large Vessel Occlusion Using Artificial Intelligence Tools: A Systematic Review and Meta-Analysis. Neuroradiology. 2025;67(9):2399–2407. DOI: 10.1007/s00234-025-03716-9.

7. Barrera Gutierrez JC, Vivian E, Shah J, Meyrat R. Impact of Artificial Intelligence-Based Triage on Stroke Workflow Metrics: A Systematic Review and Meta-Analysis. Cardiovascular and Interventional Radiology. 2026;49(9):1715–1725. DOI: 10.1007/s00270-026-04473-9.

8. Lekadir K, Frangi AF, Porras AR, et al. FUTURE-AI: International Consensus Guideline for Trustworthy and Deployable Artificial Intelligence in Healthcare. BMJ. 2025;388. DOI: 10.1136/bmj-2024-081554.

9. Sounderajah V, Guni A, Liu X, et al. The STARD-AI Reporting Guideline for Diagnostic Accuracy Studies Using Artificial Intelligence. Nature Medicine. 2025;31:3283–3289. DOI: 10.1038/s41591-025-03953-8.

10. 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.

11. Qorina F, Adamu A, Jones C, et al. Timeliness of Acute Ischaemic Stroke Care: Insights From the Registry of Stroke Quality Care (RES-Q) at a Vietnamese Stroke Centre. European Stroke Journal. 2026;11(Suppl 1). Conference abstract ESOC2026YS118. DOI: 10.1093/esj/aakag023.1204.

12. 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.

13. National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted December 10, 2025; effective March 1, 2026.

14. 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.

15. 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.

16. 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.

17. 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.

18. National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted June 26, 2025; effective January 1, 2026.

19. 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.