Clinical AI Confidence, Calibration and Uncertainty in Viet Nam: How Healthcare Organizations Should Interpret Scores, Set Thresholds and Decide When AI Should Abstain or Defer to Human Review
IMPLEMENTATION INSIGHTS
Clinical AI Confidence, Calibration and Uncertainty in Viet Nam: How Healthcare Organizations Should Interpret Scores, Set Thresholds and Decide When AI Should Abstain or Defer to Human Review
IMPLEMENTATION INSIGHT
Last reviewed: 3 October 2026
A clinical AI system says a patient has an 82% risk of deterioration. Another system reports “high confidence” that an image is abnormal. A generative-AI assistant provides a treatment recommendation and says it is 95% confident. Those numbers can look reassuringly precise, but they may mean very different things, and sometimes much less than clinicians assume.
An 82% output may be an estimated probability, a normalized model score or simply a number scaled between zero and one. “High confidence” may describe how strongly a classifier favors one category over another without showing whether that confidence corresponds to real-world correctness. For a large language model, a statement such as “I am 95% confident” may itself be generated text rather than a validated estimate of the probability that the answer is correct.
This distinction matters because healthcare organizations cannot safely govern clinical AI if they do not understand what its scores mean, how reliable those scores are, what uncertainty remains and what should happen when the system is not reliable enough to support the next clinical action. Recent research is increasingly moving beyond the question “Is the AI accurate?” toward a more operational one: “When is it reliable enough to act on, and when should it abstain or hand the case back to a human?” A 2026 npj Digital Medicine study proposed separate rule-in, rule-out and gray zones to translate reliability targets into clinical workflows, while a 2026 review of LLM abstention emphasized that knowing when not to answer is itself an important safety capability [9,11].
For healthcare organizations in Viet Nam, confidence, calibration and uncertainty should therefore be treated as governance questions, not merely model-development terminology.
A score is not automatically a probability
The first question to ask when an AI product displays a number is surprisingly basic: What exactly does this number represent?
Some predictive models produce probabilities. If a well-calibrated model assigns a predicted probability of 20% to a sufficiently large group of comparable patients, approximately 20% of those patients should experience the outcome within the specified time horizon. That is the basic idea of calibration.
But many AI outputs are not directly interpretable in this way. A classification system may expose a softmax score, similarity score, decision margin, anomaly score or proprietary confidence index. An imaging product may display “0.93” without that number having been validated as a 93% probability that the finding is correct. A vendor may convert an internal model output into labels such as “low,” “moderate” and “high confidence” without showing how those categories correspond to observed performance.
Healthcare organizations should therefore resist the intuitive interpretation that 0.90 means “the AI is 90% sure.” Before using a score clinically, the organization should understand whether it is intended to represent an absolute risk, a class probability, a ranking score, confidence in classification, an uncertainty measure or another technical quantity. If the vendor presents it as a probability, the next question is whether that probability has been empirically calibrated in a relevant population.
This distinction is particularly important because discrimination and calibration are different properties. A model can rank patients very effectively from lower to higher risk while systematically overestimating or underestimating their absolute probabilities. Conversely, recalibration may improve probability estimates without changing the model's ability to rank patients.
A 2025 Lancet Digital Health methodological review examining 32 performance measures for predictive AI explicitly separates discrimination, calibration, overall statistical performance, classification performance and clinical utility. It recommends assessing calibration directly rather than assuming that good AUC or classification performance implies reliable probabilities [7].
Calibration asks whether predicted probabilities behave like probabilities
Consider an AI model predicting a 24-hour risk of clinical deterioration. If patients given approximately a 10% predicted risk deteriorate about 10% of the time, patients given approximately 30% risk deteriorate about 30% of the time, and patients given approximately 70% risk deteriorate about 70% of the time, the model is behaving in a calibrated way across those risk levels.
