Does Clinical AI Actually Improve Care in Viet Nam? A Practical Framework for Evaluating Patient Outcomes, Workflow, Equity, Costs and Unintended Consequences
PRACTICAL FRAMEWORKS & TOOLSHEALTHCARE IMPLEMENTATION, MONITORING & EVALUATIONEVIDENCE & IMPLEMENTATION READINESS
Does Clinical AI Actually Improve Care in Viet Nam? A Practical Framework for Evaluating Patient Outcomes, Workflow, Equity, Costs and Unintended Consequences
PRACTICAL FRAMEWORK & TOOL
Last reviewed: 4 October 2026
A clinical AI system can be accurate, well calibrated, locally validated and technically stable, and still fail to improve care.
An AI alert may identify high-risk patients correctly, but clinicians may receive it too late to change treatment. An imaging system may shorten interpretation time without shortening time to diagnosis because another part of the pathway remains the bottleneck. A documentation assistant may reduce typing while increasing the time required to review and correct notes. A screening algorithm may detect more disease while generating enough false positives to increase unnecessary imaging, procedures or specialist referrals.
None of these examples necessarily means that the AI model is poor. They illustrate a different problem: model performance and healthcare impact are not the same thing.
This distinction is becoming increasingly important as clinical AI moves from pilots into routine care. A five-phase medical-AI evaluation framework published in 2026 argues that evaluation needs to progress beyond technical validation toward operational robustness, human–AI interaction, clinical evidence and real-world system integration. Its final phase focuses not simply on whether the model continues to perform, but on sustained effectiveness, safety, fairness and organizational impact in routine care [7].
A separate 2026 conceptual outcome-measurement framework makes the point even more directly: discrimination, calibration, sensitivity and specificity are necessary but insufficient for determining whether deployed AI improves patient-relevant outcomes, workflow, resource use, economic performance, clinician and patient experience, organizational learning or equity [8].
For healthcare organizations in Viet Nam, the question after implementation should therefore move beyond “Is the AI still working?” The more important question is:
“Is care actually better because we use it, and what else changed at the same time?”
The AI model is only one part of the causal pathway
A useful way to evaluate clinical AI is to follow the chain between the algorithm and the eventual healthcare outcome.
The AI first needs to encounter the right patient or case. Its input data need to be sufficiently accurate and timely. The model needs to generate an appropriate output. That output needs to reach the right person at the right time. The clinician or team needs to understand it, trust it appropriately and have the authority and resources to act. The action then needs to change a clinical process in a way capable of influencing an outcome.
The pathway can therefore be represented as:
eligible patient → AI exposure → AI output → human interpretation → clinical action → workflow change → patient outcome → health-system impact
A failure anywhere along this chain can weaken or break the relationship between good model performance and better care. Suppose a deterioration model identifies 90% of patients who later become critically ill. That sounds impressive. But if only half of eligible patients are processed by the system, clinicians review only 60% of the alerts, and treatment changes after only one third of reviewed alerts, the effect on patient outcomes may be far smaller than the model's sensitivity suggests.
Impact evaluation should therefore measure reach, exposure, uptake and action, not simply outcomes at the end of the pathway. If mortality, length of stay or complications do not improve, the organization needs to understand whether the model failed, whether eligible patients never reached the AI, whether clinicians did not use the output, whether the resulting intervention was ineffective, or whether another bottleneck prevented benefit.
A 2026 framework for AI implementation research similarly emphasizes that effectiveness depends on implementation context, organizational resources, workflow integration, user-centeredness, quality and safety, and equity rather than model performance alone [11].
Start by defining what “better care” is supposed to mean
Healthcare organizations should define the intended benefit before implementation rather than deciding afterward which available metric makes the project appear successful. A clinical AI system should have a plausible impact hypothesis.
If AI is introduced to identify sepsis earlier, the intended benefit might be shorter time to appropriate treatment and ultimately fewer complications or deaths. If AI prioritizes radiology worklists, the intended benefit might be shorter time to interpretation for clinically urgent studies, not simply faster average reporting across every case. If an ambient documentation system is introduced to reduce clinician burden, relevant outcomes may include documentation time, after-hours work, note quality, correction burden and clinician experience rather than transcription accuracy alone.
