Evaluating AI-Assisted Diagnostic Systems in Viet Nam: Local Validation, Human Oversight and Post-Deployment Monitoring
MEASUREMENT, LEARNING & IMPROVEMENTDIAGNOSTIC SAFETY & RISK MANAGEMENT
11/5/202628 min read


Evaluating AI-Assisted Diagnostic Systems in Viet Nam: Local Validation, Human Oversight and Post-Deployment Monitoring
Consider a hospital that introduces an artificial intelligence system to help identify abnormalities on medical images. The vendor presents evidence of strong diagnostic performance, the software integrates with the hospital's imaging system, and clinicians begin using its findings during routine reporting.
Initially, the technology appears to work as expected. Several months later, however, the hospital introduces new imaging equipment and changes some examination protocols. The characteristics of patients undergoing these examinations also begin to change. The AI system continues producing results, but no one has established whether its performance remains consistent with the evidence available before implementation.
The software may not have malfunctioned. It may still be operating exactly as designed. Yet the reliability of its outputs can change when clinical conditions differ from those under which the system was originally evaluated.
This hypothetical situation illustrates a fundamental challenge in AI-assisted diagnosis. Good performance during development or validation does not guarantee equally reliable performance in every healthcare setting or throughout a system's operational life.
Artificial intelligence is increasingly being used to support the detection, classification and interpretation of diagnostic findings. Applications include medical imaging, digital pathology, electrocardiographic analysis and selected forms of clinical decision support. These technologies may help healthcare professionals recognize abnormalities, prioritize examinations and manage diagnostic information more efficiently.
Their clinical value, however, depends on more than the algorithm's reported accuracy. The diagnostic task, quality of the supporting evidence, characteristics of the patient population and way professionals interact with the technology all influence whether AI contributes meaningfully to patient care.
Regulatory authorization, where applicable, provides an important part of the compliance framework. It does not remove the need for healthcare organizations to understand whether a particular system is suitable for its intended use within their clinical environment.
Recent studies published in 2026 have provided further evidence of variations in AI performance across imaging equipment, the sometimes modest benefits of AI-assisted interpretation and changes in model behavior after clinical deployment.
These findings are relevant to healthcare organizations in Viet Nam, where diagnostic services differ considerably in equipment, digital infrastructure, staffing and clinical workflows. The country's evolving AI regulatory framework also places greater emphasis on patient safety, human oversight and reliability under real-world conditions.
Responsible evaluation therefore requires more than reviewing a product demonstration or a published accuracy figure. Healthcare organizations need to determine whether the technology performs adequately under local conditions, whether clinical professionals can meaningfully assess its outputs and whether its reliability can be maintained after implementation.
From Published Accuracy to Clinical Reliability
AI-assisted diagnostic systems are frequently evaluated using sensitivity, specificity and the area under the receiver operating characteristic curve. These measures provide useful information about an algorithm's ability to distinguish between specified clinical categories.
Their meaning, however, depends on how the evaluation was conducted.
A model assessed using carefully selected examinations may behave differently when exposed to the full range of cases encountered in routine practice. Medical images may contain artifacts, technical limitations or unusual findings that were uncommon in the development dataset. Patients may also have characteristics or conditions that were excluded from earlier studies.
The reference standard used to establish whether a finding is present matters as well. If the comparison diagnosis is uncertain or inconsistently determined, reported performance may not provide a dependable estimate of the algorithm's accuracy.
Differences in disease prevalence, clinical presentation and severity can further influence the interpretation and consequences of AI outputs.
Consider a system developed to identify particular abnormalities on radiographs. False-positive findings may lead to additional review or investigations without corresponding clinical benefit. False-negative outputs may create greater concern if clinicians assume that an examination without an AI alert contains no clinically important abnormality.
The consequences depend on the system's intended role, the clinical context and how healthcare professionals use its findings.
This is why standalone AI performance must be distinguished from AI-assisted clinical performance.
Standalone performance describes how an algorithm performs against an appropriate reference standard. AI-assisted performance examines what happens when professionals use the technology within a defined clinical workflow.
The two are not necessarily equivalent. An algorithm may perform well independently but provide limited additional value to experienced clinicians. Another system may have imperfect standalone performance yet still improve the clinical team's ability to recognize selected abnormalities.
The relevant question is whether introducing the technology makes the diagnostic process more reliable or provides another meaningful clinical benefit, rather than simply whether the algorithm achieves a high accuracy score.
What Recent Studies Tell Us About AI-Assisted Diagnosis
A major umbrella review published in npj Digital Medicine on October 27, 2026, examined the evidence concerning artificial intelligence for fracture detection.
Kuo and colleagues included 50 systematic reviews and meta-analyses covering 387 primary studies. The pooled estimates suggested sensitivity of approximately 89.7% and specificity of 90.4% for AI fracture detection.
These figures appear encouraging, but the broader findings require caution.
Evidence directly comparing AI with clinicians or evaluating clinicians assisted by AI was sparse and inconclusive. The certainty of evidence was rated low or very low for almost all assessed outcomes, and the methodological quality of the included meta-analyses was generally low or critically low.
None of the included reviews reported adverse events or other patient safety outcomes. This represents an important gap in the evidence, not proof that AI-assisted fracture detection is free from potential harm.
