From AI Monitoring to a Learning Health System Approach in Viet Nam: Turning Real-World Evidence, Frontline Feedback and Quality Signals into Continuous Improvement
IMPLEMENTATION INSIGHTSHEALTHCARE IMPLEMENTATION, MONITORING & EVALUATIONRESEARCH & CROSS-SECTOR COLLABORATION
IMPLEMENTATION INSIGHT
Last reviewed: 4 October 2026
A clinical AI dashboard can show drift, calibration, latency, override rates, incidents and patient outcomes. But those measurements, by themselves, do not mean that a healthcare organization is learning. Monitoring tells an organization that something may have changed. Learning begins when that signal is interpreted, its significance is understood, practice changes in response, and the organization then checks whether the change actually made care safer or better.
This distinction is becoming increasingly important as healthcare AI moves from pilots into routine operations. In a 2026 BMJ Innovations analysis, Jeffrey Braithwaite describes a shift from periodic measurement toward continuous sensing, while emphasizing that continuous sensing does not automatically produce continuous improvement. Prediction is not prevention, and a learning health system is not created simply by installing more software or generating more data [8].
Recent learning-health-system scholarship reaches a similar conclusion. Bian and colleagues argue that the major bottleneck is no longer simply a shortage of data or algorithms, but insufficient technical, human and cultural infrastructure to turn data and AI into continuous, bidirectional learning. They describe this institutional capability as a “learning utility”: evidence informs practice while practice continually generates information that can refine evidence and improve the system [9].
For healthcare organizations in Viet Nam, the next step after establishing AI monitoring should therefore not simply be another dashboard. It should be a learning loop:
real-world care → data and signals → interpretation → knowledge → improvement action → changed practice → new outcomes → new learning
The central question should move beyond “Are we monitoring the AI?” toward “When monitoring tells us something important, can our organization learn from it and make care better?”
A Learning Health System approach is an organizational capability, not a technology label
A Learning Health System, or LHS, is generally understood as a system in which information generated through healthcare is converted into knowledge, that knowledge is translated back into practice, and the resulting practice produces new information for the next cycle of learning. In simple terms:
data → knowledge → practice → new data → better knowledge → improved practice
The concept is much broader than AI. Steel and colleagues describe an LHS in terms of the infrastructure, stakeholders and culture required to support continuous cycles of evidence generation, translation and healthcare improvement [15]. This is why the word approach matters in the title of this article.
A healthcare organization does not become a Learning Health System because it has an EHR, a data warehouse, an AI model or a real-time dashboard. Nor should Viet Nam be described as already having a national AI-enabled Learning Health System simply because more healthcare information is becoming digital. An LHS is fundamentally a sociotechnical and organizational capability. It depends on data and technology, but also on governance, leadership, frontline participation, evaluation methods, quality-improvement capability, multidisciplinary collaboration and a culture willing to change when evidence shows that current practice is not working.
The HALO model published in 2026 provides a useful contemporary example of how these ideas can be connected to AI. Developed through conceptual system modeling, HALO aligns the traditional LHS cycle with an AI lifecycle and five dimensions of continuous assurance covering performance, safety, equity, transparency and security [10]. It treats AI as a dynamic knowledge artifact embedded within a wider learning and oversight system rather than as a one-time technology deployment. HALO is a conceptual framework rather than an empirically validated standard or a legal requirement, but it illustrates an important principle: responsible AI governance should ultimately connect with the healthcare organization's wider ability to learn and improve care.
Monitoring is the sensing layer, not the learning system
Clinical AI monitoring can generate many useful signals. Calibration may deteriorate. Sensitivity may fall. Performance may differ across patient groups. Clinicians may increasingly override recommendations. Data missingness may rise. Response time may worsen. One site may begin generating far more alerts than another. Staff may report that a generative AI system behaves differently after an update. These signals become useful for learning only when there is a pathway from signal to interpretation to action.
A monitoring program therefore needs more than metrics and thresholds. It needs ownership. Someone must know who reviews an abnormal signal, how quickly review should occur, what additional information is required, whether technical, clinical, quality, privacy, cybersecurity or vendor investigation is needed, what decisions can follow, and how the organization will determine whether its response worked. Without that pathway, a healthcare organization can accumulate increasingly sophisticated data while remaining poor at improvement.
