When Clinicians and AI Disagree in Viet Nam: Governing Overrides, Second Review and Clinical Safety Learning
IMPLEMENTATION INSIGHTS
When Clinicians and AI Disagree in Viet Nam: Governing Overrides, Second Review and Clinical Safety Learning
PRACTICAL FRAMEWORKS & TOOLS
Last reviewed: 10 October 2026
A hospital introduces an AI application to help identify fractures on X-rays. During an emergency department visit, the AI highlights a possible fracture, but the examining clinician reviews the image and concludes that no fracture is present. The patient is ready to go home.
Should the clinician disregard the AI result, request another professional opinion or arrange further assessment?
In another case, a radiologist identifies a potentially significant abnormality on an examination, while the AI system returns a negative result. The radiologist remains concerned and recommends further investigation.
These situations illustrate a practical challenge in clinical AI implementation. Healthcare organizations need to establish what should happen when an AI output and a healthcare professional's assessment do not agree.
The answer is not to assume that the clinician is correct because they are human, or that AI is correct because its performance has been supported by research. Both can be wrong, and the consequences of disagreement depend on the clinical circumstances.
For healthcare organizations in Viet Nam, the important question is how to resolve disagreements in a way that protects patients without creating unnecessary tests, delays, second reviews or administrative work. This requires a practical process for recognizing clinically meaningful differences, determining when additional assessment is justified, reaching an appropriate decision and learning from cases where an important finding or risk may have been overlooked.
Clinical–AI disagreement does not automatically mean a clinical error
Clinical AI applications may produce diagnostic findings, risk estimates, classifications, recommendations or alerts intended to support professional decisions. A disagreement occurs when an AI output and a clinician's assessment differ in a way that may influence patient care.
However, not every difference is clinically significant.
An AI application may highlight a subtle imaging feature that the clinician considers insignificant. A clinician may describe a finding differently from AI without disagreeing about its clinical meaning. A risk score may also fall close to a decision threshold, resulting in a different classification despite only a small difference in the underlying estimate.
Some apparent disagreements arise because the clinician and AI are not evaluating the same clinical question.
For example, an AI application designed to detect pulmonary embolism should not be expected to identify every abnormality on a CT examination. If a radiologist detects a condition outside the AI system's intended scope, the absence of an alert for that condition is not necessarily evidence that the AI has failed.
The organization should therefore establish whether the AI successfully analyzed the relevant information, whether the case was within its intended use and whether the clinician and AI were assessing the same target.
A negative AI result must also be distinguished from a system that did not process the examination, could not generate an output or indicated that the case was outside its supported scope.
The timing of the AI output matters as well. A clinician who interprets an examination before seeing the AI result is in a different situation from one who has reviewed the AI findings before reaching a conclusion. Similarly, an AI alert generated after a report has been finalized cannot be managed in exactly the same way as one available during the initial assessment.
These distinctions help determine whether a meaningful disagreement has occurred and what response is appropriate. Disagreement should be treated as a reason to examine clinically relevant uncertainty, not as automatic evidence that either the clinician or AI has made an error.
What international evidence tells us about clinician–AI disagreement
International guidance increasingly recognizes the need for clear arrangements when clinicians and AI produce different assessments.
In January 2025, the United Kingdom's National Institute for Health and Care Excellence (NICE) published HealthTech guidance HTG739 on AI technologies intended to support fracture detection on X-rays in urgent care.
NICE recommended that four selected technologies could be used in the National Health Service during an evidence-generation period, subject to specified conditions, including appropriate regulatory approval and the collection of further evidence. The technologies were intended to support healthcare professionals rather than replace their interpretation or clinical decisions [1].
The guidance emphasized the importance of clear local protocols so healthcare professionals understand what action to take when their assessment differs from the AI result. It also recognized concerns about overreliance on AI and the possibility that AI-assisted interpretation could influence professional judgment.
NICE's accompanying evidence-generation plan described approaches for evaluating diagnostic disagreement, including expert-panel assessment, arbitration of discordant findings and subsequent clinical follow-up. It also emphasized evaluating performance against an appropriate reference standard rather than relying only on agreement between AI and clinicians [2].