The review therefore highlights a distinction that healthcare organizations should not overlook. Relatively high diagnostic accuracy does not automatically demonstrate that introducing AI into clinical practice improves patient outcomes.
A separate study by Herpe and colleagues, published in the American Journal of Roentgenology in 2026, examined a commercial fracture detection system deployed across three emergency departments.
The researchers identified significant variation in AI performance among radiographs obtained using equipment from seven manufacturers.
This finding provides a practical reason for local evaluation. Evidence generated in one technical environment may not be fully transferable to hospitals operating different imaging systems or acquisition protocols.
The study does not establish that every radiology AI system is affected by equipment differences to the same extent, nor does it necessarily establish that equipment manufacturer alone explains every observed difference. It does, however, demonstrate why compatibility and performance under actual operating conditions deserve attention.
Another prospective investigation, published in npj Digital Medicine on October 27, 2026, examined AI-assisted fracture detection during routine emergency department practice.
Tan and colleagues evaluated 2,496 consecutive musculoskeletal examinations interpreted by 13 radiographers in a radiographer-led screening program.
The standalone AI system demonstrated sensitivity of 75.6%, which decreased to 46.2% for subtle fractures. The participating radiographers had baseline sensitivity of 91.6%, while their overall accuracy improved modestly from 97.6% to 98.0% with AI assistance.
The study also found that no AI output or incomplete output was returned for 8.8% of acquired examinations.
These findings are informative because they reveal aspects of clinical performance that a single accuracy statistic cannot capture.
The algorithm had lower standalone sensitivity than the radiographers, yet its assistance produced a small improvement in their overall accuracy. At the same time, incomplete or unavailable outputs represented an operational limitation that could affect the technology's usefulness in routine practice.
The investigators did not measure downstream patient outcomes, and their findings should not be generalized to other diagnostic AI systems without appropriate evidence.
Taken together, these studies suggest that AI can provide useful support for selected diagnostic tasks, but its incremental benefit depends on the technology, clinical environment, professional expertise and workflow.
Strong algorithmic performance is an important consideration. It is not sufficient evidence, by itself, that a system will improve diagnostic care.
Defining the Intended Use Before Evaluating the Technology
A healthcare organization cannot meaningfully evaluate an AI-assisted diagnostic system without first establishing what the technology is intended to do.
This requires a clear understanding of the clinical task, target population, input data, expected outputs and role of the system in the diagnostic process.
An application developed to detect selected fractures on conventional radiographs should not automatically be used to evaluate every musculoskeletal examination. Similarly, a system designed to recognize abnormalities on adult chest radiographs should not be assumed suitable for pediatric imaging or other modalities.
Even systems working with similar examinations may serve different purposes.
One application may flag suspected abnormalities for professional review. Another may prioritize examinations according to the likelihood of urgent findings. A third may estimate the probability of a condition to support clinical assessment.
These functions have different clinical consequences.
An examination prioritization system can influence how quickly certain patients receive attention even when it does not issue a diagnosis. A classification tool may affect a clinician's interpretation, while a probability model may influence decisions based on estimated risk.
The evaluation must therefore reflect how the system's outputs are expected to influence clinical practice.
Healthcare organizations should review the manufacturer's intended use, documented limitations and relevant regulatory status before determining whether the proposed application is appropriate.
They also need to understand what happens when the system receives unsuitable inputs.
Some AI applications recognize unsupported examinations or inadequate data and decline to produce an output. Others may generate a prediction despite poor image quality, missing information or unexpected data formats.
A system that appropriately recognizes when it cannot provide a reliable assessment may be safer than one that consistently generates apparently confident outputs under unsuitable conditions.
Understanding these limitations is essential before integrating AI into a diagnostic workflow.
Local Validation: Establishing Fitness for the Clinical Setting
A hospital may have extensive manufacturer documentation and access to published studies but still face uncertainty about whether an AI system will perform adequately with its own patients, equipment and workflows.
Local validation can help address this uncertainty.
In this article, local validation refers to a risk-based assessment of whether the available evidence and operational characteristics of an AI-assisted diagnostic system are suitable for its intended clinical use within a particular healthcare organization.
It should not be confused with formal regulatory conformity assessment, nor does it imply that every hospital must conduct an independent clinical trial before implementing every AI product.
The appropriate evaluation depends on the clinical purpose, potential consequences of error, available evidence, applicable regulations and changes introduced during local implementation.
For diagnostic imaging, local assessment may consider examinations obtained using the hospital's actual scanners, acquisition protocols and image-processing systems.
For an AI application interpreting electrocardiographic information, relevant factors may include recording devices, signal quality, patient characteristics and the way outputs are incorporated into clinical review.
Systems processing clinical text may face additional challenges involving Vietnamese language usage, medical terminology, abbreviations, documentation practices and data completeness.
These issues are relevant in Viet Nam, where healthcare facilities differ substantially in their digital capabilities, equipment and operational arrangements.
A meaningful evaluation should begin with a clearly defined clinical question. The organization needs to determine what performance is necessary for the intended application and which failures could have the greatest clinical consequences.
The cases selected for evaluation should reasonably represent the patients and examinations the system is expected to encounter.
Where feasible, consecutive or systematically selected examinations may provide a more realistic assessment than datasets consisting mainly of obvious positive and negative examples.