A 2026 npj Health Systems perspective proposed the concept of continuous quality observability, in which AI agents could continuously review EHR narratives and surface emerging quality and safety signals earlier than traditional mechanisms such as manual case finding, self-reporting or billing-code review [13]. The concept is promising, but it primarily strengthens detection. Detection alone does not improve care. A faster signal becomes valuable when the organization can interpret it and respond appropriately.
Start with a learning question, not a dashboard
One of the easiest ways to create an ineffective monitoring program is to collect everything that can be measured. A better approach begins with a defined learning question.
For a deterioration model, the question might be: “Are deteriorating patients still being identified early enough for clinicians to intervene?” Answering it may require model sensitivity, alert timeliness, acknowledgment, time to clinical review, treatment escalation, ICU transfer and patient outcomes.
For an ambient AI scribe, the question may instead be: “Is the system reducing documentation burden without degrading the accuracy, usability or safety of the clinical record?” That requires a different signal set, including use rates, correction burden, clinically important documentation errors, clinician time, patient complaints and possibly downstream clarification requests.
The learning question determines which data matter. This helps prevent dashboard accumulation: measuring dozens of technically available indicators without knowing which decisions those indicators are meant to inform. Every important monitoring domain should therefore connect to a question, an owner and a possible action. If nobody knows what decision could follow an abnormal metric, the organization should ask whether that metric is genuinely useful.
Real-world data are not automatically real-world evidence
Once AI is deployed, healthcare organizations gain access to increasingly rich sources of real-world data: electronic medical records, AI logs, clinical outcomes, incident reports, complaints, audit findings, device information, workflow timestamps, clinician overrides and operational data. These data can be highly valuable, but their existence alone does not make them evidence.
The U.S. Food and Drug Administration distinguishes real-world data, or RWD, from real-world evidence, or RWE. RWD are routinely collected data relating to patient health status or healthcare delivery, while RWE is clinical evidence regarding the use, benefits or risks of a medical product generated through analysis of RWD [11]. These are international evidence-generation concepts rather than Vietnamese statutory definitions, but the distinction is useful for healthcare improvement.
An EHR dataset is real-world data. An appropriately designed analysis showing that an AI-assisted pathway was associated with earlier treatment, while accounting for relevant patient mix, workflow changes, missingness and confounding, may contribute to evidence about real-world performance. By contrast, a dashboard showing fewer adverse events after an AI system was introduced does not, by itself, establish that the AI caused the improvement.
This distinction matters in a learning organization because faster feedback can create pressure to act on weak evidence. Learning should be fast enough to remain useful but rigorous enough to avoid learning the wrong lesson. Healthcare organizations should therefore consider data quality, completeness, representativeness, timing, denominator definition, confounding and other simultaneous changes in the healthcare system before turning an observed pattern into organizational knowledge.
Frontline and patient feedback are learning signals, but they need structure
Clinical AI operates inside real work. Clinicians, nurses, pharmacists, technicians, quality teams, administrative staff and patients may therefore recognize problems before formal monitoring does. A nurse may notice that an AI alert repeatedly appears only after the clinical team has already recognized deterioration. A physician may observe that a recommendation works well for most patients but poorly in one specialty subgroup. A pharmacist may identify a recurrent medication-documentation error. Staff may develop a workaround because AI has introduced an unexpected bottleneck.
These observations matter because they can reveal sociotechnical failure modes that model metrics may not capture. But anecdote alone is not enough. A mature learning process should turn frontline observations into structured information. Useful context may include the system involved, workflow setting, frequency, patient or case type, observed consequence, whether the problem is reproducible, whether staff overrode the AI and what workaround was used. Repeated patterns can then be distinguished from isolated impressions.
Frontline feedback should also continue after the pilot phase. Once implementation becomes routine, trust changes, workarounds emerge, patient populations shift, new staff begin using the system and vendor updates may alter behavior. A 2026 report describing two years of AI governance at Oregon Health & Science University illustrates this progression. After reviewing more than 300 AI requests, the organization connected AI governance with education, evaluation and Learning Health System strategy, including an ecosystem intended to support monitoring and learning over time [12].
Patients are part of this learning architecture as well. A patient may notice that an AI-generated summary contains something that was never discussed. An AI-enabled communication pathway may unintentionally make access to a clinician more difficult. Automated information may be technically accurate while remaining confusing or inaccessible for people with particular language or health-literacy needs.
