Bias, Fairness and Equity in Clinical AI in Viet Nam: How Healthcare Organizations Should Assess and Manage Performance Differences Across Patient Groups
IMPLEMENTATION INSIGHTS
Bias, Fairness and Equity in Clinical AI in Viet Nam: How Healthcare Organizations Should Assess and Manage Performance Differences Across Patient Groups
IMPLEMENTATION INSIGHTS
Last reviewed: 2 October 2026
A clinical AI system can perform well overall and still perform differently across groups of patients. That difference may be clinically unimportant. It may reflect real differences in disease prevalence, clinical presentation or measurement conditions. It may be an unstable finding caused by a small sample. Or it may reveal a genuine problem: some patients may be more likely to have a condition missed, receive an unnecessary alert, be incorrectly classified or gain less benefit from the technology than others.
Looking only at average performance can hide these patterns. For healthcare organizations, the practical question is therefore not simply “Is this AI accurate?” It is “Does it perform acceptably across the patients and conditions in which we intend to use it, and what will we do if important differences appear?”
This question is increasingly relevant in Viet Nam. Article 4 of the Law on Artificial Intelligence No. 134/2025/QH15, effective since 1 March 2026, establishes principles including fairness, transparency, avoidance of bias, non-discrimination, prevention of harm and accountability for AI decisions and consequences.[1] The National Artificial Intelligence Ethics Framework, issued under Circular No. 05/2026/TT-BKHCN in March 2026, provides more operational direction by calling for identification and mitigation of data bias, model bias and operational bias, while requiring appropriate consideration of impacts on vulnerable groups including children, older people, people with disabilities and other disadvantaged groups.[2]
The scope of that Circular needs to be understood correctly. It applies to state agencies and to organizations or individuals acting as AI developers, providers, deployers or users for state-management activities or public-service delivery. Other Vietnamese and foreign organizations carrying out AI activities in Viet Nam are encouraged to apply it.[2] It should therefore not be presented as if every provision has exactly the same mandatory legal status for every private healthcare organization.
In this article, equity is used as a health-system and implementation concept. Viet Nam’s AI Law establishes principles of fairness, avoidance of bias and non-discrimination, but it does not prescribe a single clinical equity metric or statutory fairness test. The practical governance challenge is to identify performance differences that may matter, understand their causes and determine whether they create avoidable or unjustified disadvantage.
Performance differences, bias, fairness and equity are related but not identical
These concepts are connected, but they are not interchangeable. A performance difference is an observed difference in how an AI system performs between patient groups, sites or operating conditions. For example, sensitivity may be higher in one subgroup than another. Bias describes a systematic distortion that may arise from data, model development, measurement, workflow, implementation or human interaction; a biased training dataset is only one possible source.
Fairness asks whether the distribution of benefits, errors and risks is acceptable and justified in context. It has both technical and normative dimensions, and there is no single mathematical definition that is correct for every clinical use. Equity is broader still. IMDRF describes health equity as the absence of unfair and avoidable or remediable differences in health among population groups and encourages attention to differential impacts, disaggregated data and underrepresented populations throughout the medical-device lifecycle.[5]
An observed performance difference therefore deserves investigation, but it should not automatically be labeled unfair, biased or discriminatory. The organization first needs to understand why the difference exists, how certain the finding is and whether it creates a clinically meaningful disadvantage.
Equal performance does not mean identical numbers for every group
Clinical fairness does not require every metric to be identical across all patient groups. Disease prevalence, presentation, referral patterns, severity and measurement quality can differ, and those differences may affect performance measures even when the model itself behaves consistently.
Positive predictive value provides a simple example. Two groups can have the same sensitivity and specificity but different positive predictive values when disease prevalence differs. That alone does not demonstrate bias. There are also trade-offs between metrics: increasing sensitivity may increase false positives, equalizing one error rate may worsen another, and a model can be well calibrated while still producing different error rates in populations with different underlying risk.
There is therefore no universal fairness score that determines whether clinical AI is equitable. The more useful question is whether observed differences are clinically justified, sufficiently understood and appropriately managed, rather than whether every number is identical.
Start with the intended population and plausible pathways to unequal performance
Subgroup assessment should begin before running statistical tests. Healthcare organizations should first define who the system is intended to serve and where meaningful differences could plausibly arise.
