Multimodal Clinical AI in Viet Nam: Governing Systems That Combine Text, Images, Audio, Physiological Signals and Structured Clinical Data
IMPLEMENTATION INSIGHTS
Multimodal Clinical AI in Viet Nam: Governing Systems That Combine Text, Images, Audio, Physiological Signals and Structured Clinical Data
IMPLEMENTATION INSIGHT
Last reviewed: 4 October 2026
Clinical care is already multimodal. A clinician rarely makes a decision from one isolated source of information. The patient speaks. The medical record provides history and previous notes. Laboratory values add structured measurements. Imaging contributes another layer of evidence. ECG, EEG, pulse oximetry and other physiological signals show patterns over time, while medication records, diagnostic codes and device data add still more context.
It is therefore understandable that clinical AI is moving in the same direction. Instead of analyzing only an image, only a note or only a table of laboratory values, multimodal AI can combine several forms of clinical information and use them to produce a prediction, recommendation, summary or other output.
But more modalities do not automatically mean better decisions. A system may rely much more heavily on text than on the image clinicians assume it is interpreting. One modality may be current while another is several days old. A missing modality may silently change system behavior. A radiology image, radiology report and progress note may disagree. Audio may be transcribed incorrectly. A physiological signal may contain artifact. Structured data may use a code or unit that changed after a software upgrade. If the healthcare organization cannot understand how these inputs were assembled and used, the output can appear more clinically informed while becoming harder to govern.
That is the central challenge of multimodal clinical AI: combining more information also combines more assumptions, dependencies and opportunities for failure.
Multimodal AI is broader than large multimodal models
A multimodal AI system broadly refers to an AI system that uses more than one type of input or information representation. In healthcare, this might mean combining chest radiographs with laboratory results, clinical notes with ECG signals, pathology images with molecular information, or audio from a consultation with structured clinical data.
A large multimodal model, or LMM, is more specific. WHO describes LMMs as a type of generative AI that can accept one or more types of data input, such as text or images, and generate outputs that are not limited to the same type of data provided to the model [8]. Not every multimodal clinical AI system is therefore an LMM, and not every multimodal system is generative or agentic.
This distinction matters for governance. A fixed predictive model combining imaging and laboratory values creates different risks from a multimodal generative assistant that reads an image, interprets laboratory data, retrieves medical knowledge and writes a treatment recommendation. An agent that can additionally order a test, retrieve patient information or update a clinical record introduces another layer of authority and risk.
Healthcare organizations should therefore govern the actual system architecture, intended purpose and permitted actions, not simply the label “multimodal AI.”
Clinical capability is advancing faster than the evidence for routine deployment
Multimodal clinical AI is progressing rapidly, but the evidence base remains uneven. A 2026 npj Digital Medicine scoping review examined 37 peer-reviewed studies of multimodal AI agents published between January 2022 and June 2025. Applications included clinical decision support, monitoring and health management, documentation and report generation, and medical education. Most systems, however, remained prototypes or were undergoing early technical evaluation, while evidence relating to safety, fairness, prospective outcomes and real-world clinical deployment remained limited [10].
This distinction between technical capability and clinical readiness is important. A model may perform impressively on a curated benchmark while never having been tested across the fragmented, incomplete and asynchronous information environment of an actual hospital.
A 2025 Lancet Digital Health review reached a similar conclusion from a broader implementation perspective. It identified important opportunities for multimodal AI while highlighting data heterogeneity, integration complexity, fusion methods, model selection, generalization, fairness, safety and security as major challenges for responsible adoption [9].
The practical lesson is not that healthcare organizations should avoid multimodal AI. It is that the number of modalities should not be confused with the strength of the evidence supporting the system.
More modalities can improve performance, but they can also mislead the model
It is tempting to assume that an AI system with access to more clinical information will always reason better. Recent research shows why that assumption is unsafe.