Patient complaints, corrected records, satisfaction information and communication problems can therefore provide meaningful learning signals where AI is involved. The objective is not to classify every complaint as an AI problem. It is to ensure that, when AI materially affects a patient's experience or care pathway, that information does not disappear into a customer-service process disconnected from AI governance and quality improvement.
Quality and patient-safety signals should feed the same learning architecture
Clinical AI monitoring should not become a separate universe of model metrics. Quality indicators, medical incidents, near misses, complaints, audit findings, mortality reviews, clinical audits, tracer findings, documentation reviews and corrective-action effectiveness checks can all reveal how AI is functioning inside care.
Viet Nam already has important components of this architecture. Circular No. 19/2013/TT-BYT remains recorded as in force and provides a hospital quality-management structure covering quality improvement, quality indicators, internal assessment and organizational responsibilities [6]. Circular No. 43/2018/TT-BYT provides a framework for reporting, analyzing, responding to and preventing medical incidents and expressly recognizes near misses. The national legal database currently records Circular No. 43 as partially expired, so individual provisions should be applied according to their current legal status and together with later applicable law [7].
The practical opportunity is to connect these existing quality and patient-safety mechanisms with AI monitoring. Suppose monitoring shows a rise in clinician overrides. That should not automatically be interpreted as clinicians failing to follow AI recommendations. Quality review may reveal that clinicians have correctly recognized a shift in the patient population that the system has not accommodated. Conversely, repeated overrides may expose poor training, an inappropriate alert threshold, confusing interface design or a workflow problem. The signal does not explain itself; the learning process provides the context needed to understand it.
A closed learning loop requires action and remeasurement
Healthcare organizations sometimes investigate a problem, assign an action and consider the work complete. A genuine learning loop goes further:
detect → interpret → investigate → decide → improve → remeasure → learn → spread or revise
Consider a recurring AI documentation error. Monitoring or frontline feedback identifies the pattern. Investigation finds a source-data mapping problem. The interface is corrected. The organization then remeasures the error rate. If the problem disappears, there is evidence that the corrective action worked. If it persists, the learning cycle continues.
The intervention may not be technical. The response might involve redesigning workflow, changing escalation rules, retraining staff, improving data quality, restricting AI use, modifying a threshold, replacing a vendor component or temporarily suspending the system. The loop closes only when the organization evaluates what happened after the change. That is the difference between recording corrective actions and creating organizational learning.
Learning from AI does not always mean changing the AI
A clinical AI system can generate valuable learning even when the model itself remains unchanged. Monitoring may reveal that the real bottleneck lies in the response pathway rather than prediction. Frontline feedback may show that escalation rules need to change. Outcome analysis may expose an access problem affecting one site. Incident review may find that interface design is encouraging automation bias despite acceptable underlying model performance.
In each of these situations, care may improve by changing workflow, policy, staffing, training, interface design, data quality or governance rather than model parameters. This distinction is central to Learning Health System thinking. The objective is not simply to optimize an algorithm; it is to improve the health system in which the algorithm operates.
Recent work on AI embedded within Learning Health Systems reinforces this point. Gourishankar and Shalhout, writing in October 2026, describe an LHS cycle in which AI contributes to data capture, interpretation and action while real-world outcomes feed back into monitoring, drift management, governance and adaptation [14]. They emphasize that AI should be evaluated not only through technical performance but also through clinical, operational and patient-centered outcomes.
Organizational learning is not the same as autonomous model learning
The phrase “continuous learning” can create confusion in clinical AI. A Learning Health System continuously learns as an organization. That does not mean an AI model should continuously retrain or modify itself without deliberate governance.
A healthcare organization can continuously monitor outcomes, learn from incidents and frontline experience, and improve workflows while leaving the deployed model unchanged until a formally reviewed update is introduced. This distinction is particularly important under Viet Nam's current AI regulatory framework.
Law No. 134/2025/QH15 on Artificial Intelligence has been effective since 1 March 2026, and Decree No. 142/2026/NĐ-CP has been effective since 1 May 2026 [1,2]. For AI systems classified as high-risk, Article 15 of Decree No. 142 establishes risk-management responsibilities for providers and operational risk-management responsibilities for deployers. Among other requirements, deployers are expected to monitor system operation to detect deviations, risks or incidents, maintain appropriate human oversight and apply risk controls within their authority [2].