These recommendations were developed for particular fracture-detection technologies within the UK urgent care system. They provide useful international reference points but are not binding Vietnamese clinical procedures.
More recent research demonstrates why resolving disagreements requires careful evaluation.
Goldberg-Stein and colleagues published a study in Radiology: Artificial Intelligence in 2026 examining the implementation of a commercial AI system for pulmonary embolism detection across 32,501 CT pulmonary angiography examinations involving 29,492 patients [3].
Overall agreement between AI and radiologists was 97.79%. However, 719 examinations, representing approximately 2.21%, showed disagreement and underwent expert thoracic radiologist adjudication.
The adjudication process supported the original radiologist's interpretation in 638 of the 719 discordant examinations, or 88.73%, while supporting the AI interpretation in the remaining 81 cases.
The direction of disagreement provided additional information. Among 538 cases where the radiologist considered pulmonary embolism present but AI was negative, adjudication supported the radiologist in 483 cases and AI in 55. Among 181 cases where AI was positive but the radiologist was negative, adjudication supported the radiologist in 155 cases and AI in 26.
These findings demonstrate that either assessment could be supported after further review. Neither professional interpretation nor AI output should automatically be treated as correct whenever they disagree.
The study involved one AI system, one diagnostic target and a defined clinical workflow in which radiologists had AI information available. It was not a fully independent, blinded comparison of human and AI interpretations. Expert adjudication was used for discordant cases, while concordant cases did not undergo the same independent assessment.
The findings therefore should not be interpreted as establishing that radiologists are correct in 88.73% of disagreements with clinical AI generally, or that the observed agreement rate proves that every concordant examination was interpreted correctly.
Another study by Topff and colleagues, published in European Radiology in 2024, evaluated AI-assisted double reading of 25,104 chest radiographs across two institutions [4].
The AI-supported system initially identified discrepancies between imaging findings and radiology reports in 5,289 examinations, representing 21.1% of the total. Following external radiologist review, 47 cases were considered to contain clinically relevant missed findings. Subsequent assessment by radiologists at the participating institutions confirmed 35 of those cases as clinically relevant, representing approximately 0.1% of all examinations.
This distinction is important for implementation. A relatively large number of initial discrepancies did not translate into an equally large number of confirmed clinically significant reporting errors.
A study by Chekmeyan and colleagues, published in the Journal of the American College of Radiology in 2023, provides another example. The researchers evaluated a targeted quality-assurance workflow across 111,674 high-acuity CT examinations. The process identified 46 cases in which the radiologist's report was negative, AI indicated a high probability of a significant abnormality and the radiologist had not viewed the AI output. Subsequent review confirmed 26 initially missed findings, leading to appropriate report amendments and communication [5].
This was a targeted workflow for selected findings involving intracranial hemorrhage, cervical spine fracture and pulmonary embolism. It should not be interpreted as evidence that every AI-positive, clinician-negative disagreement requires the same process.
Taken together, these studies show that disagreements can help identify opportunities to improve clinical safety, but their significance must be established through appropriate assessment. Neither the number of disagreements nor the overall agreement rate is sufficient to demonstrate that an AI-enabled clinical workflow is safe.
Two directions of disagreement require different responses
One of the most important distinctions is whether AI identifies a possible abnormality that the clinician does not recognize, or whether the clinician identifies a concern that AI has not flagged.
These situations may appear similar, but they can create different risks.
An AI-positive, clinician-negative disagreement raises the possibility that the clinician has overlooked an important finding. It may also represent an AI false positive that could lead to unnecessary investigation, referral or treatment.
A clinician-positive, AI-negative disagreement may indicate that AI has missed a clinically relevant finding. However, it may also occur because the clinician is considering information the system did not receive, or because the finding falls outside the AI application's intended scope.
Healthcare organizations should therefore avoid applying one identical response to every disagreement.