The selected outcomes should therefore correspond to the reason the organization adopted the AI.
This sounds obvious, but clinical-AI evaluations frequently remain centered on technical or intermediate metrics. A 2026 systematic evidence map of real-world and clinical studies involving large language models found that the evidence remains heterogeneous, with many evaluations emphasizing functional performance and operational measures while robust evidence on clinical outcomes remains comparatively limited [10].
The right evaluation question is not simply “What can we measure?” It is:
“What would have to improve for us to conclude that this AI was worth introducing into care?”
Patient outcomes should remain visible even when the AI acts upstream
Not every AI tool will directly reduce mortality, complications or readmissions. Requiring every administrative or documentation tool to demonstrate a mortality benefit would make little sense. But clinically consequential AI should maintain a visible connection to patient outcomes, even when its immediate function occurs earlier in the care pathway.
Outcome measures may include mortality, morbidity, complications, diagnostic delay, readmission, emergency escalation, treatment success, functional status, symptom burden, quality of life or other patient-relevant endpoints appropriate to the use case. Patient-reported outcomes and patient experience may also matter when the AI changes communication, access or care delivery.
Clinical benefit is possible, but it is highly context dependent. A 2026 systematic review of 32 randomized trials of AI interventions in cardiovascular care reported pooled improvements across workflow, patient engagement and selected clinical outcomes in the included trials. The authors also found important heterogeneity, design limitations and a translation gap: improved diagnostic support did not automatically translate into better hard outcomes unless it was linked to an actionable clinical intervention [9].
Better detection is not the same as better care.
Detection becomes valuable when it leads to an appropriate action early enough to change what happens to the patient.
Workflow outcomes matter because AI often changes the care system around the decision
Clinical AI often changes workflow, sometimes in ways that are less visible than its effect on the immediate clinical task. It may change who reviews a case first, how work is prioritized, when information reaches a clinician, which cases require escalation, how documentation is produced, how often another professional is consulted or how long a patient waits for the next step in care.
Workflow evaluation should therefore examine more than average time saved. It should ask whether total work decreases or increases, where work moves after implementation, whether one professional group saves time while another absorbs additional review, whether urgent cases receive faster attention, whether false positives create new work, whether clinicians repeatedly verify AI outputs, and whether communication or handoffs change.
For example, an AI scribe may save ten minutes of drafting but require eight minutes of careful correction. A radiology AI system may reduce interpretation time but increase downstream specialist consultations. An alert may bring treatment forward by 30 minutes but generate enough false alarms that staff gradually stop responding.
Human–AI research reinforces the importance of evaluating the combined workflow. A 2026 systematic review and meta-analysis of human–LLM collaboration found that benefits varied across tasks and studies rather than being universal [14]. A broader 2026 scoping review found that successful human–AI collaboration depends on task fit, workflow integration, training and appropriately calibrated trust rather than model accuracy alone [15].
A useful operational question is:
Where did the work go, not just how many minutes did the AI appear to save?
Measure reach and implementation fidelity before blaming or crediting the AI
An organization cannot interpret impact correctly without knowing how the system was actually used.
A clinical-AI tool may technically be available for every eligible patient but reach only a fraction. Some clinicians may rely on it heavily while others ignore it. One hospital site may integrate it into routine workflow while another treats it as optional. Staff may initially use the system as intended and later develop workarounds.
Impact evaluation therefore needs to capture both reach and implementation fidelity. How many cases were eligible? How many actually reached the AI? How many generated an output? How many outputs reached the intended user? How many were reviewed? How often were recommendations accepted, modified, rejected or ignored? How often did an AI output produce a measurable clinical action?
These measures help distinguish intervention failure from implementation failure.
If the AI was technically available but rarely reviewed or acted upon, absence of patient benefit does not necessarily mean that the algorithm lacked potential value. Conversely, if reach and uptake were high and outcomes still did not improve, the organization should be cautious about attributing value to the system simply because clinicians used it.
This distinction becomes especially important when scaling a successful pilot. The people, training, leadership attention and workflow discipline present during a pilot may not persist when the system expands across departments or facilities.