A 2026 Nature Communications study evaluated eight multimodal foundation models across 1,090 medical cases and found that image predictions were often strongly influenced by accompanying text. When misleading clinical text was introduced, performance could deteriorate substantially even when a model had previously interpreted the image correctly. In one experiment, accuracy for one evaluated model fell from 84% to 28% after a misleading clinical vignette was added [12].
That result should not be generalized mechanically to every multimodal system. It does, however, demonstrate an important governance problem: different modalities do not necessarily contribute equally, and the model may prioritize information differently from the clinician using it.
Imagine a chest image that suggests one diagnosis while the referral note strongly suggests another. A clinician may consciously recognize and investigate the discrepancy. A multimodal model may instead allow the text to dominate the image without making that weighting apparent.
Adding a modality can therefore improve performance, contribute little, or make the output worse depending on the model, task and information supplied. Healthcare organizations should ask not only “What modalities can this system use?” but also “How does its behavior change when those modalities agree, disagree, disappear or become misleading?”
Source, identity and time must remain aligned across modalities
For a single-modality model, identifying the input may be relatively straightforward. For a multimodal system, the input can become a clinical information package assembled from several systems at different times.
A single AI output might depend on a CT image acquired at 09:15; a radiology report signed at 10:40; laboratory values collected at 08:50 but resulted at 10:05; vital signs streamed continuously from a bedside monitor; a progress note copied forward from the previous day; an audio transcript generated during the current consultation; a medication list last reconciled three months ago; and structured diagnostic codes imported from another system. Technically, all of these may be “available.” Clinically, they do not have the same meaning, provenance or freshness.
A healthcare organization should therefore be able to reconstruct which clinically important modalities were available to the AI for a particular output, where they came from, when they were created or measured, what transformations occurred and whether any were missing or substituted. This extends ordinary data lineage into a multimodal environment.
Time is particularly important because clinical information does not arrive simultaneously. Suppose a patient's oxygen saturation falls sharply at 14:00. The AI receives the current waveform and vital signs, but the accompanying progress note was written at 08:00, the laboratory values are from the previous evening and the medication list has not yet been updated after a treatment change. A system may technically combine these inputs even though they describe different states of the patient.
The reverse problem occurs during model development. A later note may contain information that was not known when an earlier decision was made. If future information is inadvertently combined with earlier inputs during training or evaluation, temporal leakage can make a model appear more accurate than it could be prospectively.
Healthcare organizations should therefore treat time as part of provenance. For important inputs, it should be possible to distinguish when the clinical event occurred, when the information was recorded, when it became available to the AI and whether it remained relevant to the decision being supported. “Latest available” does not necessarily mean “appropriate for the current decision.”
Identity integrity is equally important. A radiology image belongs to a patient and study. A laboratory result belongs to a specimen and encounter. A physiological waveform comes from a device connected to a particular patient. Audio may contain several speakers. A note may have been copied forward from another encounter, and external documents may enter the record through scanning or manual upload.
Multimodal governance therefore needs to preserve alignment between patient, encounter, specimen, study, device and source document across the full pipeline. A sophisticated AI model cannot compensate for a perfectly formatted CT image attached to the wrong patient.
This becomes especially important when the system assembles context automatically. The more information an AI retrieves without a clinician explicitly selecting each item, the greater the need for reliable context binding, lineage and auditability.
Each modality has different quality risks, and modalities will often be missing
“Data quality” becomes more complicated when the underlying data are fundamentally different.
For text, important problems include missing sections, copied-forward material, ambiguous abbreviations, negation errors, translation issues, outdated notes and incorrect transcription. For medical images, quality can depend on acquisition protocol, orientation, resolution, compression, artifacts, equipment, imaging parameters and whether the image represents the intended anatomy. For audio, background noise, overlapping speakers, regional accents, microphone quality, incomplete capture and speech-to-text errors can change what information reaches the AI.
For physiological signals, electrode placement, motion artifact, sampling frequency, signal dropout, filtering, calibration and sensor malfunction may matter. For structured clinical data, units, coding systems, reference ranges, mappings, missing values, duplication, timestamps and patient matching remain important.