Article 13 separately requires providers of high-risk AI systems to reassess conformity when significant changes could affect the basis of the previous conformity assessment. The specified circumstances include potentially material changes to the system's purpose or scope, system architecture or AI model, principal data sources or input types, data-processing methods, integration with other systems, or the operating environment [2]. Routine updates, ordinary technical fixes or performance optimization that do not alter compliance with the applicable requirements are not automatically treated as significant changes under that provision.
A learning organization should therefore not interpret continuous improvement as permission for uncontrolled model evolution. Organizational learning can be continuous; material model change should remain governed.
Serious incidents also have a separate legal pathway. Article 19 of Decree No. 142 sets out the serious-incident reporting and handling framework, including events involving death or serious harm to human health, significant property or organizational harm, serious infringement of lawful rights and interests, or serious disruption of essential or public services and certain public-security interests [2]. Providers and deployers have defined responsibilities for response and cooperation, and relevant system logs, data and information must be maintained for verification, evaluation and remediation.
Not every clinical AI error, complaint, override or near miss will meet the statutory threshold for a serious AI incident. Internal quality and patient-safety learning can therefore appropriately operate at a much lower threshold than formal serious-incident reporting.
The system should learn from outcomes, not only from AI performance
An AI model can improve its technical metrics while the patient pathway remains unchanged. A Learning Health System approach should therefore connect AI monitoring with outcomes.
For a sepsis or deterioration system, the learning loop should eventually ask whether recognition became earlier, treatment occurred sooner and clinically meaningful outcomes changed. For a documentation system, the organization should ask whether staff burden, documentation quality and communication improved. For imaging prioritization, the relevant question may be whether clinically urgent patients actually receive faster care rather than whether the algorithm ranks images accurately in isolation.
This distinction remains highly relevant because clinical AI research still relies heavily on process outcomes. A 2026 Lancet Digital Health comment noted that randomized trials of healthcare AI commonly measure diagnostic accuracy, detection rates, turnaround time and workflow efficiency, whereas improvements in such process measures do not necessarily guarantee improvements in mortality, morbidity, quality of life or other patient outcomes [16].
The learning loop should therefore connect:
model behavior → human response → care process → patient outcome → system learning
Otherwise, an organization may become increasingly sophisticated at improving AI performance while remaining uncertain whether AI is improving healthcare.
Learning also requires knowing when not to act
Continuous learning does not mean reacting to every fluctuation. A metric can change through random variation. A small subgroup may generate an unstable signal. A temporary clinical situation may explain a sudden increase in overrides. One patient complaint may represent an isolated experience rather than a systemic failure.
Organizations therefore need decision thresholds and evidentiary discipline. Some signals require immediate action because the potential consequence is severe. Others warrant investigation. Others may need observation until more information accumulates. The ability to distinguish among these situations is itself part of organizational learning. Without that discipline, continuous monitoring can become continuous intervention and, ultimately, continuous instability.
Multi-site learning should travel without erasing local context
AI may be approved centrally but used locally. A healthcare group can deploy the same system across several hospitals or clinics while encountering different patient populations, staffing levels, equipment, documentation practices, workflows and response capacity.
A problem detected at one site can provide early warning to others. If one hospital discovers that an AI tool fails under a particular documentation pattern, that lesson should be shared. But local findings should not automatically be generalized. A workflow intervention that succeeds in one hospital may fail in another because staffing differs. A threshold may behave differently because disease prevalence is different. An interface problem may exist only where a particular information-system configuration is used.
Learning across sites therefore requires two capabilities at the same time: shared institutional learning and preserved local context. Central governance can distribute lessons, standardize minimum controls and identify recurring patterns, while local quality teams need enough authority and visibility to recognize when the shared AI system behaves differently in their environment.
The strongest multi-site organization is not necessarily one in which every site behaves identically. It is one in which local variation becomes visible, interpretable and useful to the wider system.
Organizational memory prevents the same lesson from being learned repeatedly
Healthcare organizations can investigate a problem, solve it and later forget why a particular control was introduced. Staff leave. Committees change. Vendors change. Policies are rewritten. A familiar failure can eventually return.
A Learning Health System therefore needs institutional memory. For important AI systems, this may include the AI inventory, risk register, validation evidence, monitoring history, incident investigations, model and vendor changes, corrective actions, effectiveness checks and the rationale behind major governance decisions.