The appropriate action depends on the disputed finding, the available evidence, the potential consequences of an incorrect decision and the uncertainty that remains after the initial professional assessment.
When AI is positive but the clinician disagrees
Consider an AI application that identifies a possible fracture on an X-ray while the examining clinician considers the image normal.
The AI output may be useful because it draws attention to a finding that might otherwise be missed. However, an AI-positive result does not establish that the suspected fracture is present.
The clinician should evaluate the original examination alongside the patient's symptoms, physical findings and other relevant clinical information.
The AI application's intended purpose and limitations also matter. Some systems are developed for particular anatomical regions, examination types or patient populations. A result generated outside those conditions may require additional caution.
Where AI highlights a particular region, the clinician may use that information to guide further review without treating the highlighted feature as definitive evidence.
If the clinician can establish an adequately supported interpretation and the consequences of an incorrect decision are limited, another professional review may not be necessary.
The response may be different when the suspected finding is serious, the examination is difficult to interpret or the clinician remains uncertain.
In those circumstances, additional imaging, specialist review or another clinically appropriate method of verification may be justified.
The organization also needs to consider potential harm from unnecessary action. Automatically investigating every AI-positive result could increase referrals, expose patients to avoidable tests and generate additional costs or anxiety.
The objective is not to maximize acceptance of AI suggestions. It is to determine whether the suspected finding warrants a change in the patient's care.
When the clinician identifies a concern but AI is negative
The opposite situation requires a different assessment.
A clinician may identify an abnormality even though the AI system has not generated an alert. For example, a radiologist may identify a clinically important finding that AI failed to detect, or a treating physician may remain concerned because of the patient's symptoms and examination findings despite a negative AI result.
An AI-negative output should not automatically override the clinician's assessment.
The clinician should continue the appropriate diagnostic or treatment process using the available clinical evidence and applicable professional guidance.
It is important to distinguish a true negative AI output from an examination that was not successfully processed. If the system failed to analyze the case or considered it outside its intended scope, the absence of an alert should not be interpreted as confirmation that the target condition is absent.
A similar distinction applies when the clinician has identified a finding that the AI was never designed to detect.
Where the clinician's concern is supported by the clinical evidence, a negative AI output does not justify abandoning that concern.
If significant uncertainty remains, additional professional review or investigation may be appropriate. However, the purpose is to resolve the clinical question, not to require the clinician to prove that the AI was wrong.
This distinction also has implications for monitoring. Organizations that examine only positive AI alerts may overlook cases where clinicians identified abnormalities despite negative AI results. They may also miss situations where both clinicians and AI failed to recognize the same condition.
A meaningful evaluation therefore requires more than counting accepted and rejected alerts.
Neither the clinician nor AI is an automatic reference standard
Disagreement often creates pressure to identify which assessment was correct.
However, the clinician's initial interpretation and the AI output are the two observations being compared. Neither should automatically serve as the definitive reference standard.
Clinicians can make diagnostic errors, especially when findings are subtle or working conditions are demanding. AI systems can produce false-positive and false-negative results, and their performance may vary across clinical settings, patient groups and examination types.
An appropriate resolution process should therefore seek supporting evidence relevant to the clinical question.
Depending on the circumstances, this may involve review of the original examination, assessment by another qualified professional, additional investigations or follow-up information that becomes available later.
Expert adjudication can be useful, but it also has limitations. A second reviewer may reach the same incorrect conclusion as the initial clinician. Reviewers may also be influenced if they know which finding was suggested by AI or how the first clinician interpreted the case.
Where an independent assessment is feasible and clinically appropriate, organizations should consider how to minimize those influences.
The appropriate reference standard will differ across clinical applications. Expert consensus may be adequate for some purposes, while others require additional diagnostic evidence, treatment response or longer-term clinical follow-up.
Not every disagreement can be resolved with certainty at the time it occurs.
Healthcare organizations should therefore allow for unresolved clinical uncertainty rather than forcing every case into a conclusion that either the clinician or AI was correct.
The goal is to reach a safe and adequately supported clinical decision, not to identify a winner in every disagreement.