A multimodal system inherits these modality-specific failure modes and then adds another: a problem in one modality can change the way another modality influences the final output.
Missing modalities also need to be treated as routine conditions rather than rare edge cases. Not every patient has imaging. Not every test is ordered. ECG may be available for one encounter but not another. Audio capture can fail. A referral letter may arrive as an unreadable scan. One hospital may capture a concept as structured data while another documents it only in narrative text.
A 2025 Journal of Biomedical Informatics study was specifically developed around the fact that missing modalities are common in medical datasets and can adversely affect multimodal model performance [13]. Its proposed architecture is one technical approach rather than a universal solution, but the underlying problem is directly relevant to implementation.
Healthcare organizations should know whether the system stops when a required modality is absent, continues using the remaining information, lowers confidence, requests human review, substitutes historical information or generates an estimate for the missing input. Those behaviors should be understood before deployment rather than discovered after an incident.
If the system generates or imputes missing information, that inferred content should not silently acquire the same clinical status as information that was actually observed or measured. A reconstructed image, waveform or textual finding remains an inference.
Multimodal conflict needs an explicit governance strategy
Clinical information frequently disagrees, and disagreement is not necessarily an error. An imaging study may suggest infection while laboratory markers remain normal. The radiology report may differ from the treating clinician's interpretation. A patient's spoken history may conflict with the medication list. A device signal may suggest deterioration while the nursing note describes the patient as stable.
Humans deal with these contradictions by considering reliability, timing, context and clinical plausibility. A multimodal AI system also needs some mechanism to combine or represent conflicting evidence, but healthcare organizations should not assume that its strategy mirrors clinical reasoning.
There is no universal hierarchy in which imaging always overrides text, laboratory data always override symptoms, or the newest source always wins. The appropriate weighting depends on the clinical task and quality of the information.
A well-governed system should therefore have defined behavior when important modalities conflict. Depending on the use case, this may include displaying the discrepancy, reducing confidence, requesting additional information, triggering human review or following a predefined clinical rule.
What the system should not do is quietly merge contradictory evidence into one highly confident conclusion while hiding the disagreement from the clinician.
Multimodal AI should be tested by removing, degrading and contradicting modalities
Standard validation often evaluates a system with all expected inputs present. That is not enough for multimodal AI.
Evaluation should deliberately alter the information environment. The organization should examine what happens when imaging is missing, the image remains available but the accompanying report is removed, laboratory values arrive late, one physiological channel fails, audio quality degrades, a Vietnamese note uses unfamiliar local abbreviations, or a structured field contradicts the narrative note. Testing should also examine the effect of misleading but plausible information in one modality.
These exercises can reveal modality dependence that aggregate performance metrics hide. A model may appear robust when every input is present but deteriorate when one commonly missing modality disappears. It may appear multimodal even though almost all useful predictive signal comes from one source. Or it may become especially vulnerable when one input is confidently wrong.
Healthcare organizations should therefore ask vendors for evidence of modality-ablation, missing-modality, degraded-modality and cross-modal conflict testing, rather than relying only on headline multimodal performance.
The test should not only ask whether the system works when all expected information is present. It should ask how the system behaves when the clinical information environment becomes incomplete, asynchronous or contradictory.
Local validation in Viet Nam should test the modality mix that will actually occur
Multimodal systems are particularly sensitive to differences between development data and deployment environments because every modality creates another potential source of variation.
A healthcare organization in Viet Nam may use imaging equipment, laboratory analyzers, coding conventions, documentation workflows, monitoring devices, audio equipment and EHR structures that differ from those represented in international model development.
Language adds another layer. Clinical records may be written in Vietnamese, English or a combination of both. Vietnamese notes may contain specialty abbreviations, product names and hospital-specific terminology. Speech systems may encounter regional accents, Vietnamese-English code-switching, background noise and local patterns of clinical communication.