The objective is not to preserve every email or minor technical ticket. It is to preserve enough context to answer practical questions: Why was this threshold changed? Why was use restricted in one department? What happened during the previous incident? Which version was involved? Why did leadership decide not to scale the system? Which control was introduced, and did it actually work?
An organization that cannot answer those questions may possess large amounts of data without possessing organizational memory.
Viet Nam is generating more digital clinical information, but digitization is not learning
The implementation environment in Viet Nam is changing quickly. Circular No. 13/2025/TT-BYT on electronic medical records has been effective since 21 July 2025 [5]. Hospitals were required to implement electronic medical records no later than 30 September 2025, so that statutory deadline has already passed. Other medical examination and treatment establishments covered by Article 4(2)(b), with patients receiving inpatient, day and outpatient treatment, are required to complete implementation by 31 December 2026 [5].
These deadlines are important because broader electronic documentation can create more opportunities for longitudinal monitoring, quality measurement and real-world evaluation. The fact that an implementation deadline has passed, however, should not be confused with uniform digital maturity. Electronic documentation can differ considerably in completeness, interoperability, usability, data quality and the organization's ability to turn those data into improvement.
Digitization is not the same as learning. An EHR can document care without improving it. A dashboard can display variation without explaining it. An AI model can predict deterioration without ensuring that anyone acts. A data warehouse can contain millions of records without answering the organization's most important improvement questions.
This is especially relevant in Viet Nam because healthcare organizations differ substantially in size, information infrastructure, data maturity, workforce capability and quality-improvement capacity. A Learning Health System approach therefore needs to be proportionate and practical. A smaller specialty organization does not need an enterprise-scale analytics platform before it can create a meaningful learning loop around one or two important AI systems. The starting point is not data volume; it is the ability to turn meaningful information into better care.
Real-world learning still requires data governance
Learning Health Systems depend on data generated through routine care. Those data remain governed data.
Viet Nam's Law No. 91/2025/QH15 on Personal Data Protection and Decree No. 356/2025/NĐ-CP have both been effective since 1 January 2026 [3,4]. Clinical outcomes, patient characteristics, recordings, longitudinal records and some AI logs may contain personal data. Health-status information is classified as sensitive personal data under the current framework [4]. Other logs, user-activity information and operational data should be classified according to their actual content and whether they relate to, identify or help identify an individual rather than being assumed automatically to be sensitive.
Decree No. 356 also contains provisions specifically relevant to AI. Data produced through AI inference must receive personal-data protection where they can be used to identify or help identify a particular person [4]. The Decree also addresses automated processing of personal data, including responsibilities related to informing data subjects and explaining the principles and potential effects of automated processing in applicable circumstances.
A desire to “learn from everything” is therefore not a governance principle. Organizations should identify what information is required, why it is needed, who may access it, how long it should be retained, whether it will be shared externally and what personal-data processing, contractual, security or cross-border requirements may apply.
This becomes particularly important where cloud analytics, external AI platforms, model providers, research collaborators or other third parties receive personal data. The Learning Health System concept does not create an exemption from personal-data law. Nor does describing an activity as quality improvement automatically determine its legal, ethical or research status. Where an activity constitutes research or falls within applicable research-governance requirements, additional ethical and institutional requirements may apply.
Responsible learning therefore requires both the ability to use data and the discipline to govern its use.
Viet Nam's AI framework supports lifecycle learning, but it does not mandate a Learning Health System
Several elements of Viet Nam's current AI framework are compatible with learning-system practice. For high-risk AI systems, Decree No. 142 requires lifecycle risk management, operational monitoring, human oversight and appropriate risk controls [2]. Significant changes can trigger conformity reassessment under Article 13, while Article 15 establishes continuing responsibilities for providers and deployers.
Separately, Article 19 establishes requirements for reporting and handling events that meet the statutory threshold for a serious AI incident, including preservation of relevant logs, data and information for verification, evaluation and remediation [2]. These provisions strengthen traceability and lifecycle governance.
They do not, however, create a statutory requirement for healthcare organizations to establish a Learning Health System, use a particular quality-improvement methodology or follow the Clinical AI Learning Loop proposed in this article. A healthcare organization may adopt Learning Health System principles because they strengthen its ability to learn from real-world use and improve care, not because Vietnamese AI legislation prescribes that organizational model.