Second review should be proportionate to clinical risk
Requiring an independent second opinion for every disagreement may appear cautious, but it can create unnecessary workload, delay decisions and divert professional attention from more consequential cases.
The need for additional review should reflect the potential risk.
A difference involving a finding with little clinical significance may be resolved during ordinary professional assessment. In contrast, uncertainty about a potentially life-threatening condition or a decision with substantial treatment consequences may justify prompt further evaluation.
The organization should consider the seriousness of the suspected condition, the possible consequences of a missed or incorrect finding, the quality of the available evidence and the uncertainty remaining after the initial assessment.
The expected benefit of another review is also relevant. A second interpretation that is unlikely to change management may add less value than a targeted investigation or follow-up assessment.
The availability of an appropriately qualified reviewer matters as well.
A large hospital may have specialist support and established peer-review arrangements. A smaller clinic may need to use referral, additional testing or another safe clinical pathway when immediate specialist review is unavailable.
For urgent clinical situations, necessary care should not be delayed solely to complete an AI-specific disagreement procedure.
The purpose of second review is to resolve important uncertainty, not to create another mandatory administrative step whenever a clinician and AI differ.
A practical Clinical–AI Discordance Resolution pathway
Healthcare organizations in Viet Nam can organize their response through a proposed Clinical–AI Discordance Resolution pathway.
This is a practical organizational approach intended to support safe clinical decisions. It is not a prescribed Ministry of Health procedure, an internationally validated assessment instrument or a requirement that every disagreement undergo formal adjudication.
The pathway begins by establishing whether a meaningful disagreement exists, assessing its potential clinical consequences, determining whether the available evidence is sufficient and arranging additional review when justified. It ends with an appropriate clinical decision and, where necessary, quality-improvement follow-up.
Establish what the disagreement actually concerns
The first consideration is whether the clinician and AI were evaluating the same information and clinical target.
The organization should establish whether the AI successfully processed the relevant data, whether the case was within its intended use and whether the clinician had access to the AI output at the appropriate point in the workflow.
A difference may reflect incomplete AI processing, an unsupported use case or different diagnostic targets rather than genuinely conflicting assessments.
Where a meaningful disagreement exists, the team should identify its direction and potential clinical significance. The decision about what to do next should depend on the possible consequences of that disagreement, not merely on whether the two assessments differ.
Determine whether immediate clinical action is necessary
The next consideration is whether the disputed finding could affect urgent care or another high-consequence decision.
The responsible clinician should assess the patient using the available clinical information and established professional procedures.
If immediate care is necessary, it should not be delayed merely to resolve the AI disagreement. Where the patient is clinically stable and uncertainty remains, additional review or investigation may be arranged according to the circumstances.
The urgency of the response should follow the patient's actual condition and the consequences of a possible error, not simply the presence of an AI alert.
Review the evidence and determine whether uncertainty remains
The clinician should review the underlying evidence rather than compare only the final AI classification with the professional conclusion.
Depending on the application, this may involve reassessing original images, reviewing relevant clinical findings, confirming examination limitations or considering information that the AI did not receive.
If the difference can be resolved through an appropriate assessment within the clinician's professional competence, a further review may not be necessary.
Where important uncertainty remains, the organization should determine whether an independent professional assessment, additional investigation or other escalation is justified.
The review should seek to establish the most appropriate clinical decision rather than automatically confirm the initial clinician's view or the AI output.
Resolve or escalate according to clinical significance
The organization should determine who has responsibility for resolving different categories of disagreement and when additional expertise is required.
Depending on the service, this may involve a senior clinician, an appropriately qualified specialist, another radiologist or an established multidisciplinary process.
A second opinion is not always definitive. In some situations, follow-up information or another diagnostic investigation may provide more reliable evidence than an immediate additional interpretation.
The final clinical decision should reflect the available evidence, the patient's circumstances and any uncertainty that remains.
Where a disagreement cannot be resolved immediately, the clinical team should establish an appropriate interim plan. This may include additional assessment, monitoring, follow-up or referral, depending on the potential risk.