The local-validation question should therefore not simply be whether a product “supports Vietnamese.” It should be whether the combinations of language, devices, data sources and modalities encountered in the intended workflow behave reliably.
If a vendor validated a system using structured laboratory values, high-quality images and standardized text, but a hospital's actual workflow contains incomplete structured data, scanned external documents and highly variable narrative notes, the local information environment is materially different.
Validation should therefore include realistic modality combinations, including incomplete and imperfect ones.
Multimodality can create new fairness and access questions
Multimodal AI may improve performance by allowing one source of information to compensate for limitations in another. It can also make unequal performance more difficult to recognize.
Access to modalities is not uniform. One hospital may routinely provide advanced imaging while another does not. Some patients may have rich longitudinal records while others arrive with fragmented histories. Audio quality can differ between environments. Diagnostic testing patterns may vary by site, clinical service or pathway.
The AI may therefore effectively operate under different information conditions for different patients or sites because the available modality set is different.
The current evidence base provides limited reassurance. A 2026 systematic review of fairness in multimodal clinical decision-support studies found that fairness evaluation was uncommon. Among 88 multimodal studies in its chest X-ray and sepsis focus areas, only seven — approximately 8%, included fairness evaluations, despite many studies containing sensitive-attribute information that could potentially support such analysis [11].
This does not establish that multimodal systems are inherently unfair. It means healthcare organizations should not assume that adding modalities automatically improves equity.
Evaluation should examine whether performance changes when modalities are absent or degraded and whether modality availability or quality creates clinically important differences across relevant sites or patient groups.
Human oversight requires decision-relevant visibility into the multimodal context
Meaningful clinical oversight becomes difficult when a system presents only its final conclusion without enough information about the context behind it.
Suppose an AI recommends urgent treatment. The clinician may need to know that the system used the current CT scan, yesterday's laboratory results and an automatically transcribed conversation, but did not receive the medication order entered ten minutes earlier. That information can materially change how the recommendation should be interpreted.
Oversight therefore benefits from appropriate visibility into which important modalities were used, which were unavailable, how current they were and whether significant conflicts were detected.
This does not mean an interface should expose every model weight, attention map or fusion coefficient. Excessive technical detail can increase cognitive burden without improving clinical judgment.
The goal is decision-relevant transparency: enough information for the clinician to judge whether the AI had an adequate and coherent context for the task it was asked to perform.
A 2026 scoping review of human-AI collaboration in healthcare supports this broader principle. It found that effective collaboration depended on factors including task fit, workflow integration, training and appropriately calibrated trust, rather than on model accuracy alone [15].
Multimodal capability does not justify collecting every available data stream
One attraction of multimodal AI is its ability to process increasingly rich data. That does not mean a healthcare organization should collect every modality the technology is capable of ingesting.
Viet Nam's Law on Personal Data Protection No. 91/2025/QH15 has been effective since 1 January 2026, together with Decree No. 356/2025/NĐ-CP [4,5]. Under the current framework, health status and biometric data, as defined by law, are among categories of sensitive personal data. Decree No. 356 also includes specific requirements relevant to personal data used in AI systems and provides that AI-derived information remains subject to protection where it can identify or help identify an individual [5].
This is particularly relevant for multimodal systems because one workflow may simultaneously process clinical notes, diagnostic images, audio recordings, physiological signals and structured health information.
These modalities should not be treated as interchangeable from a data-protection perspective. A voice recording may contain clinical content, identity information and speech from other people in the room. Images and audio may also be processed for biometric purposes in some architectures. Whether an image, voice recording or other signal constitutes biometric data requires assessment of the nature of the data and how it is processed and used; in any event, the same material may still contain personal or sensitive health information.
The broader governance principle is simpler: technical ability to ingest a modality does not, by itself, justify processing it.
Healthcare organizations should understand why each modality is needed, what role it plays in the intended clinical function, where it is processed, how long it is retained, which organizations or vendors receive it and whether comparable clinical value can be achieved with less data.