Improvement should be measured against a credible counterfactual
If outcomes improve after AI is introduced, it is tempting to credit the AI. Hospitals, however, change continuously. Staffing changes. New protocols are introduced. Teams receive training. New medicines become available. Seasonal disease patterns change. Other quality-improvement initiatives may begin at the same time.
Impact evaluation therefore needs a credible answer to the question:
What would probably have happened without the AI?
The most appropriate design depends on the use case, the decision being made and the strength of the causal claim the organization needs to support.
Randomized or cluster-randomized designs can provide strong causal evidence when feasible. Pragmatic trials can evaluate effectiveness in ordinary workflows. Stepped-wedge designs may be useful when an intervention is introduced sequentially across sites. Controlled before-and-after studies, prospective cohorts and interrupted time-series designs can also provide useful evidence when randomization is impractical.
Simple pre–post comparison may still be operationally useful, but its limitations should be understood. If length of stay falls from 5.2 to 4.7 days after AI deployment, that change does not by itself prove that AI caused the improvement.
The evaluation design should be proportionate to the claim being made. A healthcare organization does not necessarily need a randomized trial to determine whether an AI documentation tool reduced after-hours clerical work. A system expected to influence treatment decisions in ways that may affect serious patient outcomes deserves a stronger evidence strategy.
The five-phase evaluation framework published in 2026 similarly distinguishes technical testing, controlled interaction studies, clinical evidence generation and real-world effectiveness evaluation rather than assuming that one study design is sufficient for every stage [7].
Equity means evaluating who receives the benefit, not only whether model performance is fair
AI fairness is often evaluated by comparing model performance across demographic groups. That is important, but equitable healthcare impact is broader than model fairness.
Two groups can have similar model sensitivity while receiving very different benefits from the system. One site may have specialists available to respond immediately to AI alerts while another does not. Patients in one pathway may reliably receive the additional diagnostic test triggered by AI while others face access barriers. A digital tool may perform accurately among people who use it while being less accessible to patients affected by language, digital-literacy or technology-access barriers.
Impact evaluation should therefore ask at least two distinct questions:
Does the AI perform appropriately across relevant groups?
Are the benefits and burdens of the AI distributed equitably?
Although focused on the WHO European Region, a 2026 WHO review illustrates a broader implementation challenge: equity is increasingly recognized in digital-health policy and evaluation but remains inconsistently incorporated into practice, while standardized equity metrics and robust evaluation remain limited [13]. A 2026 landscape analysis of 5,452 digital-health randomized-trial outcomes also found uneven reporting of participant characteristics, making it difficult to determine whether apparent benefits apply across populations [12].
For healthcare organizations in Viet Nam, relevant factors depend on the use case. Site, age, sex, clinical service, disease severity, referral source, language, access to follow-up or other characteristics may matter. Subgroup analyses should be clinically justified and interpreted carefully, particularly when sample sizes are small.
The objective is not to produce dozens of unstable subgroup comparisons. It is to determine whether AI could systematically make care better for some patients while leaving others behind or shifting burdens toward them.
Costs should include what the organization spends because the AI exists
AI products are frequently justified through efficiency or cost-saving claims. Those claims deserve careful evaluation because the cost of an AI system is not simply the annual license fee.
Total organizational cost can include procurement, integration, cloud and API charges, hardware, cybersecurity, local validation, staff training, workflow redesign, data engineering, monitoring, quality assurance, incident investigation, vendor management, recalibration, maintenance and eventual decommissioning.
AI may also create downstream costs. Higher detection can lead to additional testing, biopsies, consultations or procedures. False positives consume capacity. Human review of AI outputs has a cost. If staff must verify every generated note, apparent automation may simply move work rather than remove it.
Benefits may arise through shorter stays, fewer adverse events, earlier diagnosis, reduced duplication, improved productivity or avoided unnecessary procedures.
A 2025 systematic review of economic evaluations of clinical AI found generally promising evidence across several specialties, including studies reporting favorable cost-effectiveness and cost savings. It also identified important methodological limitations, including incomplete consideration of implementation and infrastructure costs and limited attention to equity in some evaluations [16].
Healthcare organizations should also distinguish cost-effectiveness from budget impact.
