Governing AI-Assisted Post-Stroke Dysphagia Screening in Viet Nam: Clinical Swallow Assessment, Aspiration Risk Management and Safe Oral Intake Decisions

Governing AI-Assisted Post-Stroke Dysphagia Screening in Viet Nam: Clinical Swallow Assessment, Aspiration Risk Management and Safe Oral Intake Decisions

A patient arrives at a stroke unit following an acute ischemic stroke. The patient is awake, follows simple instructions and speaks with family members, although speech is slightly slurred. During the first nursing shift, the patient asks for water. A family member assumes drinking should be safe because the patient appears alert and has not been coughing.

The hospital is evaluating an artificial intelligence (AI) system that analyzes voice recordings and clinical information to identify possible swallowing difficulties. The system assigns the patient a relatively low probability of dysphagia.

That result may appear reassuring, but it does not establish that the patient can safely drink water or take medication by mouth. Swallowing impairment can occur even when someone is alert and communicates reasonably well. Some patients also aspirate without coughing or showing obvious warning signs.

The patient still needs appropriate swallowing screening before oral intake, followed by further professional assessment when indicated. An AI prediction cannot independently authorize eating, drinking or oral medication administration.

For healthcare organizations in Viet Nam, the important question is whether AI can help identify swallowing problems earlier without creating false reassurance, delaying necessary assessment or imposing avoidable restrictions on nutrition and hydration.

The answer depends on more than a model's ability to recognize abnormal voice patterns or clinical risk factors. It depends on whether the hospital can connect early screening with professional swallowing assessment, aspiration risk management, appropriate nutrition and medication plans, and reassessment as the patient recovers.

Why Swallowing Safety Matters After Stroke

Swallowing requires precise coordination among the muscles and structures of the mouth, pharynx, larynx and esophagus, together with neurological mechanisms that protect the airway. A stroke can disrupt this coordination, affecting the patient's ability to control food or liquid, initiate swallowing, transport material through the pharynx or prevent it from entering the airway.

Post-stroke dysphagia is common, particularly during the acute phase. Its reported prevalence varies substantially with stroke severity, assessment methods, timing and patient characteristics.

The consequences extend beyond coughing or choking during meals. Dysphagia may contribute to aspiration, respiratory complications, dehydration, malnutrition, difficulty administering medication and prolonged hospitalization. It can also affect rehabilitation, social interaction and quality of life.

Aspiration occurs when material enters the airway below the vocal folds. The material may include food, liquid, saliva or other secretions. However, aspiration does not always lead to pneumonia, and pneumonia after stroke can develop through several interacting mechanisms.

Swallowing function, airway protection, oral hygiene, mobility, respiratory health, immune status and the patient's overall condition all influence the likelihood of aspiration-related complications.

Silent aspiration deserves particular attention. A patient may aspirate without coughing, choking or displaying immediate respiratory distress. Someone who looks comfortable and speaks normally may still have clinically important swallowing impairment.

This creates competing risks for the clinical team. Missing dysphagia may expose the patient to preventable harm, while unnecessarily prolonged restrictions on eating or drinking may compromise nutrition, hydration, medication delivery and recovery.

The objective of dysphagia management is therefore not to classify every patient as safe or unsafe using a single measurement. It is to understand swallowing function sufficiently well to support appropriate decisions about nutrition, hydration, medication administration and rehabilitation.

AI may contribute information to that process, but its clinical role must be clearly defined.

The Clinical Framework Already Available in Viet Nam

Vietnamese healthcare organizations have national professional guidance relevant to swallowing disorders after stroke.

Ministry of Health Decision No. 3312/QD-BYT, dated 5 November 2024, issued Diagnosis and Treatment of Stroke (Hướng dẫn chẩn đoán và điều trị đột quỵ não). The Decision took effect on the date of issuance and replaced Decision No. 5331/QD-BYT dated 23 December 2020.

The 2024 guidance addresses stroke diagnosis, treatment, rehabilitation and prevention of complications. Its swallowing-related recommendations support early screening in acute stroke, assessment before eating, drinking or taking oral medication, and further investigation when aspiration is suspected.

The guidance recognizes that trained healthcare professionals, including speech-language therapy personnel, may participate in swallowing screening. It also addresses instrumental evaluation, oral hygiene, nutritional support and rehabilitation.

These recommendations provide an important foundation for hospital swallowing safety pathways.

A further development was Ministry of Health Decision No. 3554/QD-BYT, dated 14 November 2025, issuing Guidance on Rehabilitation Technical Procedures, Volume 2 (Hướng dẫn quy trình kỹ thuật về Phục hồi chức năng, Tập 2).