Relevant patient subgroups, technical variations and uncommon but clinically consequential presentations should also be considered.
The quality of the reference standard is particularly important. Depending on the diagnostic task, an appropriate reference may involve qualified expert interpretation, adjudication of disagreements, pathology findings, longitudinal clinical information or another justified method.
Unreliable reference labels can compromise an evaluation regardless of the statistical techniques applied.
Data independence must also be maintained. Cases used to train or fine-tune a model, or to select its operating threshold, should not subsequently be presented as an independent evaluation of the final system.
Sample size and statistical uncertainty deserve similar attention. A small dataset may produce favorable accuracy estimates while providing little assurance about performance in uncommon clinical conditions.
Confidence intervals, the number of positive and negative cases and subgroup sample sizes should therefore be considered when interpreting the findings.
Local evaluation should also examine operational reliability.
A model may perform well on images it successfully processes but be unsuitable for clinical use if results are frequently unavailable, arrive too late or cannot be reliably linked to the correct examination and patient.
The purpose of local validation is not simply to reproduce the manufacturer's published performance figures. It is to establish whether the technology can be used appropriately and dependably under the conditions in which the organization intends to deploy it.
Choosing Performance Measures That Reflect Clinical Risk
Different AI systems require different performance measures.
For diagnostic classification, sensitivity and specificity may provide useful information about the ability to identify cases with or without the target condition.
Positive and negative predictive values can help explain the meaning of results in a particular population, but they depend on disease prevalence.
A model evaluated in a selected research population may have a substantially different positive predictive value when used in a lower-prevalence screening setting.
The area under the receiver operating characteristic curve describes discrimination across different thresholds. However, a high area under the curve does not establish that the threshold selected for clinical use is appropriate.
The operating threshold influences the balance between false-positive and false-negative findings. Its suitability depends on the clinical purpose and the consequences of different errors.
For systems producing individual probability estimates, calibration is another important consideration.
A model may correctly distinguish patients at higher and lower risk while systematically overestimating or underestimating the probability of an outcome. Such a system may demonstrate acceptable discrimination but provide misleading absolute risk estimates.
Calibration is particularly relevant when predicted probabilities influence clinical thresholds or risk categories. It is not necessarily the primary performance measure for every form of AI image detection.
Evaluation should also include failed, missing or uninterpretable outputs.
A system may perform well among examinations for which it generates a usable result while failing to process a clinically important proportion of routine examinations. Excluding these failures without explanation can create an overly favorable impression of reliability.
Subgroup performance also matters. Differences associated with age, sex, disease severity, imaging equipment or other relevant characteristics may affect clinical usefulness.
However, subgroup findings based on small numbers can be statistically unstable and should not be overinterpreted.
A credible evaluation should explain what the system does well, where its performance is uncertain and what those limitations mean for its intended clinical use.
Human Oversight Must Be Meaningful
Human oversight is frequently presented as a central safeguard for AI-assisted diagnosis.
A healthcare organization may require a qualified clinician to review an AI output and retain responsibility for the final clinical decision. This is an important principle, but the presence of a human reviewer does not automatically establish that the system is being used safely.
Oversight is meaningful only when the professional has the knowledge, information and practical opportunity to evaluate the output and intervene when necessary.
Clinicians need to understand the system's intended use, relevant limitations and circumstances in which its findings may be unreliable.
The workflow must also allow sufficient opportunity to examine the underlying clinical information rather than simply approve an automated recommendation.
A professional expected to review large numbers of AI-generated findings under severe time pressure may have limited ability to recognize errors, even when formal responsibility remains with that individual.
Human oversight should therefore be evaluated as an operational capability, not merely documented as a professional obligation.
One concern is automation bias, in which users may place excessive confidence in computerized recommendations.
A clinician reviewing an image labeled negative by AI may become less likely to recognize a subtle abnormality that the system did not flag.
The opposite problem can also occur. Repeated false alerts, unfamiliarity with the technology or insufficient trust may cause clinicians to disregard useful AI findings.
Both patterns can reduce the potential value of the system.
The presentation of AI information may influence these behaviors.
A system that displays findings before the clinician begins interpretation may affect the review process differently from one that provides a second opinion after an initial independent assessment.
Neither approach is universally superior. Its suitability should be evaluated according to the diagnostic task, user experience and available evidence.
Interface design also matters. Confidence scores, colored overlays, heatmaps and automated explanations can influence how users interpret an AI output.
However, these features should not automatically be regarded as reliable explanations of the model's reasoning or proof that its conclusions are correct. A displayed confidence score may also require appropriate evaluation before it can be interpreted as a calibrated clinical probability.
Training should therefore address appropriate reliance on AI rather than focus only on software operation.
Healthcare professionals may need guidance on known failure patterns, circumstances requiring independent review, appropriate responses to uncertain findings and procedures for reporting suspected problems.
The organization should establish who can challenge, disregard or override an AI recommendation and how concerns are investigated.
Where clinically appropriate, professionals should also be able to continue the relevant diagnostic assessment when the AI system is unavailable or produces an unreliable output.
Meaningful oversight becomes particularly important when similar problems are observed across multiple examinations. A single discrepancy may be isolated, while repeated findings may indicate a broader technical or operational problem.
The quality system should allow those patterns to be recognized and addressed.