A model can instead be overconfident. Patients assigned 80% risk may actually experience the outcome only 55% of the time. Or it may be underconfident: a group assigned 30% risk may experience the outcome 50% of the time. Calibration is therefore not about whether every individual prediction turns out correctly. A patient with a genuinely calibrated 10% risk can still experience the event, while a patient with 90% predicted risk can still avoid it. Probability describes uncertainty across comparable situations; it does not determine an individual patient's future.
Several approaches can be used to assess calibration. Calibration plots, calibration-in-the-large, calibration intercept and calibration slope assess different aspects of calibration. Proper scoring rules such as the Brier score can provide complementary information about overall probabilistic performance, but they do not isolate calibration on their own [7,8].
Healthcare leaders do not need to become biostatisticians to govern these systems, but somebody involved in validation and oversight needs to understand the distinction. A dashboard showing an AUC of 0.94 and “accuracy 91%” does not answer the question of whether an output displayed as 35% risk actually behaves like a 35% risk in the population where the system is being used.
High discrimination does not tell you where the AI can safely act
AUC is useful because it measures how well a model separates cases with different outcomes. But it does not by itself determine whether an AI output is reliable enough for a particular clinical action.
This becomes obvious when the costs of errors differ. A cancer-screening system used to deprioritize apparently negative studies may need exceptionally strong rule-out reliability because a false negative could delay diagnosis. A system used only to prioritize cases for earlier review may tolerate more false positives because its output does not remove patients from clinical assessment. A deterioration alert used to summon a rapid-response team creates different consequences from the same model used only to reorder a clinician's worklist.
The same underlying AI model can therefore require different operating thresholds depending on what the healthcare organization allows the output to do.
Research published in 2026 illustrates this distinction. The proposed Safety-Aware ROC framework separates a Rule-in Safe Zone, a Rule-out Safe Zone and an intermediate Gray Zone in which human review remains necessary. In a comparison involving two FDA-cleared cancer-screening AI systems, the system with the better conventional AUC was not always the one with the better operational profile under stringent reliability requirements [9].
That framework is a research proposal, not a universal healthcare standard and not a requirement under Vietnamese law. But the underlying governance principle is highly relevant: the operational question is not simply how well the model performs overall, but where its outputs are reliable enough for the action the organization intends to attach to them.
Thresholds translate model outputs into clinical policies
A decision threshold is sometimes presented as though it belongs naturally to the model. In practice, a threshold is the point at which a model output is translated into a workflow or clinical action.
Suppose a deterioration model produces risk estimates between 0% and 100%. A hospital might decide that patients above 70% trigger immediate escalation, patients between 30% and 70% require clinician review, and patients below 30% remain within the ordinary workflow. Those cutoffs are not simply statistical properties of the model. They create an operating policy around the model.
Choosing thresholds requires judgment about the intended clinical use, consequences of false positives and false negatives, available clinical capacity, time sensitivity, alternative pathways and the reliability of estimates around each threshold. Depending on the product, an operating point may be predefined and validated by the manufacturer, configurable within an approved range or determined locally by the healthcare organization. Whatever the arrangement, the healthcare organization needs to understand what clinical consequences follow from the threshold actually being used.
The costs of error can also be asymmetric. Missing one deteriorating patient may matter more than reviewing several additional false alerts. In another workflow, excessive false positives may themselves create harm by overwhelming specialists, delaying higher-priority cases or generating alert fatigue.
Threshold selection therefore requires balancing safety, clinical utility and workload, rather than simply maximizing an abstract statistical metric. The 2025 Lancet Digital Health review makes the same point by recommending assessment not only of discrimination and calibration but also of clinical utility, including methods such as decision-curve analysis that explicitly relate predictions to decisions and consequences [7].
A threshold that works elsewhere may not work the same way in Viet Nam
This is particularly relevant when AI developed overseas is introduced into a Vietnamese healthcare organization. Clinical prevalence may differ. Referral patterns may differ. A tertiary hospital will see a different patient population from a screening clinic. Documentation practices, laboratory testing, disease severity at presentation and access to follow-up care may all affect the distribution of cases arriving at the model.
These differences matter because measures such as positive predictive value and negative predictive value depend on the population in which the model operates. A threshold that produces a highly reliable positive result in a high-prevalence specialist population may behave differently when the same system is deployed for broad screening.