Ensure that necessary clinical action and follow-up occur
Once an appropriate decision has been reached, any required investigation, treatment, communication or follow-up should proceed through the usual authorized clinical process.
Clinically important disagreements may also warrant later quality review, particularly where they reveal a missed finding, an unsupported AI recommendation or a weakness in how the system was implemented.
The review should consider not only which interpretation was eventually supported, but also whether the disagreement was identified, handled and resolved appropriately.
The objective is to establish a reliable process without converting every difference between clinician and AI into a formal patient-safety investigation.
Disagreements discovered after a clinical decision require particular attention
Some disagreements are recognized before a report is signed or a treatment decision is completed. Others are discovered only after clinical information has been released or acted upon.
The second situation may require a different response because patient care may already have been affected.
For example, an AI-supported quality-assurance process may identify a potentially important abnormality that was not included in a finalized radiology report.
The organization should determine whether the finding is clinically significant, whether the original information has been used in patient management and whether further professional assessment or communication is necessary.
Where a material error is confirmed, the appropriate clinical service should use a controlled correction or amendment process consistent with applicable documentation and professional requirements.
The corrected information should also reach the healthcare professionals who may have relied on the original result when the change has implications for patient care.
The Cross-Check QA study by Chekmeyan and colleagues illustrates how targeted review can identify selected initially missed findings and support appropriate report amendments and communication [5].
However, not every delayed AI alert represents a missed diagnosis, and not every apparent disagreement justifies recalling a patient or amending a report.
The response should reflect the evidence supporting the finding, the potential clinical consequences and the patient's circumstances.
Document clinically meaningful overrides without creating unnecessary work
Healthcare organizations should distinguish between an ordinary professional decision not to follow an AI suggestion and a disagreement that has meaningful patient-safety implications.
The term override is not used consistently across all clinical AI systems. In some applications, it refers to an explicit rejection of an AI recommendation. In others, it may describe dismissing an alert, choosing another action or simply not following the AI output.
These actions should not automatically be treated as equivalent.
A clinician may reject an AI suggestion because it is outside the system's intended purpose, because the clinical evidence does not support it or because patient-specific circumstances justify a different course of action.
Such a decision does not automatically indicate a safety incident.
Documentation should be proportionate to clinical significance and consistent with applicable medical-record requirements.
Where a disagreement materially influences care, it may be appropriate to record the finding in dispute, relevant evidence considered, any additional review undertaken and the final professional decision. Necessary patient follow-up or communication should also be documented through the applicable clinical process.
Organizations may maintain appropriate operational records of AI outputs and their disposition for quality-assurance purposes, subject to information-management and confidentiality requirements.
However, requiring lengthy explanations for every minor AI disagreement could create unnecessary workload and discourage meaningful engagement with the technology.
Documentation should support clinical accountability and learning, not become a mechanism for penalizing clinicians whenever they disagree with AI.
Monitoring should examine clinical significance, not just override rates
After implementation, organizations may measure how often clinicians accept or disagree with AI outputs.
These measures provide operational information, but they are not sufficient indicators of patient safety.
A high override rate may indicate that the AI system produces irrelevant or inaccurate recommendations. It could also reflect appropriate professional judgment, differences in the types of cases being evaluated or limitations in how the AI output is presented.
Similarly, a low override rate may reflect strong performance. It could also indicate insufficient scrutiny, overreliance on AI or a workflow that makes disagreement difficult.
The percentage alone cannot establish which explanation is correct.
Organizations should therefore interpret disagreement measures alongside the clinical significance of cases, the evidence supporting decisions and the actions taken.
Where feasible, monitoring should distinguish AI-positive/clinician-negative cases from clinician-positive/AI-negative cases. These two directions may reveal different issues involving system performance, clinical interpretation or workflow integration.
It may also be useful to examine how often clinically significant disagreements are resolved during initial review, how often further assessment is required and whether confirmed missed findings or inappropriate clinical actions occur.
The time required to resolve high-consequence disagreements can be informative, particularly where delays may affect treatment.