Evaluating the Human-AI Team and Clinical Value
The same AI system may have different effects when used by clinicians with different experience, responsibilities and workflows.
An application that provides additional value to one clinical team may offer little improvement in another setting where baseline performance is already high.
The prospective fracture detection study by Tan and colleagues illustrates this distinction. The participating radiographers demonstrated high baseline diagnostic performance, and AI assistance produced a small improvement despite the algorithm's lower standalone sensitivity.
This finding does not imply that AI assistance is generally ineffective. Rather, it suggests that the technology's incremental contribution depends on how it interacts with the people using it.
For some applications, the expected benefit may involve improved recognition of selected abnormalities. For others, it may involve more appropriate prioritization of urgent examinations, shorter reporting delays or reduced workload.
These outcomes should be evaluated separately.
A reduction in reporting time does not necessarily demonstrate improved diagnostic accuracy. Similarly, identifying additional abnormalities does not automatically establish that fewer patients experience diagnostic harm.
Where feasible and justified, prospective evaluation of the human-AI workflow may provide useful information beyond retrospective algorithm testing.
An initial assessment may involve operating the system in a non-interventional or silent mode, where AI outputs are recorded and evaluated without being used to direct clinical care.
Such an approach can help characterize performance before routine deployment. However, silent-mode evaluation is not automatically exempt from applicable institutional approval, research ethics, privacy or other regulatory requirements.
Clinical implementation should proceed when the available evidence, safeguards and permissions are appropriate for the intended application.
The evaluation should also consider potential unintended consequences.
False-positive alerts may increase unnecessary investigations or referrals. False-negative outputs may create unwarranted reassurance. Excessive alerts may reduce attention to important findings, while growing dependence on AI may alter how professionals examine the underlying clinical information.
Costs and resource requirements are relevant as well.
A system that reduces reporting time may still require substantial staff effort for monitoring, incident investigation, technical maintenance and periodic reassessment. These activities are part of responsible implementation and should be considered when evaluating overall value.
Where appropriate clinical outcome data are available, they should be examined alongside technical and workflow measures.
However, an improvement in a process indicator should not automatically be presented as proof of fewer diagnostic errors or better patient outcomes.
The objective is to establish whether the technology contributes to meaningful clinical improvement, not simply whether the algorithm performs accurately when examined in isolation.
Why Validation Before Deployment Is Not Enough
Even an AI system that performs adequately during initial validation may behave differently after several months or years of routine use.
Patient populations evolve, equipment is replaced, examination protocols change and information systems are updated. The completeness, timing or structure of data entering the model may also change.
These developments can influence performance without any modification to the algorithm itself.
A study published in PLOS Digital Health on July 27, 2026, examined this problem.
Kopanitsa conducted a longitudinal retrospective observational study involving four clinically deployed AI systems operating across different clinical domains and workflows within a large healthcare organization.
The investigation compared performance observed during validation with behavior after deployment.
Changes in calibration emerged during routine use and often preceded detectable changes in discrimination. Operational signals involving missing input data, delayed information and changes in feature availability were also associated with performance deterioration.
These findings matter because confirmed diagnostic outcomes may not become available immediately.
Monitoring based entirely on outcome labels can therefore identify some problems only after considerable delay. Operational information may offer earlier indications that the system is encountering conditions different from those present during validation.
The findings should nevertheless be interpreted within the study's limitations. It was observational and involved four systems within one healthcare organization. It does not establish that every clinical AI model will deteriorate or that a single monitoring strategy is appropriate for all applications.
Its broader contribution is to demonstrate why acceptable pre-deployment performance should not be treated as a permanent guarantee of reliability.
For healthcare organizations, monitoring arrangements should be considered before an AI application enters routine use rather than introduced only after a serious problem occurs.
Monitoring Model Performance and Clinical Conditions
Post-deployment monitoring should examine both the diagnostic performance of the AI system and the conditions under which it operates.
The appropriate measures depend on the model's intended use and the availability of reliable outcome information.
Where suitable reference diagnoses or outcome labels can be obtained, the organization may periodically evaluate sensitivity, specificity, calibration or other relevant performance measures.
However, confirmed outcomes may be delayed, incomplete or available only for selected patients.
For example, definitive diagnostic confirmation may occur primarily among patients receiving additional investigations. Performance estimates based only on these patients may not represent the entire population exposed to the AI system.
Verification and selection bias should therefore be considered when interpreting monitoring results.
Operational measures provide complementary information.
Relevant indicators may include the proportion of examinations receiving usable AI outputs, processing failures, data completeness, system latency, equipment compatibility and unexpected changes in the distribution of outputs.
For imaging applications, changes in scanner models, acquisition protocols or image-processing software may be particularly important. Systems using clinical records may be affected by changes in documentation practices, data coding or variable availability.
Different forms of change should be distinguished.
Data drift generally refers to changes in the distribution or characteristics of input data. Concept drift may occur when the relationship between inputs and the clinical outcome being predicted changes. Calibration drift concerns changes in the agreement between predicted probabilities and observed outcomes.
These concepts are related but not interchangeable.
Not every AI system produces probabilities requiring calibration assessment, and a change in input data does not automatically establish clinically important performance deterioration.
Monitoring should therefore connect observed changes with their potential clinical implications.