Depending on the technology, relevant characteristics may include age, sex, disease severity, important comorbidities, inpatient versus outpatient setting, emergency versus elective care, device type, image quality, hospital site, operator, language or documentation pattern. Other characteristics such as ethnicity, geography, disability or socioeconomic circumstances may matter in particular use cases where assessment is clinically relevant, feasible and appropriately governed.
The point is not to create the longest possible demographic checklist. A subgroup belongs in the evaluation because there is a clinical, technical, evidentiary or equity reason to examine it.
For an imaging system, equipment manufacturer, acquisition protocol or image quality may matter as much as patient demographics. For Vietnamese-language clinical NLP, terminology, abbreviations and mixed Vietnamese–English documentation may be important. For remote monitoring, internet access, device availability and digital literacy may influence who benefits even if the underlying algorithm performs well.
Fairness assessment should therefore follow the actual pathway through which harm or reduced benefit could occur.
External evidence should show who was represented
Before generating local subgroup analyses, healthcare organizations should first understand what the existing evidence already shows. Where was the system developed? Which populations were represented? Were relevant groups sufficiently included in training and evaluation? Were subgroup results reported? Were some groups poorly represented or excluded? Were devices, data and workflows similar to those used locally?
IMDRF’s 2025 Good Machine Learning Practice principles emphasize that datasets used in clinical evaluation should sufficiently represent the intended patient population, use environment and measurement inputs, using samples of adequate size so that findings can reasonably generalize to the intended population. The principles also specifically call for identifying circumstances and subgroups in which a model may underperform and considering the overall population, relevant subgroups, uncertainty and variability.[6]
This is particularly relevant when AI developed outside Viet Nam is introduced into Vietnamese healthcare. Development elsewhere does not mean a system will perform poorly here. Equally, strong international performance should not automatically be treated as proof of equivalent performance in Vietnamese patients.
The better question is: How similar are the people, data, devices and clinical conditions represented in the existing evidence to the environment in which we will actually use the system?
Do not import subgroup frameworks mechanically from other countries
Much international AI fairness literature focuses heavily on race and ethnicity. Those dimensions can be important, but healthcare organizations in Viet Nam should not simply import a U.S. or European subgroup framework and assume it captures the most relevant local risks.
The groups worth examining should follow the technology, use case and Vietnamese clinical context. Age may matter for one system and sex for another. Equipment type, hospital site or disease severity may matter more for an imaging algorithm. Language quality may be central for clinical NLP or conversational AI. Connectivity and access may matter more for remote monitoring.
Ethnicity may also be clinically or operationally relevant in some situations, but its use should be scientifically justified and appropriately governed. IMDRF N79 supports this contextual approach: disaggregated data can help identify differential performance across relevant groups, but meaningful analysis depends on collecting sufficient information rather than treating every possible subgroup category as universally applicable.[5]
Subgroup analysis needs enough data to be meaningful
One of the easiest ways to generate misleading fairness conclusions is to compare groups with too few cases. A dataset can be large overall while containing very few people in one clinically important subgroup. A small number of errors can then create an apparently dramatic difference even though the estimate is highly uncertain.
Healthcare organizations should therefore look beyond point estimates. Sample size, numbers of positive and negative cases, uncertainty and confidence intervals can be as important as the headline percentage. IMDRF’s health-equity principles similarly link meaningful disaggregated analysis to having enough information across the relevant categories.[5]
Where evidence is insufficient, the responsible conclusion may simply be “Performance in this group remains uncertain.” That is more useful than claiming equivalence without evidence or labeling a model biased on the basis of an unstable estimate.
Subgroup questions should also be planned where possible. Testing many small groups only after seeing the results increases the likelihood of apparent differences arising by chance. Exploratory analysis still has value, but unexpected findings should generally be treated as signals for further investigation rather than immediate proof of inequity. The objective is not to produce the largest possible number of subgroup comparisons; it is to identify clinically important differences that could affect safe or equitable care.
Choose metrics according to the clinical harm that matters
The relevant assessment depends on what the AI is intended to do. For a screening system, differences in sensitivity or false-negative rates may be particularly important because missed disease is the principal concern. For an alerting system, false-positive rates may matter because excessive alerts can lead to unnecessary investigation or alert fatigue. For risk prediction, calibration across clinically relevant groups may be important because the same predicted probability should retain a meaningful relationship with observed risk.