However, these measures need clear definitions and appropriate denominators. Differences in patient populations, disease prevalence, clinical workflow and the proportion of cases actually processed by AI can influence the results.
The organization should also distinguish cases in which clinicians viewed the AI output before making a decision from cases in which AI results were unavailable, delayed or not opened. Without this distinction, an apparent override rate may not accurately represent how clinicians responded to AI recommendations.
Organizations should avoid setting arbitrary targets for clinician acceptance or override rates without understanding the expected performance and clinical context of the application.
The aim is to evaluate whether disagreements are being managed safely, not to encourage clinicians to agree with AI more frequently.
High agreement does not necessarily mean that care is safe
Another limitation deserves attention: the clinician and AI may agree and still be wrong.
If both fail to identify an abnormality, the case will not appear in a dataset containing only disagreements.
Likewise, a high agreement rate can be influenced by the prevalence of negative cases. When the target condition is relatively uncommon, two systems may agree on many negative examinations while still differing substantially in their ability to detect important positive findings.
Organizations that examine only discordant cases may therefore overlook errors shared by clinicians and AI.
Where feasible, quality assurance should include appropriate evaluation of selected concordant cases, particularly those involving high-consequence decisions, together with relevant follow-up findings and clinical outcome information.
The reference used to evaluate performance also matters.
The American College of Radiology's Assess-AI registry, described by Coombs and colleagues in the Journal of the American College of Radiology in 2026, provides an example of infrastructure for monitoring real-world imaging AI performance. It compares AI outputs with surrogate labels derived from radiology reports and supports local investigation of discordant findings [6].
Such approaches may help organizations identify unusual patterns and select cases for review. However, report-derived labels are not necessarily independent proof that the underlying clinical interpretation is correct.
A radiology report may contain an error, and an AI system may agree with that error. Alternatively, disagreement may arise from differences in terminology or reporting conventions rather than from a clinically significant finding.
For this reason, clinically important discrepancies may require independent assessment using appropriate professional and diagnostic evidence.
The goal is not simply to maximize agreement. It is to determine whether the combined clinical workflow produces reliable decisions.
Use disagreement patterns to improve clinical safety
An effective monitoring process should help healthcare organizations understand why important disagreements occur.
Repeated AI-positive/clinician-negative cases may suggest that the AI system produces findings that are not clinically relevant in the local setting, that its operating conditions differ from those previously evaluated or that clinicians are interpreting its outputs differently than intended.
Repeated clinician-positive/AI-negative cases may indicate limitations in the AI system's detection performance for particular findings, patient groups or examination types.
However, these explanations should be investigated rather than assumed.
Some disagreements may result from technical problems or limitations in source information. Others may reflect differences in professional experience, examination protocols, disease presentation or the clinical context available to the clinician and AI.
Organizations should also consider whether important disagreement patterns vary across departments, facilities, shifts or patient groups.
An application may behave differently when used with examinations from another imaging device or among patients whose characteristics differ from those represented in the original evaluation.
Where a meaningful pattern is identified, the organization should determine whether the appropriate response involves additional staff training, workflow adjustments, further local evaluation, changes to permitted use or action involving the AI provider.
Not every disagreement requires modifying the AI model. In some cases, the problem lies in how information is presented or how clinical staff respond to it. In others, the system may not be suitable for the intended use.
Quality improvement should focus on the actual causes and assess whether corrective measures reduce the relevant risk.
These activities can be incorporated into existing diagnostic quality assurance, professional review and patient-safety arrangements rather than creating a separate governance structure for every AI disagreement.
What Viet Nam's legal framework means for clinician–AI disagreements
Viet Nam's Law on Medical Examination and Treatment No. 15/2023/QH15 provides the principal legal foundation for professional responsibilities and clinical decision-making.
Article 44 addresses practitioners' duties toward patients, including timely medical care, respect for patient rights and appropriate information and counseling. Article 45 establishes professional obligations, including compliance with professional and technical requirements and responsibility for the medical examination and treatment provided [7].