Healthcare organizations need to establish which findings warrant further investigation and what actions may be appropriate when performance falls outside acceptable limits.
Universal numerical thresholds are unlikely to be suitable for every application.
A small fluctuation in a low-volume service may reflect statistical uncertainty, while a sustained change in a high-consequence diagnostic application may require more immediate attention.
Software updates, model replacements and modifications to input processing should also be included in monitoring arrangements.
An apparently minor technical change can affect clinical performance even when the system's name and interface remain unchanged.
The organization should know which model version is operating, what has changed and whether additional evaluation is necessary before the updated system is used clinically.
Vendor Responsibilities, System Changes and Incident Response
AI-assisted diagnostic systems often depend on software and technical services outside the healthcare organization's direct control.
A developer may train and maintain the model, a commercial provider may supply the software platform, and the hospital may integrate the system into clinical practice.
These arrangements create distinct responsibilities that need to be understood before implementation.
Healthcare organizations should obtain sufficient information about the system's intended use, compatible equipment, known limitations, validation evidence and version history.
Service agreements may also need to address software changes, technical support, access to relevant operational records, incident notification and cooperation during investigations.
The allocation of responsibilities should reflect the actual service arrangement and applicable law.
Not every software update requires a new clinical validation study.
Some changes may be addressed through appropriate compatibility checks or technical verification. Others may require additional clinical assessment because they materially affect the model, its input processing, outputs or intended use.
The response should be proportionate to the potential consequences.
An organization should not assume that an externally managed AI system can change in clinically significant ways without any local review.
Incident management is equally important.
Potential AI-related incidents may involve clinically significant false-negative findings, incorrect examination prioritization, software failures, data mapping errors or unexpected performance changes.
These events are different and may require different investigations.
A disagreement between an AI system and a clinician does not automatically establish that the algorithm malfunctioned. Likewise, an incorrect AI output does not necessarily mean that an incorrect diagnosis was made or that a patient suffered harm.
An appropriate review should examine the clinical and technical circumstances, information available to users, the intended function of the system and whether its output influenced patient management.
Where a broader problem is suspected, the organization may need to introduce additional professional review, temporarily restrict use, revert to an appropriately approved previous version or suspend the system while the concern is investigated.
The appropriate action depends on clinical risk, available alternatives and applicable regulatory and organizational requirements.
A dependable AI diagnostic service should have practical arrangements for maintaining appropriate care when the technology cannot be used reliably.
International Guidance for Responsible Evaluation
Several international resources can help healthcare organizations evaluate AI-assisted diagnostic systems.
The World Health Organization published Regulatory Considerations on Artificial Intelligence for Health in October 2023. The publication examines important considerations for healthcare AI, including safety, effectiveness, transparency, data quality, risk management and oversight throughout the system lifecycle.
The International Medical Device Regulators Forum (IMDRF) published Good Machine Learning Practice for Medical Device Development: Guiding Principles in January 2025.
The document identifies ten principles intended to support the development of safe, effective and high-quality AI-enabled medical devices.
These principles address multidisciplinary expertise, intended use, software engineering, representative data, appropriate training and testing practices, clinically relevant evaluation, human-AI team performance, user information and ongoing monitoring.
They are particularly relevant when reviewing the evidence and quality arrangements established by developers of medical AI systems.
The IMDRF also conducted a public consultation between April and July 2026 on a proposed Technical Framework for Artificial Intelligence Life Cycle Management.
As of November 2026, this consultation document should be distinguished from finalized guidance. It represents continuing international work rather than a completed regulatory requirement.
Scientific reporting guidelines provide additional support for evaluating published evidence.
The STARD-AI statement, published in Nature Medicine in September 2025, provides reporting recommendations for diagnostic accuracy studies involving artificial intelligence.
It addresses AI-specific considerations relevant to dataset characteristics, evaluation methods, bias, fairness and the applicability of study findings.
For clinical prediction models, the TRIPOD+AI statement published in The BMJ in 2024 provides reporting guidance for studies developing or evaluating multivariable diagnostic or prognostic prediction models using regression or machine learning methods.
These documents have different scopes.
STARD-AI is particularly relevant to AI-centered diagnostic test accuracy studies, while TRIPOD+AI applies to clinical prediction model studies within its defined scope.
Neither guideline replaces appropriate clinical evaluation, regulatory authorization or continuing performance monitoring.
Their role is to improve transparency and help readers evaluate the strengths, limitations and applicability of published evidence.
International guidance provides a useful foundation for quality improvement, but it should not automatically be treated as establishing mandatory legal obligations for every healthcare organization in Viet Nam.
Evaluating AI-Assisted Diagnosis Within Viet Nam's Legal Framework
The introduction of AI-assisted diagnostic systems in Viet Nam must be considered within the applicable legal framework for artificial intelligence, medical examination and treatment, medical devices and personal data protection.
The Law on Artificial Intelligence No. 134/2025/QH15 was adopted on December 10, 2025 and took effect on March 1, 2026.
The law establishes principles governing the development, provision, deployment and use of AI systems, including risk management, human oversight, safety, transparency and accountability.
Article 4 sets out fundamental principles for AI activities, including a human-centered approach, human oversight and intervention capabilities, information security and responsibility for AI-related activities.
Article 6 addresses the application of AI across different sectors.