Multimodal AI also frequently depends on external cloud, speech-recognition, image-processing or model services. Where an architecture involves cross-border transfer of personal data within the meaning of Vietnamese law, the organization should assess the applicable requirements under the Personal Data Protection Law and Decree No. 356, including any relevant exemptions, impact-assessment, documentation and filing obligations [5].
Circular No. 13/2025/TT-BYT, which has governed electronic-medical-record implementation since 21 July 2025 and remains in force, adds another relevant layer. EMR implementation must operate within the applicable framework covering healthcare information, data, information technology, electronic transactions, information security, cybersecurity and personal-data protection [6].
Multimodality is not itself a legal risk category in Viet Nam
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 classifies AI systems according to risk; it does not establish a separate statutory category called “multimodal AI.”
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. A system is therefore not automatically legally high-risk because it combines text, imaging, audio and structured clinical information.
Being outside the current high-risk list does not mean that a multimodal AI system falls outside Viet Nam's AI framework. Other risk-classification, transparency and governance requirements may apply depending on the nature of the system, its functions and deployment context. The high-risk list and related regulatory requirements can also evolve over time.
For high-risk AI systems, multimodal architecture can become directly relevant to regulatory risk management. Article 15 of Decree No. 142 requires providers to maintain a risk-management system and contains requirements relating to risk identification, the quality, suitability and representativeness of training, testing and evaluation data, human oversight and risk controls [2].
Article 13 is also relevant. A provider of a high-risk AI system must reassess conformity where a significant change affects the basis of the original assessment. This can include changes to data sources, principal input data types or data-processing methods where those changes materially affect system operation or risk [2].
For a multimodal high-risk system, adding a new modality, materially changing an important data source or changing the way modalities are processed or combined may therefore become relevant to conformity reassessment where the legal significant-change criteria are met.
Even where formal statutory reassessment is not required, such changes can reasonably trigger an internal clinical AI change review because they may alter the assumptions under which the system was originally validated.
Medical-device status is a separate legal question
AI risk classification under the AI Law and classification as a medical device are different legal questions.
Under Viet Nam's medical-device framework, software can fall within the definition of a medical device when it satisfies the applicable criteria and has an intended purpose relating to areas such as diagnosis, prevention, monitoring or treatment [7].
Circular No. 24/2026/TT-BYT, effective from 1 July 2026, provides current provisions concerning risk determination and management measures for products and goods that are medical devices and amends parts of the existing implementing framework [7].
A multimodal system does not become a medical device simply because it processes several modalities. Conversely, calling a system a “foundation model,” “copilot” or “clinical assistant” does not by itself place it outside medical-device regulation.
The assessment depends on the product's intended purpose, functions and applicable legal criteria. Healthcare organizations should therefore assess AI risk classification and medical-device status separately.
Vendors should explain how modalities are combined, not merely list them
Multimodal product demonstrations often emphasize capability: the system “reads imaging,” “understands clinical notes,” “processes laboratory values,” “listens to the consultation” or “integrates the full patient context.”
Those statements are not sufficient for clinical due diligence.
Healthcare organizations should understand the multimodal architecture at a level appropriate to the system's clinical risk. Does each modality have a separate processing pathway? At what stage are the inputs combined? Can the system operate when one modality is unavailable? Does it automatically generate or impute missing information? Are there minimum image-, audio- or signal-quality requirements? Can one modality dominate the output? Does the system distinguish current information from historical context? What happens when modalities conflict?
The vendor should also identify clinically important transformations that occur before information reaches the core model. Audio may first pass through speech recognition. A scanned document may pass through OCR. Laboratory codes may be mapped. Images may be compressed, resized or normalized. Physiological signals may be filtered. Structured records may be aggregated into derived features.
Those preprocessing components can affect clinical behavior even when they are supplied by another technology provider. A multimodal system can therefore have multiple failure points before the core model begins its final inference.