An intervention may provide good value per health outcome achieved while still requiring more annual expenditure than the organization can realistically absorb. Conversely, a tool may reduce one departmental cost without improving patient outcomes enough to justify broader adoption.
For Viet Nam, economic evaluation should use local costs, staffing patterns, clinical pathways and financing arrangements rather than importing economic conclusions from another health system.
Patient and clinician experience should be treated as impact outcomes
AI can improve measurable efficiency while making care worse to experience. A documentation assistant may reduce keyboard time but distract clinicians during conversations. An AI triage tool may improve prioritization while making patients feel that access to a person has become more difficult. A decision-support system may improve diagnostic accuracy while increasing anxiety among clinicians who feel accountable for recommendations they cannot adequately interrogate.
Patient and clinician experience should therefore be considered meaningful impact domains when the use case can affect them.
For clinicians, relevant measures may include cognitive workload, usability, alert burden, trust, perceived autonomy, after-hours work and time available for direct patient care. For patients, relevant outcomes may include communication quality, understanding, involvement, access, waiting time, continuity, privacy concerns and trust.
These measures should not replace safety or clinical outcomes. They help explain why an otherwise technically effective intervention succeeds, fails or becomes unsustainable in routine care.
Improvement should be assessed alongside balancing measures
Quality improvement has long recognized that changing one part of a healthcare system can create problems elsewhere. Clinical AI is no exception.
If an AI system is intended to shorten imaging turnaround time, balancing measures might examine false-priority cases or workload displacement. If the objective is increased cancer detection, balancing measures may include additional biopsies, false-positive referrals and patient anxiety. If the goal is reducing documentation burden, balancing measures should examine correction time, note errors and downstream clarification requests.
This matters because an AI system can optimize exactly what it was designed to optimize while creating harm outside that metric. A system designed to identify more cases may increase unnecessary intervention. A scheduling model optimizing utilization may worsen continuity of care. A documentation model optimized for completeness may produce excessively long notes that obscure clinically important information.
Impact evaluation should therefore ask two questions at the same time:
Did the intended outcome improve?
What might have become worse as a consequence?
The second question is often where unintended consequences first become visible.
Unintended consequences should be actively searched for
Healthcare organizations should not wait for unintended consequences to become serious incidents.
Potential effects include false reassurance when AI does not alert; automation bias; diagnostic anchoring; alert fatigue; deskilling; new documentation errors; increased testing; shifts in workload; changed referral patterns; inequitable access; new privacy risks; delayed decisions during AI downtime; or clinicians changing their behavior in ways that alter the original validation assumptions.
Some effects can be paradoxical. If clinicians intensify treatment for patients identified as high risk, those patients may subsequently have better outcomes. Over time, the original relationship between the risk score and outcome can change precisely because the organization acted on the score.
A 2026 implementation-research framework identifies unintended harms and technical errors among issues relevant to healthcare AI implementation and emphasizes quality monitoring together with mechanisms for frontline feedback and complaints [11].
Frontline staff are therefore an important detection system. Nurses, physicians, pharmacists, technicians and administrative teams may notice workflow consequences before those effects appear in formal performance dashboards.
Impact evaluation should combine quantitative measurement with a structured way to capture unexpected signals from real work.
Evaluate the whole human–AI intervention, not the algorithm in isolation
Once clinical AI enters routine care, the meaningful intervention is often no longer simply “the AI.” It is:
AI + clinician + workflow + data + policy + training + escalation pathway + surrounding health system
A model may be identical at two hospitals while producing different results because the surrounding intervention is different. This is why replication matters: a positive study at one center should not automatically be treated as proof that another healthcare organization will obtain the same outcomes.
A 2026 quantitative analysis of 4,667 primary medical-AI studies found that only 113, or 2.4%, were randomized controlled trials. Among the RCTs with non-favorable findings, authors reported a range of possible explanations, including study-design or endpoint issues, human–computer interaction and workflow integration, model limitations, statistical power and implementation-related factors [17].
These findings do not prove that workflow or implementation caused every unfavorable result. They do show how difficult it is to interpret clinical AI independently of its study design and implementation context.