This does not mean every healthcare organization needs to develop or recalibrate its own AI model. It means local validation should examine the operating policy, not simply reproduce the vendor's global AUC. Where the AI produces probabilities and sufficient local data are available, calibration should be evaluated in data that reasonably reflect the intended deployment context. The organization should also examine the actual thresholds it plans to use and the clinical workload those thresholds create.
A useful practical question is: If we apply this threshold to our patients, how many cases will be escalated, deferred, missed or unnecessarily reviewed, and are those consequences clinically acceptable?
Uncertainty is not one thing
Healthcare discussions often refer to “AI uncertainty” as though it were a single measurable variable. In practice, several kinds of uncertainty may exist at the same time.
Aleatoric uncertainty arises from inherent variability or ambiguity in the underlying clinical problem. Two patients with apparently similar characteristics can still have different outcomes because medicine itself is uncertain. Epistemic uncertainty reflects limitations in the model's knowledge — for example, because relevant situations were poorly represented during development or because the current case differs from the data on which the system learned.
There is also uncertainty about the estimated risk itself. Prediction models are developed from finite datasets, and model parameters and individual risk estimates can have substantial statistical uncertainty. A 2025 BMJ article argues that clinical prediction models frequently provide point estimates without adequately communicating the uncertainty around those estimates [8].
In deployed systems, another practical layer can be just as important: operational uncertainty. This term is used here as an implementation concept rather than as a universal statistical taxonomy. The model may not know that a laboratory result is delayed, an imaging study is incomplete, an interface dropped a field or a patient belongs to a workflow different from the one assumed during validation.
These forms of uncertainty require different controls. More training data may reduce some epistemic uncertainty but cannot eliminate biological randomness. Calibration can improve probability estimates but cannot repair incorrect upstream data. Human review can help resolve an ambiguous case but may not detect a silent unit conversion in the data pipeline. The organization therefore needs to understand what a vendor's “uncertainty” measure is actually measuring.
Confidence should not be confused with certainty about the individual patient
Even a perfectly calibrated model does not know an individual patient's outcome in advance. If an AI predicts a 15% risk of readmission, that should not automatically be communicated as “the AI thinks the patient probably will not be readmitted.” Fifteen percent may be relatively low within the model's distribution but still high enough to trigger intervention in a particular pathway. Likewise, a 90% predicted probability does not make the remaining uncertainty irrelevant. Whether a threshold is actionable depends on what happens when the system is wrong.
Healthcare organizations should therefore avoid interfaces that collapse prediction, confidence and recommendation into one visual signal. A red icon labeled “HIGH RISK” can make a probabilistic estimate look like a clinical conclusion. A green icon may imply safety even when the model has simply placed a patient below a cutoff.
A better interface makes the distinction visible: what the model predicts, how reliable that prediction is, what important uncertainty remains and what action the organization has authorized at that level of evidence.
Generative AI creates a different confidence problem
The word “confidence” becomes even more problematic with large language models. A predictive classifier can sometimes be calibrated against observed outcomes. Generative AI, however, produces sequences of language. A model can write, “I am highly confident that this is the correct treatment,” but that sentence should not be assumed to represent a calibrated probability of clinical correctness.
Research published in February 2026 evaluated self-reported confidence across 48 commercial and open-source LLMs using 300 gastroenterology board-style questions. Accuracy varied substantially, but the models generally reported high confidence whether answers were correct or incorrect. All tested models showed systematic overconfidence, and the study concluded that self-reported confidence should not be used as a standalone indicator of reliability [10].
The study has important limitations: board-style multiple-choice gastroenterology questions are not the same as real-world, open-ended clinical conversations. But the governance implication remains important. If an LLM displays “confidence: 95%,” the healthcare organization should ask how that number was generated and validated. Was it simply requested in the prompt? Was it derived from token probabilities? Was an external uncertainty model used? Was the score calibrated against actual correctness in comparable clinical tasks?