The Decision took effect on 1 July 2026 and introduced 169 rehabilitation technical procedures, replacing several earlier decisions concerning rehabilitation techniques.

The document includes procedures for bedside dysphagia assessment, assessment using modified food textures and thickened liquids, Flexible Endoscopic Evaluation of Swallowing (FEES), and videofluoroscopic swallowing assessment.

The bedside assessment procedure recognizes the participation of rehabilitation physicians, speech-language therapy personnel, rehabilitation professionals and nurses. It specifies patient readiness requirements and identifies circumstances in which direct swallowing trials should not proceed.

These details matter because a requirement to assess swallowing must not be interpreted as permission to offer water or food to every patient regardless of clinical condition.

A patient who cannot maintain adequate alertness, protect the airway or participate safely requires an appropriate alternative assessment and care pathway.

The two Ministry of Health documents serve complementary purposes. Decision No. 3312/QD-BYT establishes the broader stroke care framework, while Decision No. 3554/QD-BYT provides more detailed rehabilitation technical procedures.

Neither establishes a requirement to use artificial intelligence for swallowing screening.

Hospitals considering AI should evaluate the technology within existing professional arrangements rather than treat it as a replacement for established clinical responsibilities.

Interpreting the 2026 AHA/ASA Guidelines

Two major American Heart Association and American Stroke Association (AHA/ASA) guidelines published in 2026 provide additional evidence for swallowing assessment and management after stroke.

The 2026 Guideline for the Early Management of Patients With Acute Ischemic Stroke, initially published online in January 2026, recommends bedside swallowing screening before liquid or food intake to identify patients at increased risk of aspiration.

The guideline recognizes that screening may be performed by speech-language pathologists or other appropriately trained healthcare professionals.

For patients who fail bedside screening or cannot participate because of neurological impairment, instrumental evaluation using FEES or a Videofluoroscopic Swallowing Study (VFSS) may be appropriate to assess dysphagia severity and aspiration risk.

An August 2026 correction to the guideline clarified the recommendation concerning instrumental assessment, explicitly recognizing both FEES and VFSS. The corrected recommendations should be used when reviewing the guideline for clinical implementation.

The acute stroke guideline also recognizes that an oral hygiene protocol may be reasonable to reduce pneumonia risk.

The second document, the 2026 Guideline for Adult Stroke Rehabilitation and Recovery, published online in August 2026, provides recommendations extending across the recovery pathway.

It recommends dysphagia screening as soon as possible, within 24 hours after admission for acute stroke, and before the patient begins eating, drinking or receiving oral medication.

The guideline recommends formal, validated screening methods that include observation of swallowing attempts with different liquid volumes.

However, this recommendation must be interpreted together with the patient's readiness and the safety criteria of the selected screening protocol. It does not mean that a water swallowing challenge should be performed in someone who cannot participate safely.

For patients with suspected dysphagia, the rehabilitation guideline recommends instrumented swallowing studies rather than bedside assessment alone when the clinical objective is to verify aspiration, understand swallowing biomechanics, determine the least restrictive appropriate oral diet or guide treatment more precisely.

It also recognizes the importance of evaluating pill swallowing when an instrumental swallowing study is performed.

The guideline supports repeat dysphagia screening following admission to a new care setting and comprehensive reassessment during recovery. For patients with persistent post-stroke dysphagia, repeated comprehensive swallowing assessment within three months of discharge from acute care is considered reasonable.

These recommendations help connect swallowing safety during acute hospitalization with rehabilitation and transitions of care.

Some aspects of the acute stroke guideline's dysphagia section have also prompted professional discussion. In October 2026, the American Board of Swallowing and Swallowing Disorders raised concerns about the distinction between dysphagia screening and clinical swallowing evaluation, as well as the interpretation of evidence concerning pharyngeal electrical stimulation.

These concerns do not invalidate the broader importance of early screening and appropriate swallowing assessment. They do, however, reinforce the need to distinguish clinical activities and avoid treating every emerging intervention as an established standard of care.

The guidelines must also be interpreted according to their intended patient populations. Recommendations for acute ischemic stroke should not automatically be extended to every form of stroke, while the adult rehabilitation guideline should not be applied without qualification to pediatric patients.

For Vietnamese hospitals, AHA/ASA recommendations provide contemporary evidence that can inform locally approved procedures alongside applicable Ministry of Health guidance.

Screening, Clinical Assessment and Instrumental Evaluation Are Different Activities

Safe dysphagia management requires a clear distinction between screening, clinical swallowing assessment and instrumental evaluation.

Swallowing screening is intended to identify patients who may have swallowing impairment and require further assessment. It should be practical enough to support timely decisions within a defined clinical pathway.

A validated bedside protocol may examine alertness, neurological findings, secretion management, communication, oral motor function and swallowing performance.