Healthcare organizations should therefore avoid evaluating AI as though it were an intervention whose effect remains identical regardless of the environment into which it is introduced. Clinical AI is a sociotechnical intervention: its effectiveness emerges from the interaction between technology, people, workflow and the organization around it.
Impact can change after the organization learns to use the AI
The first month after go-live may not represent steady-state impact. Clinicians learn. Trust changes. Workflow adaptations emerge. Alert response becomes faster or slower. Teams develop shortcuts. Vendors update products. Patient populations change. Some benefits may increase as users become familiar with the system, while other problems may emerge only after novelty and intensive launch support disappear.
Impact evaluation should therefore consider time. Organizations may need an early implementation period, a stabilization period and longer-term evaluation rather than one post-launch snapshot.
The appropriate horizon depends on the intervention. A workflow tool may show meaningful changes within weeks. Readmission or mortality outcomes may require months. Changes in clinician skill, resource utilization or unintended downstream effects may take longer.
The five-phase evaluation framework treats real-world integration as a longitudinal stage in which effectiveness, fairness, safety, adoption and organizational assimilation require continued assessment rather than a one-time final evaluation [7].
AI value is not necessarily a fixed property at the moment of go-live.
Viet Nam's AI framework supports lifecycle governance but does not prescribe one universal impact study
Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 has been effective since 1 March 2026, and Decree No. 142/2026/NĐ-CP since 1 May 2026 [1,2]. The legal framework is risk-based rather than imposing identical obligations on every AI system.
For high-risk AI systems, Article 15 of Decree No. 142 establishes risk-management responsibilities for providers and distinct operational responsibilities for deployers. Providers must establish and maintain a risk-management system appropriate to the system's intended purpose, deployment scope and level of risk. Deployers must operate the system within the purpose, conditions and limitations defined by the provider; monitor for deviations, risks and incidents; maintain mechanisms for human oversight and intervention; apply risk controls within their authority; and coordinate with the provider when significant changes or incidents require risk-management measures to be reviewed or updated [2].
Article 13 separately requires providers of high-risk AI systems to reassess conformity where a significant change affects the basis of the original conformity assessment [2].
These provisions make continued safety and performance governance important, but they should not be overstated. Vietnamese AI legislation does not currently require every healthcare organization to conduct a randomized trial, use one standardized patient-outcome set, demonstrate a specified return on investment or follow the exact impact-evaluation framework proposed in this article.
Decision No. 33/2026/QĐ-TTg, effective from 15 August 2026, establishes the current formal list of high-risk AI systems [3]. As of this review, its healthcare section identifies specified AI-assisted surgical and robotic-surgery systems under defined conditions. Not every clinical decision-support, prediction or generative-AI application becomes legally high-risk merely because it is used in healthcare.
Healthcare organizations can nevertheless evaluate clinical impact much more broadly as part of responsible quality management and patient safety. Legal minimum and good healthcare governance are not always the same threshold.
Outcome evaluation must also respect personal-data and healthcare-information requirements
Real-world impact evaluation often requires linking AI exposure with clinical outcomes, workflow information, user behavior and patient characteristics. Those datasets can contain personal and sensitive health information.
Viet Nam's Law on Personal Data Protection No. 91/2025/QH15 has been effective since 1 January 2026 [4]. Healthcare organizations evaluating AI impact therefore need to ensure that collection, linkage, analysis, retention, sharing and other processing of personal data comply with the applicable personal-data framework, including Decree No. 356/2025/NĐ-CP where relevant [5].
Impact evaluation should not become an excuse to collect every available variable indefinitely. The organization should define the information required to answer its evaluation question and apply appropriate purpose, access, security, retention and governance controls.
Circular No. 13/2025/TT-BYT has governed implementation of electronic medical records since 21 July 2025 and remains in force [6]. EMR data can become an important source for real-world evaluation and, when appropriately generated, linked and analyzed to answer a defined question, can contribute to real-world evidence.
The distinction matters. Real-world data do not automatically become real-world evidence merely because they exist in the EMR, and availability of data does not remove the obligation to govern how those data are used.