For healthcare, it specifically emphasizes patient safety, reliability under real-world conditions of use and protection of health data in accordance with applicable law.
These principles are directly relevant to AI-assisted diagnosis because the outputs of such systems may influence clinical interpretation, examination prioritization and patient management.
The law also establishes a risk-based classification framework.
Article 9 distinguishes high-risk, medium-risk and low-risk AI systems according to their potential impacts and other relevant considerations.
Article 10 addresses the classification and notification of AI systems.
Providers are responsible for classifying their systems before placing them into use. Systems classified as medium-risk or high-risk must have classification documentation, and providers must notify the Ministry of Science and Technology of the classification results through the designated AI single-window portal before the systems are put into use, subject to the applicable provisions and transitional arrangements.
Deployers may rely on the provider's classification but remain responsible for safety and system integrity during operation. Where modification, integration or changes in function introduce new or higher risks, coordination with the provider may be required to reassess the classification.
These obligations must be interpreted according to the applicable legal provisions and the actual roles of the organizations involved.
For high-risk AI systems, Articles 13 and 14 establish additional requirements concerning conformity assessment, risk management, technical documentation, human oversight, monitoring and the responsibilities of providers and deployers.
Importantly, not every AI system used in healthcare should automatically be classified as high-risk solely because it operates within a clinical setting.
The applicable classification depends on the system's function, risk characteristics and relevant legal provisions.
Decree No. 142/2026/ND-CP, issued on April 30, 2026 and effective May 1, 2026, provides further implementing provisions concerning AI risk classification, conformity assessment, transparency, incident management and related responsibilities.
Decision No. 33/2026/QD-TTg, issued on June 30, 2026 and effective August 15, 2026, establishes a list of high-risk AI systems.
Within healthcare, the published list identifies specified AI-assisted surgical and robotic surgical systems subject to defined conditions.
This should not be interpreted as automatically classifying every diagnostic imaging AI application, laboratory AI system or clinical prediction tool as high-risk.
Equally, the absence of a particular diagnostic application from an explicitly listed healthcare category should not be treated as proof that no AI-related or sector-specific obligations apply.
Healthcare organizations should assess the applicable classification and compliance requirements according to the system's actual functions, intended purpose and legal context.
For high-risk systems, regulatory conformity assessment must also be distinguished from local clinical evaluation.
Article 13 of the Law on Artificial Intelligence provides different conformity assessment arrangements depending on whether a system falls within the category requiring mandatory external conformity certification before use.
For specified systems, assessment must be undertaken by an appropriately registered or recognized conformity assessment organization. Other high-risk systems may be subject to provider self-assessment or assessment through an appropriate external organization, in accordance with applicable requirements.
These provisions are important because a hospital's internal evaluation of clinical suitability does not automatically replace a legally required conformity assessment.
Conversely, regulatory conformity assessment does not eliminate the need to manage clinical risks associated with the actual deployment environment.
Medical Device Regulation and Professional Responsibilities
AI legislation does not replace the medical device regulatory framework.
Some AI-assisted diagnostic software may qualify as medical devices according to their intended medical purpose and applicable legal definitions.
Where this is the case, healthcare organizations should consider the relevant requirements concerning medical device classification, market authorization, documentation, intended use and other applicable obligations.
Decree No. 98/2021/ND-CP on medical device management, as amended and reflected in Consolidated Document No. 08/VBHN-BYT dated March 6, 2026, provides an important reference for medical device regulation in Viet Nam.
Medical device classification and AI system risk classification address related but distinct regulatory questions.
A system's classification under the Law on Artificial Intelligence should not automatically be assumed to determine its medical device classification, or vice versa.
Similarly, not every software application using AI within a healthcare organization necessarily qualifies as a medical device.
The intended purpose and applicable legal criteria must be considered.
The Law on Medical Examination and Treatment No. 15/2023/QH15, as amended and reflected in Consolidated Document No. 26/VBHN-VPQH dated February 26, 2026, provides the broader framework governing professional healthcare activities.
Healthcare facilities and practitioners remain subject to applicable requirements concerning authorized clinical activities, professional responsibilities and medical records.
Article 69 addresses the creation, updating, preservation and confidentiality of medical records.
Where AI outputs are incorporated into diagnostic reports or patient records, the healthcare organization should establish how those outputs are identified, reviewed and documented according to applicable requirements.
The use of AI should not obscure the identity or responsibilities of appropriately qualified professionals undertaking the relevant clinical activities.
The precise allocation of legal responsibilities depends on the technology, clinical service and applicable legal and organizational arrangements. It should not be inferred solely from the presence of an AI recommendation or an incorrect output.
Transitional Provisions for AI Systems
Article 35 of the Law on Artificial Intelligence establishes transitional provisions for systems already operating before the law took effect.
For relevant AI systems in healthcare, education and finance that were operating before March 1, 2026, the law provides an 18-month period from that date for providers and deployers to fulfill applicable compliance obligations.
Decision No. 33/2026/QD-TTg also establishes transitional arrangements for systems included in its high-risk list.
For covered healthcare, education and finance systems already operating before August 15, 2026, the decision specifies a compliance deadline before September 1, 2027.
The decision contains a separate transitional provision for covered systems brought into operation during the six months following its effective date. For those systems, the specified deadline is before March 1, 2027.