Without an empirical answer, the number should not be presented to clinicians as though it were a validated clinical probability.
Additional uncertainty or reliability signals may sometimes be derived from mechanisms such as evidence availability, retrieval quality, external verification, rule-based checks, dedicated uncertainty classifiers or explicit detection of missing context. Agreement across repeated outputs may provide information about output stability, but repeated agreement does not establish that the answer is correct. None of these signals should automatically be equated with the probability that an entire generated clinical response is reliable.
Abstention is a clinical capability, not simply a model failure
Clinical AI does not always need to produce a definitive answer. If the model is outside its validated domain, required information is missing, evidence is conflicting or uncertainty is too high for the intended action, the appropriate response may be to abstain.
For predictive AI, abstention may mean withholding classification or routing the case into a human-review pathway. For imaging AI, it may mean declining to analyze an image that fails quality checks. For generative AI, it may mean saying that available evidence is insufficient, asking for missing clinical information or explicitly deferring the question to an appropriate clinician.
A 2026 review in npj Digital Medicine distinguishes two broad reasons for LLM abstention in healthcare: uncertainty-driven abstention, where the system lacks sufficient confidence or information, and safety-driven abstention, where responding may create harm or violate relevant constraints [11]. The review also highlights the opposite problem: excessive refusal can reduce utility and obstruct appropriate care.
The governance objective is therefore not to make AI refuse as often as possible. It is to make abstention proportionate to risk. An AI that never abstains may continue confidently into situations it cannot handle, while a system that routinely defers straightforward cases can become unusable and overwhelm clinicians with unnecessary review.
Human review needs capacity, not just a checkbox
“Human in the loop” is often offered as the solution to uncertainty, but sending uncertain cases to people works only if the human-review pathway is itself designed to work. Suppose a hospital raises the confidence threshold so the AI handles fewer cases automatically. Reliability may improve for the cases retained by the system, but the gray zone grows and more patients require human assessment.
Who reviews them? How quickly? What happens overnight? Does the specialist have access to the original data or only the AI summary? Can the reviewer override the AI easily? Does the workflow preserve the reason the case was deferred?
The 2026 operational-safety work makes this trade-off particularly visible: increasing required reliability can reduce the proportion of cases that qualify for autonomous or semi-autonomous handling while increasing the proportion requiring human review [9]. Threshold governance therefore cannot be separated from workforce and workflow design. A statistically cautious threshold that sends 70% of cases into a review queue that cannot process them safely is not automatically a safer implementation.
Selective prediction can formalize when AI should defer
The broader technical field increasingly refers to selective prediction or prediction with a reject option: the system produces a prediction when predefined reliability conditions are satisfied and abstains or defers otherwise.
A 2026 Scientific Reports study applied calibrated probabilities, conformal prediction and cost-aware deferral to early sepsis prediction. In that research setting, the system routed lower-confidence cases to clinicians and demonstrated a favorable risk–coverage trade-off, including under temporal distribution shift [12].
This is promising, but individual methodological studies should not be treated as universal recipes for clinical deployment. Conformal methods, for example, provide specific statistical coverage properties under assumptions that need to be understood; they do not create a general guarantee that the AI is clinically safe. Distribution shift, outcome definition, dataset quality, action consequences and workflow remain important.
The broader lesson is more durable: an AI system does not have to make the same kind of decision for every case. Healthcare organizations can define pathways in which sufficiently reliable outputs support one action, intermediate outputs require human assessment, and outputs outside acceptable conditions are rejected or withheld.
Calibration should be evaluated in the deployment context, and should not be assumed to remain stable
A model that was calibrated during development can become miscalibrated after deployment. Disease prevalence may change. A hospital may expand its referral network. A new diagnostic pathway may alter which patients receive testing. A laboratory method can change. Documentation may improve. The model may be moved from an inpatient population to an ambulatory service. Even if discrimination remains relatively stable, absolute probability estimates may no longer mean what they did originally.