These responsibilities remain relevant when clinicians use AI-supported information in reaching diagnostic or treatment decisions.
Article 60 establishes responsibilities for healthcare establishments, including compliance with professional requirements, responsibility for examination and treatment results produced by practitioners under their management and the provision of necessary conditions for professional practice [7].
This matters because safe disagreement resolution depends not only on individual professional judgment but also on whether the organization has established suitable procedures, responsibilities and resources.
Article 64 governs professional consultation (hội chẩn) in specified circumstances, including cases beyond the diagnostic or treatment capabilities of a practitioner or healthcare establishment, or cases in which treatment is not progressing satisfactorily or the patient's condition is deteriorating [7].
Where formal consultation is conducted under this provision, the results must be documented in writing and retained in the medical record. The practitioner directly responsible for the patient makes the appropriate diagnostic or treatment decision based on the consultation results.
However, the mere existence of a disagreement between a clinician and AI does not automatically establish that the statutory circumstances requiring consultation under Article 64 are present.
A routine second interpretation or additional professional review used to clarify an AI-related finding should therefore not automatically be described as a legally required formal consultation.
Article 69 governs medical records, including their creation, updating, storage and confidentiality. Where a significant disagreement affects diagnosis, treatment or other clinically important information, the relevant documentation must be managed according to applicable medical-record requirements [7].
Article 71 establishes a framework for preventing medical incidents through identification, reporting, analysis of causes and preventive measures [7].
A clinically important disagreement may reveal a concern that warrants review. However, disagreement alone does not establish that a medical incident has occurred or that the circumstances meet specific incident-reporting criteria.
Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 has been effective since 1 March 2026. Decree No. 142/2026/NĐ-CP, effective from 1 May 2026, further details parts of the AI regulatory framework, including requirements that depend on the nature of the AI system and the roles of relevant participants [8,9].
These requirements must be assessed according to the system's actual intended purpose, functionality, legal classification and conditions of use.
The existing legal framework does not prescribe one universal AI-specific procedure requiring identical override documentation, second-review arrangements or disagreement thresholds across all clinical specialties and AI applications.
The Clinical–AI Discordance Resolution pathway discussed here is therefore intended as a practical organizational approach that supports safe clinical practice alongside applicable legal and professional obligations.
Making disagreement resolution workable in Vietnamese healthcare organizations
The appropriate process will depend on the healthcare setting and the AI application involved.
In a large hospital, significant disagreements may be reviewed by appropriately qualified specialists or managed through existing diagnostic quality-assurance arrangements.
In smaller facilities, immediate access to a second specialist may not always be available. The organization may need to determine when additional investigation, referral or another safe clinical pathway is more appropriate.
Responsibilities should be established before routine implementation.
Clinicians need to understand the AI system's intended use, the meaning and limitations of positive and negative outputs, and the circumstances in which a discrepancy requires additional assessment.
Clinical leaders should help define which disagreements can be resolved through ordinary professional assessment and which require further review.
Information technology personnel and AI providers should ensure that relevant outputs and supporting information can be retrieved when a clinically important disagreement needs investigation.
Quality and patient-safety teams can contribute by reviewing selected cases, identifying recurring problems and assessing whether corrective actions have improved the workflow.
The process should also account for actual clinical workload.
An AI system that generates frequent low-value alerts may interrupt care and reduce the attention clinicians can devote to important findings. Conversely, a workflow that makes significant discrepancies difficult to recognize or review may fail to provide the intended safety benefit.
Before wider deployment, healthcare organizations should test the proposed disagreement process using representative clinical scenarios. These should include cases in which AI is correct and the clinician is wrong, cases in which the clinician is correct and AI is wrong, and cases in which neither assessment can be confirmed immediately.
Testing should also establish whether the workflow can distinguish genuine disagreements from failed AI processing or findings outside the system's intended scope.
The organization should determine whether important disagreements are recognized, whether additional review is available when justified and whether necessary clinical actions occur without avoidable delay.
The value of the process should ultimately be judged by how reliably it supports clinical decisions and patient care, not simply by whether it produces another checklist or record of AI use.