These provisions apply according to their stated conditions and should not be interpreted as granting a general exemption to every newly introduced AI system.
They also do not remove responsibilities under other applicable legislation or prevent competent authorities from requiring action where a system presents a serious risk.
Healthcare organizations operating AI-assisted diagnostic applications should therefore identify the systems they use, determine the applicable legal obligations and establish appropriate compliance arrangements.
Regulatory transition should be managed alongside continuing patient safety oversight, rather than replacing it.
Protecting Clinical Data and Patient Confidentiality
AI-assisted diagnostic systems may process medical images, laboratory results, electrocardiographic recordings, clinical notes and other information associated with identifiable patients.
Depending on the technology and service model, these data may be used for clinical inference, technical monitoring, performance evaluation or other purposes.
Healthcare organizations should understand what information is processed, where it is stored, who can access it and whether patient-related data are transferred to external service providers.
These questions are particularly important for systems relying on cloud-based infrastructure or services operated outside the healthcare facility.
Viet Nam's Law on Personal Data Protection No. 91/2025/QH15 and its implementing Decree No. 356/2025/ND-CP took effect on January 1, 2026.
These instruments establish requirements relevant to lawful personal data processing, protection measures, organizational responsibilities and applicable conditions for disclosure and data transfers.
Healthcare organizations should assess the legal basis for processing patient information and the conditions governing its use for AI operation, evaluation and monitoring.
Consent requirements, applicable exceptions, other lawful processing conditions and relevant assessment or documentation obligations should be considered according to the specific circumstances.
The fact that clinical information is valuable for AI validation does not automatically mean it may be reused for unrelated research, model development or commercial purposes.
Removing direct identifiers also does not necessarily eliminate every privacy risk or automatically remove all applicable legal responsibilities.
Where patient-related information is shared with an external provider, appropriate arrangements should address access, retention, confidentiality, security and responsibilities for data processing.
Data integrity is another important consideration.
An AI system may generate a technically plausible output from incorrect, incomplete or mismatched patient information. Such a failure can compromise diagnostic reliability even when the model itself functions as designed.
Protecting patient information and ensuring the integrity of diagnostic data are therefore connected responsibilities within an AI-enabled clinical service.
Building a Practical AI Diagnostic Quality Program
Healthcare organizations do not necessarily need a large AI governance department to begin evaluating an individual diagnostic AI application responsibly.
A practical starting point is to select a defined clinical use case and establish why the technology is needed.
The organization should identify the diagnostic problem, intended users, anticipated benefit and potential consequences of an incorrect or unavailable output.
The available evidence should then be examined in relation to the proposed application. This includes the manufacturer's intended use, relevant regulatory status, validation studies, known limitations and technical requirements.
The evidence should be assessed for its relevance to the hospital's patient population, equipment and clinical workflow.
Where justified, the organization can conduct an appropriately authorized local evaluation using representative cases and a suitable reference standard.
This assessment should consider diagnostic performance, operational reliability and the interaction between healthcare professionals and AI outputs.
Before deployment, the organization should establish the conditions under which the technology may be used, circumstances requiring additional review and situations in which its outputs should not be relied upon.
Responsibilities for oversight also need to be clear.
The relevant clinical department should determine how AI findings will be incorporated into professional assessment. Technical and information technology personnel should address integration, access control and system reliability, while quality and patient safety personnel may support performance monitoring, incident review and improvement activities.
The vendor or provider should supply the information and support required under the applicable contractual and regulatory arrangements.
An appropriate monitoring plan should be established before routine implementation.
This may include diagnostic performance measures, system availability, failed outputs, clinically significant discrepancies and changes in relevant input data or equipment.
Not every measure requires continuous real-time calculation.
For some systems, periodic review of appropriately selected cases may be sufficient. Others may require more frequent assessment because of their clinical consequences, technical characteristics or operational risks.
The monitoring plan should identify who reviews performance information, which findings require further investigation and who has authority to modify or suspend use when necessary.
After implementation, the organization should assess whether the expected benefit has materialized.
A tool introduced to improve reporting efficiency should be evaluated for its effect on actual reporting work and relevant quality outcomes. A system intended to improve recognition of abnormalities should be examined for its contribution to the clinical team's performance, not merely the accuracy of the algorithm.
Feedback from healthcare professionals may reveal problems that are not immediately apparent in technical metrics, including confusing outputs, excessive alerts and difficulties incorporating AI findings into routine practice.
Such feedback should complement objective performance evidence rather than replace it.
A sustainable quality program connects evidence appraisal, local evaluation, implementation safeguards, continuing monitoring and corrective action.
The purpose is not simply to demonstrate that AI software has been installed successfully. It is to establish whether the technology continues to provide dependable support for the clinical tasks it was introduced to perform.
From Promising Technology to Accountable Diagnostic Improvement
AI-assisted diagnostic systems may offer important opportunities to improve healthcare. They can support the recognition of selected abnormalities, assist clinical interpretation and help healthcare professionals manage diagnostic information more efficiently.
However, the presence of AI in a clinical workflow does not, by itself, demonstrate diagnostic excellence.
An algorithm may achieve strong performance in a published study but encounter limitations when exposed to different equipment, patient populations or examination protocols.