Post-go-live monitoring should be modality-aware
A multimodal system can deteriorate even when its headline performance metric appears stable.
A hospital may begin sending fewer images because its workflow changes. Speech-recognition accuracy may deteriorate after a microphone replacement. A new laboratory analyzer may change a structured input. One site may begin documenting more information in narrative notes and less in structured fields. An imaging service may adopt a different acquisition protocol.
Each of these changes can alter the balance of information reaching the AI without changing the core model.
Monitoring should therefore consider not only final outputs but also modality availability, quality, freshness and distribution. Useful signals may include the proportion of cases containing each expected modality, modality-specific missingness, input freshness, image- or audio-quality failures, signal dropout, mapping success, relevant modality-specific distributions, detected cross-modal conflicts and changes in the combinations of modalities available over time.
Where practical, healthcare organizations may also monitor whether model behavior or performance changes disproportionately when an important modality becomes unavailable or degraded.
A multimodal system should not be considered stable merely because its software version has not changed.
A practical Multimodal Clinical AI Governance Gate
Before deployment, a healthcare organization can begin with intended use and modality necessity. What clinical decision or workflow is the system intended to support? Which modalities are genuinely necessary for that purpose? Which are required, which are optional and which only add contextual information? Does each additional modality provide enough clinical value to justify its technical, privacy and operational complexity?
The next layer is source, identity and timing. Where does each important modality originate? Can it be linked reliably to the correct patient, encounter, specimen, study or device? Are acquisition time, documentation time and availability time distinguishable? Can the system recognize when information is historical or stale for the intended task?
A third layer is data quality and missing modalities. What quality requirements apply to text, images, audio, physiological signals and structured data? What happens when a modality is absent, incomplete, corrupted or outside specification? Does the system degrade predictably? Does it alert the user? Does it infer missing information, and if so, is the distinction between observed and inferred information preserved?
The organization should then evaluate cross-modal behavior. What happens when modalities agree? What happens when they conflict? Does one source dominate unexpectedly? Has the system been tested with modality ablation, degraded inputs, misleading information and realistic incomplete combinations? Can clinicians see important conflicts rather than only the fused conclusion?
The next layer is local validation and fairness. Has the system been evaluated with the modality combinations, language, terminology, devices, imaging protocols, audio environments and workflows expected in the Vietnamese deployment? Does performance change materially across sites or clinically relevant patient groups? Does unequal access to particular modalities create different effective system behavior across populations?
The organization also needs human oversight and workflow design. Can clinicians see which important modalities were used, missing or stale? Can they access the underlying clinical evidence where necessary? What happens when context is insufficient? Can the clinician override, defer or stop use of the output? Is the human-review burden manageable when conflict or missing information triggers escalation?
Finally comes lifecycle governance. Which changes to modalities, acquisition equipment, interfaces, preprocessing, coding, fusion methods, model versions or workflow require review or revalidation? How will modality availability and quality be monitored after go-live? Can the organization reconstruct the multimodal context behind a clinically significant output if an incident occurs?
The Multimodal Clinical AI Governance Gate is a practical organizational framework proposed for implementation and patient-safety purposes. It is not a statutory checklist prescribed by Vietnamese law.
What this means for healthcare organizations in Viet Nam
Multimodal AI has an intuitive appeal in medicine because clinicians already reason across multiple sources of information. A system that can see an image, read the medical record, interpret laboratory values, process physiological signals and understand a conversation appears closer to the way clinical decisions are actually made.
But combining information is not the same as understanding it.
The clinical value of multimodal AI depends on whether the right modalities are available, whether they belong to the right patient and encounter, whether they describe a coherent point in the patient's journey, whether their quality is sufficient, how the system responds to disagreement and whether clinicians can recognize when context is incomplete.