This is one reason calibration should be treated as a lifecycle property, not a certificate issued once. Where sufficient outcome data become available, organizations should periodically compare predicted probabilities with observed outcomes in the population actually receiving the AI-supported service. The frequency should reflect the use case, case volume, clinical consequence and speed of change rather than following one universal interval.
Recalibration may sometimes be sufficient. Depending on the model and problem, probability estimates can be adjusted without rebuilding the complete model. In other cases, deteriorating calibration may indicate a deeper population or workflow shift requiring broader revalidation or model revision.
Any recalibration also needs change control. Changing a calibration layer or threshold can alter how patients flow through a clinical pathway even if the underlying model weights never change.
Overall calibration can hide subgroup problems
A model can appear reasonably calibrated for the overall patient population while behaving differently in smaller groups. Overprediction in one group can be offset statistically by underprediction in another, allowing an aggregate calibration curve to look acceptable even though reliability differs across the groups.
Where clinically relevant and supported by adequate sample sizes, healthcare organizations should therefore examine calibration and threshold performance across important deployment groups, for example by site, service, age or sex, rather than assuming one aggregate measure is sufficient.
This requires care. Small subgroups produce uncertain estimates and can make apparent calibration differences unstable. Subgroup analysis should therefore acknowledge statistical uncertainty and should not convert random fluctuations into strong claims of bias or safety problems. The same principle applies across hospitals: a threshold validated at one site should not automatically be treated as locally validated at another site simply because both belong to the same healthcare group.
Thresholds and calibration need explicit ownership
Clinical AI governance often identifies an owner for the system but not for the operating threshold. For clinically consequential systems, that is an important gap.
The organization should be able to answer: Who approved this threshold? What evidence supported it? Which population was used? Which action occurs above or below it? What error trade-off was accepted? When will the threshold be reviewed? Who has authority to change it?
Thresholds should not be quietly changed in an IT configuration screen simply because alert volume is inconvenient. Nor should a vendor be able to alter operational cutoffs remotely without the healthcare organization understanding whether the change affects the validated workflow.
A threshold decision should be documented alongside the system's intended use, validation evidence, clinical rationale, expected review workload and relevant fallback arrangements. For systems with multiple thresholds, each one may have a different purpose. A low threshold might trigger screening or additional assessment, while a much higher threshold could trigger urgent escalation. Treating both as simply “positive” obscures the actual operating policy.
Where software also falls within Viet Nam's medical-device framework, threshold changes may have additional implications if the operating point forms part of the manufacturer's validated or intended configuration. Healthcare organizations should therefore consider manufacturer instructions, validation evidence and applicable regulatory requirements before changing such settings [6].
What Viet Nam's current AI framework does and does not require
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 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 requirements for providers and separate operational responsibilities for deployers. The framework addresses risk identification and assessment, data quality, human oversight and intervention, monitoring and risk controls. Article 13 also requires the provider to reassess conformity where significant changes affect the basis of the original conformity assessment, including certain changes in system architecture, model configuration, data sources, processing methods or operating environment [2].
These provisions support the importance of reliability and human oversight, but they should not be overstated. Vietnamese AI legislation does not prescribe a universal calibration metric, one mandatory confidence threshold, one required abstention mechanism or one “safe” percentage that every healthcare AI system must use. The calibration, threshold and deferral controls discussed in this article are practical governance measures that should be designed according to the system's intended use and clinical risk.
Decision No. 33/2026/QĐ-TTg, effective from 15 August 2026, establishes the current 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. Clinical prediction models, generative assistants and other AI systems should therefore not automatically be described as legally high-risk simply because they operate in healthcare. Their classification must be assessed according to the actual system, intended use and legal framework in force at the relevant time.
Being outside the current high-risk list does not, however, mean that an AI system falls outside Viet Nam's AI framework. Other requirements under the AI Law and implementing instruments may still apply depending on the system, its risk classification, functions and deployment context.
Threshold changes can also matter even when formal high-risk conformity reassessment is not triggered. If changing a threshold materially changes who receives escalation, what care may be delayed or which decisions can proceed without human review, the organization should treat the change as an internal clinical AI change event proportionate to its patient-safety significance.