A clinician may remain formally responsible for a diagnostic decision without having sufficient opportunity to review an AI recommendation meaningfully.
A model may also perform adequately when first introduced but become less reliable as the surrounding clinical environment changes.
Recent research has made these limitations increasingly visible. It also demonstrates that AI can provide real, though context-dependent, benefits when appropriately integrated into clinical practice.
International guidance from WHO, IMDRF and the scientific reporting community provides useful foundations for evaluating these systems.
Viet Nam's AI regulatory framework adds an important national dimension by emphasizing risk-based management, human oversight, real-world reliability and patient safety.
For healthcare organizations, the priority should be to evaluate the specific technology against its intended clinical purpose, determine whether the evidence applies to local conditions and establish appropriate oversight throughout its operational life.
These activities should be proportionate to clinical risk, supported by credible evidence and consistent with applicable Vietnamese law.
They should also enable organizations to recognize when an AI system is no longer providing the expected benefit or when its use needs to be modified, restricted or discontinued.
Diagnostic excellence is not achieved simply by introducing artificial intelligence into healthcare. It depends on whether the technology is appropriately evaluated, used with meaningful professional oversight and continually assessed for its contribution to accurate, timely and safe diagnostic care.
Selected References and Regulatory Sources
World Health Organization. Regulatory Considerations on Artificial Intelligence for Health. October 19, 2023.
International Medical Device Regulators Forum. Good Machine Learning Practice for Medical Device Development: Guiding Principles. IMDRF/AIML WG/N88 FINAL:2025. January 29, 2025.
International Medical Device Regulators Forum. Technical Framework for Artificial Intelligence Life Cycle Management. N93 Proposed Consultation Document. Public consultation April 10 to July 10, 2026. Not finalized as of November 2026.
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.
Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385. DOI: 10.1136/bmj-2023-078378.
Herpe G, Rheins M, Davy G, Thomas C, Dumas V. Variable performance across radiography vendors of a commercial artificial intelligence system for fracture detection: a multisite evaluation. American Journal of Roentgenology. 2026;226(6). DOI: 10.2214/AJR.26.34550.
Kopanitsa G. Validation is not enough: longitudinal evidence of post-deployment fragility in clinical AI systems. PLOS Digital Health. 2026;5(7). DOI: 10.1371/journal.pdig.0001534.
Kuo R, O'Hanlon C, Freethy A, et al. Performance of artificial intelligence for fracture detection: an umbrella review and second-order meta-analysis. npj Digital Medicine. Published October 27, 2026. DOI: 10.1038/s41746-026-03303-z.
Tan JR, Tran ANT, Sim WY, et al. Radiographer fracture detection with and without AI assistance in prospective emergency department use. npj Digital Medicine. Published October 27, 2026. DOI: 10.1038/s41746-026-03313-x.
National Assembly of Viet Nam. Law on Artificial Intelligence No. 134/2025/QH15. Adopted December 10, 2025. Effective March 1, 2026.
Government of Viet Nam. Decree No. 142/2026/ND-CP, Detailing Certain Provisions and Measures for Implementing the Law on Artificial Intelligence. Issued April 30, 2026. Effective May 1, 2026.
Prime Minister of Viet Nam. Decision No. 33/2026/QD-TTg, Promulgating the List of High-Risk Artificial Intelligence Systems. Issued June 30, 2026. Effective August 15, 2026.
Government of Viet Nam. Decree No. 98/2021/ND-CP on Medical Device Management, as amended and reflected in Consolidated Document No. 08/VBHN-BYT dated March 6, 2026.
National Assembly of Viet Nam. Law on Medical Examination and Treatment No. 15/2023/QH15, as amended and reflected in Consolidated Document No. 26/VBHN-VPQH dated February 26, 2026.
National Assembly of Viet Nam. Law on Personal Data Protection No. 91/2025/QH15. Effective January 1, 2026.
Government of Viet Nam. Decree No. 356/2025/ND-CP, Detailing Certain Provisions and Measures for Implementing the Law on Personal Data Protection. Effective January 1, 2026.
Professional Scope and Disclaimer
This article is published by Diagnostic Excellence Vietnam (DxEx.vn), a professional initiative dedicated to advancing diagnostic quality, safety and reliability in Viet Nam through evidence-informed knowledge sharing, risk management, measurement and quality improvement.
The content is intended for general professional education and discussion. It does not constitute patient-specific medical advice, a clinical or laboratory diagnosis, an individualized recommendation to select, validate, deploy or use a particular AI-assisted diagnostic system, a regulatory conformity assessment, medical device authorization, an accreditation assessment, a determination of professional or technical error, or legal advice.
References to international guidelines, technical standards, reporting frameworks and research findings do not imply that they are automatically mandatory under Vietnamese law. Their applicability depends on the relevant legal requirements, intended purpose, system classification, healthcare setting and professional circumstances.
Decisions concerning AI system procurement, classification, validation, clinical implementation, human oversight, performance monitoring, incident management, result interpretation and patient care must be undertaken by appropriately qualified and authorized professionals in accordance with applicable Vietnamese law, relevant professional standards, authorized service capabilities and the patient's clinical circumstances.
Diagnostic Excellence Vietnam does not develop, approve, certify or clinically validate AI systems through publication of this educational content. It does not provide clinical diagnosis or treatment services.