From disagreement to safer clinical decisions
Clinical AI is valuable when it helps healthcare professionals make better decisions, not simply when its recommendations are accepted.
A disagreement between AI and a clinician may represent an unnecessary alert, an AI limitation, a missed clinical finding or legitimate uncertainty in a complex case. Its meaning depends on the underlying evidence and clinical circumstances.
Healthcare organizations therefore need an approach that is proportionate to risk.
Some disagreements can be resolved through ordinary professional assessment. Others justify additional review, investigation or follow-up. Clinically significant cases may also reveal information that helps improve the AI-enabled workflow.
Neither the clinician's initial opinion nor the AI output should be treated as an automatic reference standard. Both require appropriate evaluation when the consequences of a wrong decision matter.
For healthcare organizations in Viet Nam, the goal should be to establish a practical process in which clinicians can question AI, clinically important concerns are addressed, decisions remain appropriately supported and lessons from meaningful disagreements contribute to improvement.
The purpose of governing clinician–AI disagreement is not to make clinicians agree with AI more often. It is to ensure that disagreements are resolved in ways that support better clinical judgment, safer decisions and continuous learning in patient care.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized medical, diagnostic, legal, regulatory, patient-safety, accreditation or quality-management advice.
The proposed Clinical–AI Discordance Resolution pathway is a practical organizational approach. It is not a prescribed Vietnamese Ministry of Health procedure, mandatory international arbitration standard or independently validated clinical AI assessment framework.
Recommendations concerning disagreement classification, risk-proportionate second review, documentation, monitoring and clinical safety learning should not be interpreted as identical statutory requirements for every healthcare organization or AI application.
Applicable responsibilities depend on the clinical service, the AI system's intended purpose, functionality and regulatory status, professional scope of practice, the consequences of clinical decisions and relevant legal requirements.
The international studies discussed involve different clinical settings, technologies and evaluation methods. Their findings should not be generalized automatically to other specialties, AI applications or healthcare organizations in Viet Nam.
The proposed practices are intended to support existing clinical quality, professional accountability and patient-safety responsibilities, not replace clinical judgment or applicable legal requirements.
References
National Institute for Health and Care Excellence. Artificial Intelligence (AI) Technologies to Help Detect Fractures on X-Rays in Urgent Care: Early Value Assessment. HealthTech Guidance HTG739. Published 14 January 2025.
National Institute for Health and Care Excellence. Evidence Generation Plan for Artificial Intelligence (AI) Technologies to Help Detect Fractures on X-Rays in Urgent Care. Supporting guidance for HTG739. January 2025.
Goldberg-Stein S, Gandomi A, Barish MA, et al. Clinical Implementation of AI for Pulmonary Embolism Detection in over 30,000 CT Pulmonary Angiography Examinations. Radiology: Artificial Intelligence. 2026;8(4). DOI: 10.1148/ryai.250017.
Topff L, Steltenpool S, Ranschaert ER, et al. Artificial Intelligence-Assisted Double Reading of Chest Radiographs to Detect Clinically Relevant Missed Findings: A Two-Centre Evaluation. European Radiology. 2024;34(9):5876–5885. DOI: 10.1007/s00330-024-10676-w.
Chekmeyan M, Baccei SJ, Garwood ER. Cross-Check QA: A Quality Assurance Workflow to Prevent Missed Diagnoses by Alerting Inadvertent Discordance Between the Radiologist and Artificial Intelligence in the Interpretation of High-Acuity CT Scans. Journal of the American College of Radiology. 2023;20(12):1225–1230. DOI: 10.1016/j.jacr.2023.06.010.
Coombs LP, Brink L, Kim W, et al. ACR's Assess-AI: A Registry for Real-World Performance Monitoring of Clinical Imaging Artificial Intelligence. Journal of the American College of Radiology. 2026;23(8):1557–1566. DOI: 10.1016/j.jacr.2026.04.024.
National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment, as amended. Adopted 9 January 2023; effective 1 January 2024.
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.