For generative AI, conventional diagnostic metrics may be less useful. Organizations may need to examine whether clinically important factual errors, omissions, unsupported statements or harmful recommendations occur more frequently in particular patient contexts, languages or documentation patterns. For patient-facing conversational AI, performance may also include whether different users can understand and successfully interact with the system, not simply whether the generated answer is technically correct.
Healthcare organizations should therefore avoid beginning with a generic fairness metric and forcing the clinical problem to fit it. Begin with the potential harm, then choose the measures needed to detect it.
Human-AI performance can differ even when model performance does not
Clinical AI rarely acts alone. A model may have similar technical performance across groups while unequal outcomes emerge once it enters the workflow. Clinicians may respond differently to particular outputs, alerts may be reviewed less consistently in high-workload settings, image quality may differ between sites, or a system may be reliably available at one hospital but intermittently available at another.
The unit of analysis should therefore sometimes be the human-AI system, not the model in isolation. IMDRF’s GMLP principles explicitly emphasize assessing human-AI interactions and the performance of the human-AI team in the intended use environment rather than evaluating the device alone.[6]
A fairness assessment that looks only at model outputs may consequently miss inequity introduced by workflow, infrastructure or human behavior.
Equity extends beyond algorithmic bias
Suppose an AI-assisted service performs similarly at two hospitals, but only one has the infrastructure to use it reliably. Or a remote-monitoring system performs technically well, but some patients cannot participate because they lack suitable devices, stable connectivity or digital skills. Or a chatbot produces accurate responses in standard written Vietnamese but remains difficult for some people to use because of language or accessibility barriers.
Those situations may not represent model bias, but they can still create unequal benefit.
WHO/Europe’s March 2026 scoping review found that equity is increasingly recognized in digital health but remains inconsistently incorporated into regulation, implementation and evaluation. Standardized equity metrics and robust bias auditing for AI-driven technologies remain limited, while structural barriers such as access, digital literacy, infrastructure and adaptation to diverse needs continue to matter. The review recommends an equity-by-design approach across the digital-health lifecycle.[7]
WHO/Europe has also noted that equity, bias and fairness checks for AI-driven health technologies are not yet standard practice. These findings primarily concern the WHO European Region rather than Viet Nam, but the implementation lesson is relevant: equity depends on how technology interacts with the health system, not only on algorithmic performance.
Viet Nam’s AI framework provides direction, not a universal clinical fairness metric
Viet Nam’s AI Law establishes fairness, avoidance of bias, non-discrimination and prevention of harm among its basic principles.[1] The National AI Ethics Framework provides more operational direction by recommending identification and mitigation of data, model and operational bias and consideration of impacts on vulnerable groups.[2]
Neither instrument, however, establishes one universal statistical threshold that healthcare organizations must use to prove that a clinical AI system is “fair.” Organizations should therefore avoid inventing legal thresholds that do not exist. A hospital should not claim that a system is legally fair simply because sensitivity differs by less than a chosen percentage between groups unless an applicable regulation, technical standard or validated framework actually establishes that criterion.
Acceptable performance differences should instead be considered in light of intended use, clinical consequences, evidence, uncertainty and any applicable regulatory or technical requirements. The goal is accountable judgment, not false numerical certainty.
Patient-group data require careful governance
Assessing performance differences may require patient characteristics, and some of those data are sensitive. Under Viet Nam’s personal-data framework, data revealing racial or ethnic origin and health status are classified as sensitive personal data; biometric data and genetic characteristics are also included among the sensitive categories.[3,4]
Not every variable used in subgroup analysis has the same legal status. Age, hospital site, equipment type and image quality should not simply be treated as equivalent to ethnicity or health-status information.
Healthcare organizations should therefore determine which variables are genuinely necessary, what purpose supports their processing, who needs access, which safeguards apply and whether aggregated or otherwise appropriately protected analysis can achieve the objective. Fairness assessment should not become a reason to collect every sensitive characteristic available.
At the same time, refusing to examine clinically relevant performance differences simply because the analysis is difficult can leave important safety problems invisible. Responsible governance requires both data protection and enough evidence to understand performance.
Intersectionality can matter, but sample size becomes a constraint quickly
Patient characteristics do not occur one at a time. An older patient with severe disease treated at a particular site with a particular device may experience a combination of conditions that is not visible when age, severity, site and equipment are examined separately.
Examining combinations of characteristics can therefore reveal risks hidden by single-variable analyses. But every additional subgroup division reduces sample size, and estimates can become unstable quickly.