A practical Clinical AI Care Impact Framework
A healthcare organization does not need to measure every possible outcome for every AI system. Evaluation should be proportionate to intended use, clinical risk, organizational investment and the strength of the claim being made.
A practical framework can begin with the intended benefit and causal pathway. What healthcare problem was the AI introduced to address? How is the AI expected to change clinician or patient behavior? Which step connects the AI output to the eventual outcome? What could prevent that pathway from working?
The second domain is patient outcomes and safety. Which clinical outcomes are plausibly affected by the AI? Are there measures of morbidity, mortality, diagnostic delay, treatment delay, complications, readmission, functional status, quality of life or other patient-relevant outcomes appropriate to the use case? What safety measures need to remain stable even if the primary target improves?
The third domain is workflow, reach and implementation fidelity. How many eligible cases actually receive the AI intervention? Are outputs delivered, reviewed and acted upon? Does time to decision change? Where does workload move? Does AI reduce total effort or simply transfer work between professionals? Are clinicians using the system in the manner intended?
The fourth domain is equity and access. Are performance, reach, use, outcomes and access to follow-up reasonably consistent across relevant patient groups, sites and services? Does the intervention depend on resources that are less available in some parts of the organization? Are particular patients more likely to be excluded from the AI-enabled pathway?
The fifth domain is cost and resource use. What is the full cost of acquiring, integrating, validating, operating, monitoring and maintaining the AI? Does it create downstream testing, consultation or review? Does it save staff time, bed days, procedures or other resources? Is the intervention affordable at scale as well as potentially cost-effective?
The sixth domain is experience and human factors. How does the system affect clinician workload, trust, cognitive burden and professional practice? Does it improve or impair communication with patients? Does human review remain meaningful? Are override and escalation patterns revealing problems in workflow design?
The seventh domain is unintended consequences and balancing measures. What could plausibly become worse if the AI succeeds at its primary objective? Are false positives, false reassurance, over-testing, alert fatigue, privacy risk, deskilling, workload transfer and new forms of delay being actively monitored?
Finally, the organization needs attribution, duration and decision rules. What comparator will be used? How long will evaluation continue? What other changes occurred during the same period? What level of evidence is needed to continue, modify, scale, restrict or stop the AI? Who has authority to make that decision?
The Clinical AI Care Impact Framework is a practical organizational framework proposed for evaluation, healthcare quality and patient-safety purposes. It is not a statutory checklist prescribed by Vietnamese law.
The framework should end with a decision, not a dashboard
Measurement alone does not improve care. An organization may collect dozens of AI metrics without deciding what should happen when the findings are disappointing.
Impact evaluation therefore needs decision rules.
If model performance remains good but the intended patient outcome does not improve, should the workflow be redesigned? If benefit appears only at one site, should scale-up pause until the difference is understood? If time is saved but error correction increases, is the intervention still worthwhile? If clinical outcomes improve but one group receives less benefit, what needs to change? If the AI appears clinically effective but is too costly to sustain, is there another implementation model?
Possible decisions include continue, modify, retrain, recalibrate, redesign workflow, restrict use, conduct further evaluation, expand or decommission.
The purpose of evaluation is not to prove that the organization's original procurement decision was correct. It is to determine what the organization should do next.
This is where impact evaluation becomes part of quality improvement rather than a retrospective academic exercise.
What this means for healthcare organizations in Viet Nam
Clinical AI is often introduced with impressive evidence about sensitivity, specificity, AUC, time savings or benchmark performance. Those metrics can be important, but none answers the final healthcare question:
Did patients receive better care?
Sometimes the answer will be yes. Contemporary randomized-trial evidence shows that appropriately designed and integrated AI interventions can improve workflow and, in some settings, meaningful clinical outcomes [9].
Sometimes AI will improve a process but not the patient outcome. Sometimes benefit will depend heavily on human–AI interaction. Sometimes an apparent efficiency gain will simply transfer work. Sometimes cost savings will disappear once integration and monitoring are included. Sometimes one patient group will benefit more than another. And sometimes an AI system that performs exactly as designed will reveal that the underlying clinical pathway, rather than the algorithm, is where improvement is most needed.
That is not a failure of evaluation. It is the reason evaluation matters.