The recommendations in this article should therefore be understood as healthcare quality, implementation and patient-safety practices that complement legal compliance, rather than as additional statutory obligations.
AI can accelerate learning, and it can accelerate the wrong learning
AI itself may increasingly be used to identify quality signals. The 2026 concept of continuous quality observability proposes using AI agents to review EHR narratives in the background and surface potential safety signals earlier than conventional retrospective approaches [13]. This could shorten the time between an emerging problem and organizational awareness.
The opportunity is significant, but so is the risk. If clinical documentation contains bias, AI can reproduce or amplify it. If the proxy being monitored is poorly chosen, automation can scale the wrong signal. If incident reports systematically underrepresent a particular type of harm, AI cannot reliably reconstruct incidents that were never documented. If an organization reacts to noisy or confounded signals without adequate evaluation, AI may simply accelerate ineffective improvement.
Braithwaite's 2026 analysis is therefore particularly relevant: AI can strengthen feedback loops, but outcomes still depend on implementation capability, workflow design, organizational learning and accountability [8]. Faster learning is valuable only when what is being learned is sufficiently trustworthy.
A practical Clinical AI Learning Loop
Healthcare organizations do not need to declare themselves Learning Health Systems before they begin applying learning-system principles. They can begin with a practical Clinical AI Learning Loop.
The first domain is the learning question: What do we need to know about this AI system in real-world use? Are we asking whether performance remains reliable, whether care has improved, whether staff are using the system as intended, whether one patient group is receiving less benefit, or whether a new safety problem is emerging?
The second domain is signal architecture: Which signals can answer that question? These may include model metrics, data-quality indicators, workflow measures, AI use, overrides, incidents, near misses, complaints, audit findings, patient outcomes and subgroup differences. Who owns each signal, and what should trigger review?
The third domain is real-world evidence generation: Are the data sufficiently complete and reliable? Is the denominator clear? What comparator is appropriate? Could another change explain the observed pattern? Does the organization have enough evidence to distinguish an important signal from random variation?
The fourth domain is frontline and patient feedback: What are clinicians, staff and patients observing that dashboards may not reveal? Is there a structured route for feedback? Can repeated observations be aggregated and investigated rather than remaining isolated anecdotes?
The fifth domain is multidisciplinary interpretation and decision-making: Who decides what the signal means? Depending on the issue, clinical, quality, data, technical, cybersecurity, privacy, operational and vendor expertise may be needed. Is the appropriate decision to observe, investigate further, redesign workflow, retrain staff, modify the AI, restrict use or suspend the system?
The sixth domain is implementation of improvement: What exactly will change? Who owns the action? Does the intervention address the cause identified? Does the AI itself need to change, or would changing workflow, policy, staffing, interface design, data quality or escalation be more appropriate?
The seventh domain is remeasurement and continuous assurance: After the change, did the intended measure improve? Did another problem appear? Did the benefit occur across relevant sites and patient groups? Do performance, safety, equity, transparency and security remain acceptable?
The eighth domain is organizational learning and spread: What should be retained as institutional knowledge? Which other departments or sites need to know? Should policy, training, the AI inventory, risk register, procurement requirements, monitoring thresholds or future validation criteria be updated?
The loop can be summarized as:
question → signals → evidence → interpretation → action → remeasurement → learning → spread
The Clinical AI Learning Loop is proposed here as a practical organizational framework for responsible healthcare AI implementation, healthcare quality and patient safety. It is not a statutory framework prescribed by Vietnamese law.
The learning loop needs a stop rule as well as an improvement rule
Continuous improvement is sometimes discussed as though every problem should be fixed while the technology remains in use. Sometimes the correct decision is to stop.
If repeated monitoring shows poor clinical value, unresolved inequity, unacceptable safety risk or excessive operational burden, the organization should be prepared to restrict, suspend or decommission the AI. If evidence repeatedly fails to demonstrate meaningful benefit, continued use should not be justified simply because the organization has already invested money, time and reputation in the system.
Learning requires a willingness to discover that an original assumption was wrong. That is one of the clearest differences between genuine organizational learning and technology advocacy. A Learning Health System should not be designed to prove that AI works. It should be designed to learn whether, where, for whom and under what conditions AI improves care.