Personal-data rules add transparency considerations for automated processing
Viet Nam's Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP have been effective since 1 January 2026 [4,5]. Decree No. 356 includes specific provisions concerning personal-data protection in AI systems and automated processing.
Where the relevant automated-processing provisions apply, personal data controllers and controller-processors have responsibilities that include notifying the data subject about automated personal-data processing, explaining the operating principles of the algorithm and its effects on the data subject's lawful rights and interests, and providing options enabling the data subject not to participate in that automated processing [5].
This does not mean Vietnamese law requires a hospital to disclose a calibration curve, mathematical model or every technical threshold to every patient. Nor should the provision be simplified into a general statutory “right to an AI confidence score.” The precise requirements depend on the actual processing arrangement and legal context.
The provisions nevertheless reinforce a broader governance principle: healthcare organizations should understand how automated scores influence decisions and should be capable of explaining the relevant operating logic where the law requires it. A proprietary threshold that nobody in the healthcare organization understands is difficult to govern meaningfully.
AI-derived information that can be used to identify or help identify an individual also remains subject to applicable personal-data protection requirements [5].
Vendors should explain what their confidence score actually means
A vendor due-diligence discussion should not stop at the question “Does the system provide confidence scores?”
The more useful questions are: What is the mathematical meaning of the score? Is it a calibrated probability? Against which outcome and time horizon? Which population was used for calibration? How was calibration assessed? Does the score retain the same interpretation across sites? Does the vendor provide calibration plots or only a summary metric? How are thresholds selected? Can the healthcare organization change them? What happens when the output falls below the threshold? Does the system abstain, generate a warning or simply continue? Has performance been evaluated under distribution shift? Are clinically relevant subgroup results available? How are calibration changes monitored after software updates?
For generative AI, the questions should be even more skeptical. Is a displayed confidence score empirically associated with correctness, or is it generated by the LLM itself? What does “low confidence” actually cause the system to do? Can it ask for clarification? Can it distinguish insufficient evidence from a prohibited request? Can it defer to an appropriate clinician without first producing a definitive answer?
The phrase “our AI provides a confidence score” is not enough. The score needs an operational interpretation.
A practical Clinical AI Confidence, Calibration and Uncertainty Gate
Before routine use, a healthcare organization can begin with score meaning. What does the AI output represent: probability, ranking score, similarity measure, confidence index or something else? If it is presented as a probability, has calibration been demonstrated? If a generative system reports confidence, how is that confidence produced and has it been validated against actual correctness?
The next layer is deployment-context reliability. Has calibration been evaluated in data sufficiently relevant to the intended Vietnamese use? Does performance remain acceptable at the score ranges where clinical actions will occur? Do site, service or clinically relevant subgroup differences materially affect reliability? Is uncertainty around the calibration estimates themselves understood?
A third layer is threshold governance. Which clinical action is attached to each threshold? What are the consequences of false positives and false negatives? Who approved the threshold? Was clinical workload considered? Does a change in prevalence alter positive or negative predictive value at the operating point? Are separate rule-in and rule-out pathways appropriate rather than forcing every patient through one binary cutoff?
The organization should then define uncertainty and deferral behavior. When should the AI abstain? When should it request more information? Which cases require mandatory human review? What happens when data are missing, inconsistent or outside the validated domain? Is abstention visible and understandable to the clinician? Can the human reviewer see enough information to make an independent decision?
Finally comes lifecycle monitoring and change control. Which calibration, threshold-performance and workload measures will be monitored after go-live? What deterioration triggers investigation or recalibration? Which changes to the model, population, data pipeline, clinical workflow or operating threshold require review or revalidation? Who can change thresholds, and are those changes version-controlled and auditable?
The Clinical AI Confidence, Calibration and Uncertainty Gate is a practical organizational governance framework proposed for implementation and patient-safety purposes. It is not a statutory checklist or a calibration standard prescribed by Vietnamese law.