Some screening methods include controlled swallowing trials using specified liquid volumes. These trials are appropriate only when the patient meets the protocol's readiness and safety criteria.

A patient who is unconscious, cannot maintain adequate alertness, has significant respiratory instability or cannot safely participate requires an alternative clinical approach.

The purpose of screening is to identify swallowing problems without exposing patients to unnecessary risk.

Clinical swallowing assessment provides a more detailed evaluation by an appropriately qualified professional. It may include examination of oral motor function, secretion management, cough effectiveness, voice characteristics and swallowing behavior.

These findings help guide the immediate care plan and determine whether instrumental evaluation is indicated.

Instrumental swallowing assessment provides information that may not be available from bedside observation alone.

FEES uses a flexible endoscope to visualize relevant pharyngeal and laryngeal structures, secretions and swallowing events. It can provide information about airway protection and swallowing abnormalities in appropriately selected patients.

VFSS uses dynamic X-ray imaging with suitable contrast material to examine bolus movement, airway invasion and swallowing biomechanics.

Each technique has advantages and limitations. The choice depends on the clinical question, the patient's condition, available expertise, equipment and feasibility.

Neither method is universally superior for every patient or every clinical situation.

Instrumental studies are particularly valuable when findings are inconclusive, silent aspiration is suspected or more detailed information is needed to guide diet selection and rehabilitation.

AI-assisted screening should not collapse these distinct activities into one automated determination. A dysphagia risk prediction is not equivalent to a clinical swallowing assessment, and neither is a substitute for instrumental evaluation when that evaluation is clinically indicated.

Silent Aspiration and the Danger of False Reassurance

Silent aspiration is one of the strongest reasons to avoid overinterpreting a negative swallowing screen.

A patient may aspirate liquid, food or secretions without producing an obvious cough, choking episode or immediately recognizable change in voice.

Traditional bedside screening methods can miss these events because they often depend partly on observable signs.

AI systems based on voice recordings or swallowing sounds may encounter similar limitations.

For example, a model might classify aspiration risk from acoustic changes after swallowing. Yet a patient who aspirates silently may not exhibit the voice characteristics that the model was trained to recognize.

Conversely, abnormal voice quality after stroke may result from dysarthria, vocal fold dysfunction or other neurological changes unrelated to aspiration.

A voice-based system can therefore produce both false-negative and false-positive results.

A false-negative result may create inappropriate reassurance, leading to premature oral intake or delayed specialist evaluation. A false-positive result may lead to unnecessary restrictions, additional investigations or prolonged non-oral feeding in a patient who could otherwise eat safely.

Both types of error matter.

A negative result from a validated clinical screening method may contribute to an oral intake decision when the complete protocol has been appropriately performed and the patient's condition supports the decision.

However, no bedside screening result provides an absolute guarantee that aspiration cannot occur.

An unvalidated AI-generated negative prediction cannot independently authorize oral intake.

Where clinical suspicion persists despite a reassuring result, further professional assessment should remain available.

The objective is to manage uncertainty safely, using appropriate clinical evidence and reassessment rather than treating one test result as a definitive guarantee.

Where AI May Contribute to Dysphagia Care

AI research in dysphagia assessment includes several distinct applications.

Some models analyze clinical variables such as age, stroke severity and neurological findings to estimate the likelihood of swallowing impairment.

Others use voice recordings, acoustic swallowing signals or measurements collected through specialized sensors.

Computer vision and machine-learning techniques are also being investigated to support interpretation of instrumental swallowing studies.

For example, AI may help identify penetration, aspiration or pharyngeal residue in VFSS recordings. Other applications may assist clinicians with selected measurements or findings from FEES.

These functions are not interchangeable.

A model developed using instrumental swallowing images should not automatically be used to classify patients from ordinary smartphone recordings.

Similarly, a system trained on dysphagia associated with multiple neurological or head and neck conditions may not perform equally well among patients with acute stroke.

The intended medical purpose, data source, clinical population and expected action need to be specified before implementation.

A relatively straightforward initial application may involve identifying patients whose required swallowing screening has not been documented or organizing relevant clinical information for professional review.

Such functions can sometimes be achieved through conventional electronic records and programmed rules without AI.

More advanced applications, such as acoustic-based classification or automated interpretation of instrumental studies, require validation appropriate to their intended medical purpose.

Evidence requirements become particularly important when an AI result may influence further assessment, diet restrictions or permission to begin oral intake.

The hospital must determine whether the system provides a clinically meaningful improvement beyond existing screening procedures and simpler digital solutions.

What the 2026 Evidence Shows About AI-Assisted Screening

Research published in 2026 provides encouraging evidence of technical progress, but it also illustrates important limitations in clinical validation.