What this means for healthcare organizations in Viet Nam
Clinical AI monitoring is necessary, but it is not the endpoint. The more important organizational capability is what happens after a signal appears.
Can the organization connect model monitoring with clinical outcomes? Can it learn from near misses, complaints and frontline feedback? Can real-world data become credible evidence rather than another dashboard? Can quality teams and AI-governance teams interpret signals together? Can corrective action be implemented and then tested for effectiveness? Can a lesson from one department reach another before the same problem occurs there?
These are Learning Health System questions. Viet Nam does not need every healthcare organization to build a sophisticated, research-grade Learning Health System before it can begin answering them. The practical transition can be simpler:
from monitoring AI → to learning from AI → to improving the healthcare system around AI
That is also where responsible healthcare AI becomes directly connected to healthcare quality and patient safety.
The leadership question should therefore move beyond “What does our AI monitoring dashboard show?” A more useful question is: “What did we learn from real-world use, what changed because of that learning, and did the change actually make care safer or better?”
That is where AI monitoring becomes continuous health-system improvement.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, research-ethics, quality-management, statistical, cybersecurity or data-protection advice.
The terms Learning Health System approach, learning loop, learning utility, continuous sensing, real-world data, real-world evidence, quality signals, organizational learning and Clinical AI Learning Loop are used here as practical healthcare implementation, evaluation and improvement concepts. The Clinical AI Learning Loop proposed in this article is not a statutory checklist under Vietnamese law.
A Learning Health System should not be interpreted as requiring an AI model to retrain or update itself autonomously. Organizational learning and model adaptation are different processes. Material model or system changes should remain subject to applicable change-control, validation, risk-management and regulatory requirements.
Article 19 of Decree No. 142/2026/NĐ-CP establishes a separate pathway where an event meets the statutory threshold for a serious AI incident. Internal monitoring, quality improvement, near-miss learning and organizational investigation may appropriately operate at lower thresholds and should not be confused with formal serious-incident reporting.
References to real-world evidence in this article use international evidence-generation concepts for professional education. The requirements applicable to a specific evaluation, quality-improvement activity or research project in Viet Nam depend on its purpose, methodology, data, participants and applicable legal and institutional framework.
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] 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. Issued 31 December 2025; effective 1 January 2026.
[5] 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.
[6] Ministry of Health of Viet Nam. Circular No. 19/2013/TT-BYT Providing Guidance on Management of Quality of Medical Examination and Treatment Services in Hospitals. Issued 12 July 2013; effective 15 September 2013; currently recorded as in force in the national legal database.
[7] Ministry of Health of Viet Nam. Circular No. 43/2018/TT-BYT Providing Guidance on Prevention of Medical Incidents in Medical Examination and Treatment Establishments. Issued 26 December 2018; effective 1 March 2019; currently recorded in the national legal database as partially expired.
[8] Braithwaite J. Artificial intelligence, data science and healthcare improvement: what has actually changed and has it changed enough? BMJ Innovations. Published 16 August 2026.
[9] Bian J, Afshar M, Scifres CM, et al. The bottleneck was never data or algorithms: building a learning utility for AI-enabled learning health systems. npj Health Systems. 2026;3:43.
[10] Zai AH, Adibuzzaman M, McManus DD, Walkey A. The HALO Model: A Learning Health System Framework for Artificial Intelligence. Learning Health Systems. 2026;10(3):e70100.
[11] U.S. Food and Drug Administration. Real-World Evidence and related FDA resources on real-world data and real-world evidence. 2026.
[12] Zimmerman L, Dorr DA. Governance of Artificial Intelligence in an Academic Health System after 2 Years. Studies in Health Technology and Informatics. 2026;336:784–785.
[13] Patil A, Curry WT, Nahed BV. Toward continuous quality observability in medicine. npj Health Systems. 2026;3:86.
[14] Gourishankar A, Shalhout SZ. Learning Health Systems for Chronic Disease: Embedded Artificial Intelligence Tools That Continuously Improve Care. Preventing Chronic Disease. 2026;23:260103.
[15] Steel PAD, Wardi G, Harrington RA, et al. Learning health system strategies in the AI era. npj Health Systems. 2025;2:21.
[16] Andreoletti M, Senkalfa B, Vayena E, Blasimme A. Ensuring the clinical impact of medical artificial intelligence. The Lancet Digital Health. 2026;8(7):101030.