For Viet Nam, this perspective is particularly important as healthcare organizations expand digital records and integrate more AI into routine workflows. The objective should not be to demonstrate that every AI investment succeeds. It should be to build organizations capable of knowing which AI improves care, for whom, under what conditions, at what cost and with what unintended effects.
The leadership question should therefore move beyond “Is the AI performing well?”
A more useful question is:
“What became better for patients and the healthcare system because we introduced it — what became worse, and what evidence supports that conclusion?”
That is where clinical AI evaluation becomes healthcare quality improvement.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, statistical, health-economic, data-protection or medical-device advice.
The terms clinical impact, reach, implementation fidelity, balancing measures, real-world data, real-world evidence, unintended consequences and the Clinical AI Care Impact Framework are used here as practical healthcare evaluation, quality-improvement and patient-safety concepts. The framework proposed in this article is not a statutory evaluation checklist under Vietnamese law.
Article 15 of Decree No. 142/2026/NĐ-CP establishes risk-management requirements for providers and operational risk-management requirements for deployers of high-risk AI systems, while Article 13 addresses conformity reassessment by providers following specified significant changes. These provisions should not be interpreted as imposing one clinical-outcome set, economic threshold, study design or causal-evaluation method on every healthcare AI system.
Decision No. 33/2026/QĐ-TTg establishes the current formal list of high-risk AI systems and should not be interpreted as classifying every healthcare AI application as high-risk. Healthcare organizations may nevertheless use broader impact evaluation as part of responsible implementation, healthcare quality management and patient-safety governance.
The international frameworks and studies cited below are evidence and implementation references. Their methods, thresholds and findings should be adapted to the specific AI system, Vietnamese healthcare setting, population, workflow and available resources.
References
[1] National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
[2] 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.
[3] Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. Issued 30 June 2026; effective 15 August 2026.
[4] National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
[5] Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Personal Data Protection. Issued 31 December 2025; effective 1 January 2026.
[6] Ministry of Health of Viet Nam. Circular No. 13/2025/TT-BYT Providing Guidance on Implementation of Electronic Medical Records. Issued 6 June 2025; effective 21 July 2025.
[7] Ye Z, Chen Y, Huang X, et al. A five-phase evaluation framework for diagnostic and predictive medical artificial intelligence. npj Digital Medicine. 2026;9:678.
[8] Kirchhoff J, Berns F, Schieder C, Schobel J. A testable framework linking diagnostic AI contribution to outcome measurement in clinical decision support. Frontiers in Digital Health. 2026;8:1871438.
[9] Lin YE, Yang SM, Huang CJ, et al. Impact of artificial intelligence on cardiovascular workflow, engagement, and outcomes: a systematic review. npj Digital Medicine. 2026;9:536.
[10] He Z, Yang L, Liang Z, et al. Clinical outcomes and reporting quality of large language model interventions in practice: a systematic evidence map. npj Digital Medicine. 2026;9:672.
[11] Powis M, Ladak AM, Lakey A, et al. Framework for artificial intelligence implementation research in healthcare: synthesizing current evidence on barriers and facilitators. npj Digital Medicine. 2026;9:543.
[12] Harrison TB, Hu D, Chen F, et al. A landscape analysis of access, equity, and reporting trends in digital health randomized controlled trials. npj Digital Medicine. Published 6 August 2026.
[13] World Health Organization Regional Office for Europe. Equity across the regulation, implementation and evaluation of digital health: scoping review. WHO Regional Office for Europe; 2026.
[14] Wang G, Zhang K, Jiang J, et al. Human–large language model collaboration in clinical medicine: a systematic review and meta-analysis. npj Digital Medicine. 2026;9:195.
[15] Strong J, Rogers H, Sun E, et al. Human-AI Collaboration in Healthcare: A Scoping Review. npj Digital Medicine. Published 20 June 2026.
[16] El Arab RA, Al Moosa OA. Systematic review of cost effectiveness and budget impact of artificial intelligence in healthcare. npj Digital Medicine. 2025;8:548.
[17] Wang Q, Yang Y, Qiu L, et al. A quantitative analysis of global AI medical studies: gaps in randomized controlled trials. npj Digital Medicine. 2026;9:403.