Current research illustrates both the opportunity and the limitation. Multimodal systems are becoming increasingly capable, including research demonstrations of conversational diagnostic reasoning using images, ECGs and clinical documents [14]. Yet the broader evidence base still contains limited prospective safety, fairness and real-world deployment evidence, while recent work shows that additional modalities can sometimes mislead rather than strengthen model performance [10–12,14].
For Viet Nam, these issues are particularly important because multimodal clinical AI will enter healthcare organizations with different EHR architectures, device ecosystems, documentation practices, patient populations, language patterns and levels of interoperability.
The leadership question should therefore move beyond “How many types of clinical data can this AI use?”
A more useful question is:
“Do we know what information the AI actually received, whether those inputs belong together, what happens when one is missing or misleading, and whether the final output remains safe when the clinical context becomes messy?”
That is where multimodal capability becomes responsible clinical implementation.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, cybersecurity, data-protection or medical-device advice.
The terms multimodal AI, large multimodal model, modality, modality fusion, modality conflict, missing modality, temporal alignment and the Multimodal Clinical AI Governance Gate are used here as practical clinical AI evaluation, implementation and governance concepts. The organizational framework proposed in this article is not a statutory checklist under Vietnamese law.
Multimodality alone does not determine an AI system's legal risk classification or medical-device status in Viet Nam. Those assessments depend on the actual system, intended purpose, functions, data processing, deployment architecture and applicable legal framework.
Article 15 of Decree No. 142/2026/NĐ-CP contains risk-management requirements applicable to providers of high-risk AI systems, while Article 13 addresses conformity reassessment following specified significant changes. These provisions should not be generalized into a legal requirement that every multimodal healthcare AI system follow the exact modality-governance framework proposed here.
Similarly, not every image, audio recording or physiological signal becomes a separate category of sensitive personal data merely because a multimodal AI system processes it. The applicable personal-data analysis depends on the nature of the information, whether and how it relates to an identifiable person, the purpose and manner of processing, and the other relevant circumstances under Vietnamese law.
References
[1] National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
[2] Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Artificial Intelligence. Issued 30 April 2026; effective 1 May 2026.
[3] Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. Issued 30 June 2026; effective 15 August 2026.
[4] National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
[5] Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Personal Data Protection. Issued 31 December 2025; effective 1 January 2026.
[6] Ministry of Health of Viet Nam. Circular No. 13/2025/TT-BYT Providing Guidance on Implementation of Electronic Medical Records. Issued 6 June 2025; effective 21 July 2025.
[7] Government of Viet Nam. Decree No. 98/2021/NĐ-CP on Management of Medical Devices, as amended; 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.
[8] World Health Organization. Ethics and Governance of Artificial Intelligence for Health: Guidance on Large Multi-Modal Models. World Health Organization; 2024.
[9] Azarfar G, Naimimohasses S, Rambhatla S, et al. Responsible adoption of multimodal artificial intelligence in health care: promises and challenges. The Lancet Digital Health. 2025;7(12):100917.
[10] Yu K, Zhou S, Hou Y, et al. Multimodal artificial intelligence agents in healthcare: a scoping review. npj Digital Medicine. Published 8 August 2026.
[11] Saha A, Shao L, Kim J, Dunn AG. Fairness in multimodal machine learning applications in clinical decision support: a systematic review. npj Digital Medicine. Published 14 July 2026.
[12] Buckley TA, Diao JA, Srivastava CN, Brodeur PG, Rajpurkar P, Rodman A, Manrai AK. Multimodal foundation models exploit text to make medical image predictions. Nature Communications. 2026;17:7475.
[13] Wang M, Fan S, Li Y, Xie Z, Chen H. Missing-modality enabled multi-modal fusion architecture for medical data. Journal of Biomedical Informatics. 2025;164:104796.
[14] Saab K, Park C, Strother T, et al. Advancing conversational diagnostic AI with multimodal reasoning. Nature Medicine. 2026;32:1726–1736.
[15] Strong J, Rogers H, Sun E, et al. Human-AI Collaboration in Healthcare: A Scoping Review. npj Digital Medicine. Published 20 June 2026.