Intersectional analysis should be used selectively, guided by plausible risk and sufficient evidence. Where local datasets are too small, multicenter evidence, longer monitoring periods, provider data or pooled evidence from genuinely comparable settings may be more informative than increasingly granular local comparisons.
When a performance difference appears, investigate before intervening
A subgroup difference is a signal, not a diagnosis of the problem. The cause may lie in training data, but it may also involve prevalence, case mix, image quality, missing data, measurement differences, equipment, site configuration, workflow, user behavior or simply inadequate sample size.
A practical investigation can ask: Is the difference statistically credible? Is it clinically meaningful? Does it persist after considering relevant case mix? Does it occur across sites or only in one environment? Are input quality or device differences involved? Was the group adequately represented during development? Does the problem arise from the model itself or from how the system is being used?
Changing a model or threshold before understanding the mechanism can simply shift harm elsewhere. Investigation should therefore come before mitigation.
Mitigation can occur at several levels
Not every fairness problem requires retraining the model. Possible responses include improving data quality, obtaining more representative evidence, adjusting configuration or thresholds where appropriate, restricting use to populations for which performance is sufficiently supported, strengthening human oversight, modifying workflow, improving staff training, improving accessibility or asking the provider to correct the underlying system.
In some situations, enhanced monitoring may be appropriate while evidence accumulates. In others, use may need to be restricted or suspended. The mitigation should match the cause.
If performance differs because one site uses incompatible imaging settings, changing the model threshold may not solve the problem. If older patients were poorly represented in the evidence, redesigning the interface alone will not remove that uncertainty. If patients cannot access a remote service because of connectivity or language barriers, improving model discrimination will not solve the equity problem.
Fairness and equity are therefore system-level governance issues.
Monitoring should continue after go-live
A model that performs acceptably across groups at launch may not remain that way. Patient populations can change, prevalence can shift, new equipment can be introduced, deployment can expand to new sites, clinical practice can evolve and the provider may update the model.
Post-deployment monitoring should therefore consider whether clinically important performance differences emerge over time. Not every subgroup metric needs continuous monitoring; attention should focus on groups and conditions associated with plausible risk, important evidence gaps or findings identified during local validation.
Triggers for review may include widening performance differences, clusters of incidents in a particular population, patient or clinician complaints, new provider evidence, expansion into a new population or site, or a material system change.
In its 2025 Request for Public Comment on real-world performance of AI-enabled medical devices, FDA noted that changes in patient demographics, clinical practice, data inputs and healthcare infrastructure can affect AI performance, while user behavior, workflow integration and changing clinical guidelines may also alter system behavior in practice. FDA expressly states that this document is for discussion purposes and does not represent draft or final guidance.[8]
Incidents and complaints can be equity signals
Unequal performance may first appear not in a dashboard but in an incident report. A clinician may notice repeated missed findings in a particular type of patient. A department may observe poorer outputs under certain acquisition conditions. Patients using one language may struggle with a chatbot. One group may receive substantially more false alerts than expected.
These signals should not automatically be dismissed as anecdotes. Healthcare organizations should connect bias and equity concerns with existing incident-reporting, quality-improvement and patient-feedback mechanisms so that relevant signals can be investigated against performance data.
The National AI Ethics Framework also encourages users to use feedback or complaint channels when they identify signs of bias, discrimination or harmful content, and encourages developers, providers and deployers to establish accountability and remediation mechanisms.[2]
Bias and equity should connect to the AI risk register
A finding has little value if nobody owns it. Where subgroup performance differences create meaningful risk, they should be linked to the organization’s AI risk-management process. The record should identify the affected system and population, evidence supporting the concern, potential patient impact, existing controls, accountable owner, mitigation action and monitoring plan.
Uncertainty should also be recorded. An entry such as “Performance in patients over 80 remains uncertain because of limited local sample size; enhanced monitoring required during early deployment” may be considerably more useful than prematurely classifying the system as either “fair” or “biased.”
Fairness governance should support decisions, not produce labels.
A practical Bias, Fairness and Equity Gate
Before routine deployment or wider scale, a healthcare organization can apply a simple Bias, Fairness and Equity Gate:
Population definition: Is the intended patient population clearly defined, including groups or operating conditions that may plausibly affect performance?
Evidence representation: Do development and validation data adequately represent the patients, use environments and measurement inputs relevant to the intended local use?