What this means for healthcare organizations in Viet Nam
Healthcare AI is often presented through impressive numbers: AUC 0.94, sensitivity 96%, confidence 92%, predicted risk 78%. The hardest governance question is not whether the number looks good. It is whether the healthcare organization understands what the number means in the clinical workflow in which it will be used.
A probability needs calibration. A threshold needs a clinical rationale. An uncertainty signal needs a validated interpretation. A gray zone needs enough human capacity to review it. An abstention pathway needs somewhere safe to send the patient or case. And a confidence statement from generative AI should not be treated as evidence simply because it sounds precise.
Recent research is pushing clinical AI toward more explicit selective prediction, abstention and operational safety. But these approaches do not remove the need for organizational judgment. A hospital still has to decide what reliability it requires for a particular action, how much uncertainty it will tolerate and when a human must take over.
That is especially important in Viet Nam, where clinical AI is entering healthcare organizations with different patient populations, information systems, workflows, workforce capacity and levels of digital maturity. A threshold developed elsewhere cannot simply become a local clinical policy because it arrived inside a validated software product.
The leadership question should therefore move beyond “How confident is the AI?” A more useful set of questions is: “What does its confidence actually mean? Is it reliable in patients like ours? What happens around the threshold? And when the system is uncertain, does it know when to stop and return the decision to a human?”
That is where confidence becomes governance rather than decoration.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, statistical, data-protection or medical-device advice.
The terms confidence, calibration, uncertainty, operational uncertainty, threshold, abstention, deferral, rule-in zone, rule-out zone, gray zone and the Clinical AI Confidence, Calibration and Uncertainty Gate are used here as clinical AI evaluation, implementation and governance concepts. The specific organizational framework proposed in this article is not a statutory checklist under Vietnamese law.
Article 15 of Decree No. 142/2026/NĐ-CP establishes risk-management responsibilities for high-risk AI systems, while Article 13 addresses conformity reassessment following significant changes. These provisions should not be interpreted as prescribing one calibration metric, confidence score, threshold-setting method or abstention policy for every healthcare AI system. Decision No. 33/2026/QĐ-TTg establishes the current high-risk list and, as of this review, does not classify every clinical AI application as high-risk merely because it is used in healthcare.
Research frameworks cited in this article — including safe-zone, selective-prediction, conformal and abstention approaches — are included as evidence and implementation references. Their assumptions, validation requirements and limitations need to be understood before they are adapted to local clinical use.
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] Government of Viet Nam. Decree No. 98/2021/NĐ-CP on Management of Medical Devices, as amended, including by Decree No. 07/2023/NĐ-CP and Decree No. 04/2025/NĐ-CP; Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Risk Determination and Management Measures for Products and Goods That Are Medical Devices and Amendments to Circular No. 05/2022/TT-BYT. Issued 30 June 2026; effective 1 July 2026.
[7] Van Calster B, Collins GS, et al. Evaluation of performance measures in predictive artificial intelligence models to support medical decisions: overview and guidance. The Lancet Digital Health. 2025;7(12):100916.
[8] Riley RD, Collins GS, Kirton L, Snell KIE, Ensor J, Whittle R, et al. Uncertainty of risk estimates from clinical prediction models: rationale, challenges, and approaches. BMJ. 2025;388:e080749.
[9] Kim YT, Kim H, Bahl M, Lev MH, González RG, Gee MS, et al. Defining operational safety in clinical artificial intelligence systems. npj Digital Medicine. 2026;9:281.
[10] Naderi N, Safavi-Naini SAA, Savage T, Khalafi MA, Lewis PR, Atf Z, et al. Across generations, sizes, and types, large language models poorly report self-confidence in gastroenterology clinical reasoning tasks. npj Gut and Liver. 2026;3:6.
[11] Presacan O, Nik A, Ojha J, Thambawita V, Ionescu B, et al. When silence is safer: a review and decision-theoretic framework for LLM abstention in healthcare. npj Digital Medicine. 2026;9:749.
[12] Kwon H, Kim DJ. Conformal selective prediction with cost aware deferral for safe clinical triage under distribution shift. Scientific Reports. 2026;16:10016.