Performance assessment: Have clinically important subgroup differences been evaluated where evidence and sample size allow meaningful analysis?
Clinical interpretation: Where differences exist, have we determined whether they are clinically meaningful, expected, uncertain or potentially harmful rather than assuming every numerical difference represents unfairness?
Data governance: Is the collection and processing of demographic, health, ethnic or other subgroup information relevant, proportionate and appropriately protected?
Human–AI and workflow assessment: Could workflow, equipment, clinician behavior, language, accessibility or infrastructure create unequal performance or benefit even when model performance is similar?
Mitigation: Are identified problems being addressed at the correct level—data, model, workflow, training, access, oversight, communication or approved-use restrictions?
Monitoring: Are important performance differences and evidence gaps included in post-deployment monitoring and escalation triggers?
Accountability: Is there a clear owner responsible for investigating and managing clinically meaningful bias or equity concerns?
This Gate is a practical organizational framework. It is not a statutory fairness test or mandatory checklist under Vietnamese law and should be adapted to the AI system, intended population, clinical risk and available evidence.
What this means for healthcare organizations in Viet Nam
Healthcare organizations in Viet Nam should not begin by asking whether an AI system is “biased” in the abstract. They should begin with the patients, devices, data and workflows in which the system will actually be used.
Who was represented in the evidence? Which patient or technical characteristics could plausibly affect performance? Are there important local populations for which evidence is weak? Do observed differences matter clinically? Could the workflow itself create unequal benefit? What remains uncertain?
This is especially important when AI developed elsewhere is introduced into Vietnamese healthcare settings. International evidence is valuable and should not be discarded, but responsible implementation requires understanding whether that evidence transfers adequately to local patients, devices, language, workflows and sites.
Healthcare organizations should resist two extremes. The first is to ignore subgroup differences because the overall performance metric looks good. The second is to treat every numerical difference between patient groups as proof of bias or discrimination. Both are too simplistic.
Responsible clinical AI requires something more disciplined: identify plausible differences, measure them where the evidence allows, understand why they occur, judge their clinical significance, mitigate avoidable harm and continue monitoring after deployment.
For healthcare leaders, the practical question is therefore not simply “Does this AI perform well?” It is “Does it perform well enough across the patients we intend to serve—and would we know if some patients were benefiting less or being harmed more than others?”
That is the equity question responsible clinical AI governance needs to answer.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, statistical, ethical, medical-device or personal-data advice.
The definitions, subgroup-assessment approach and Bias, Fairness and Equity Gate described in this article are practical organizational recommendations. They are not statutory definitions of bias or fairness and do not create mandatory subgroup metrics or numerical thresholds under Vietnamese law.
The National Artificial Intelligence Ethics Framework under Circular No. 05/2026/TT-BKHCN applies within the scope stated in that Circular; other organizations conducting AI activities in Viet Nam are encouraged to apply it.
Healthcare organizations should separately assess applicable obligations under the Law on Artificial Intelligence, personal-data protection legislation, medical-device regulation, professional requirements and other specialized legislation. Collection or processing of sensitive patient characteristics for bias or equity evaluation should not be assumed to be automatically permissible; organizations should determine the appropriate legal basis, necessity, proportionality, security and governance for the actual activity.
References
[1] National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
[2] Ministry of Science and Technology of Viet Nam. Circular No. 05/2026/TT-BKHCN promulgating the National Artificial Intelligence Ethics Framework. 10 March 2026; effective 10 March 2026.
[3] National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
[4] 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. 31 December 2025; effective 1 January 2026.
[5] International Medical Device Regulators Forum. Guiding Principles to Support Medical Device Health Equity. IMDRF/MC/N79 FINAL:2024. Final document dated 6 February 2024.
[6] International Medical Device Regulators Forum. Good Machine Learning Practice for Medical Device Development: Guiding Principles. IMDRF/AIML WG/N88 FINAL:2025. Final document dated 27 January 2025; published by IMDRF on 29 January 2025.
[7] World Health Organization Regional Office for Europe. Equity across the Regulation, Implementation and Evaluation of Digital Health: Scoping Review. 17 March 2026. WHO/EURO:2026-13153-52927-82472.
[8] U.S. Food and Drug Administration. Measuring and Evaluating Artificial Intelligence-Enabled Medical Device Performance in the Real World: Request for Public Comment. 2025. Discussion document only; not draft or final FDA guidance.