A scoping review published in Annals of Biomedical Engineering in July 2026 examined artificial intelligence and data science applications in post-stroke dysphagia screening and assessment.

De Andrade and colleagues reviewed evidence concerning conventional bedside screening, acoustic voice analysis and computational approaches for dysphagia and aspiration risk assessment.

Some acoustic-based models reported strong discrimination, including area-under-the-curve values above 0.90 for selected tasks.

However, much of the evidence came from relatively small datasets, single-center populations and internal validation.

The authors identified overfitting, inconsistent outcome definitions, limited external validation and differences between clinical environments as barriers to adoption.

The review also highlighted the limitations of individual voice characteristics as markers of aspiration.

Wet voice following swallowing, considered alone, has relatively low sensitivity. Incorporating this signal into a more sophisticated algorithm does not automatically eliminate the underlying limitation.

The review called for further evaluation of multimodal approaches, culturally and linguistically adapted models and integration into real clinical workflows.

Another scoping review, published in BMJ Open in July 2026, examined 27 studies involving 8,842 participants.

The studies covered AI applications using VFSS, FEES, acoustic information, wearable sensors and structured clinical data.

Most focused on screening and assessment. Only one included randomized controlled trial evaluated an AI-based dysphagia intervention.

Several studies reported promising technical accuracy for selected swallowing abnormalities, but evidence concerning external validation, routine clinical implementation and patient-centered outcomes remained limited.

Importantly, the review included dysphagia arising from several clinical conditions rather than stroke alone.

Its findings should not therefore be interpreted as direct evidence that all included technologies are suitable for acute stroke care.

The review's literature search covered studies published through May 2025, even though the review itself appeared in 2026. Its conclusions must be understood in that context.

Taken together, the available reviews support continued development and evaluation of AI-assisted dysphagia assessment. They do not establish that routine deployment improves swallowing recovery, reduces aspiration pneumonia or makes oral intake decisions safer in Vietnamese hospitals.

A September 2026 Study Illustrates Why the Reference Standard Matters

A study published in PLOS Digital Health on 29 September 2026 examined machine-learning methods for screening oropharyngeal dysphagia.

Rech and colleagues analyzed a final dataset of 465 adults and older adults, including 153 classified as having dysphagia.

The researchers evaluated several algorithms using a proposed 15-item clinical screening model.

Among the better-performing approaches, accuracy ranged from approximately 76% to 79%, while sensitivity ranged from approximately 74% to 84%.

These findings suggest that machine learning can identify patterns associated with dysphagia.

However, the study was not limited to patients with acute stroke. Its results should therefore not be presented as validated performance estimates for an acute stroke unit.

There was also an important distinction in how the reference diagnosis was established.

Participants with relevant clinical indications underwent videofluoroscopic swallowing assessment. Those without swallowing complaints, clinical signs or another indication for instrumental assessment could be classified as not having dysphagia based on the available clinical findings.

Consequently, not every participant underwent identical confirmation using VFSS.

This distinction matters when interpreting the strength of the reference standard and the possibility that some swallowing abnormalities were not detected.

The study assessed classification performance rather than the clinical impact of deploying the algorithms.

It did not demonstrate that AI-assisted screening reduced aspiration pneumonia, improved nutrition or shortened hospitalization.

Nor did it establish that a negative model result could safely authorize oral intake without an appropriate clinical assessment.

The findings support additional research, particularly through external validation and evaluation in clearly defined patient populations, rather than immediate adoption as an independent swallowing safety decision tool.

AI-Assisted Rehabilitation Is a Separate Clinical Application

AI research in dysphagia extends beyond screening and diagnosis into rehabilitation.

A systematic review published in npj Digital Medicine in June 2026 examined human–AI collaboration in swallowing rehabilitation.

Yang and colleagues reviewed 31 studies involving 1,012 participants, including adults with oropharyngeal dysphagia and healthy volunteers evaluating technologies intended for clinical application.

The review found evidence of short-term improvements in selected functional and physiological outcomes with AI-supported rehabilitation approaches.

However, sustained benefits and adherence after supervised interventions were less consistently demonstrated. Implementation challenges included digital literacy, cognitive impairment, interface usability and the need for ongoing clinical support.

These findings are relevant to stroke rehabilitation, where patients may have cognitive, communication or physical impairments affecting their ability to use digital technologies.

Nevertheless, evidence supporting AI-assisted rehabilitation must not be confused with evidence supporting AI-assisted screening.

A technology that helps deliver swallowing exercises does not automatically establish that a patient can safely begin drinking water.

Likewise, improvement in rehabilitation outcomes does not validate an unrelated AI model for aspiration risk classification.

Each application requires evaluation according to its intended clinical purpose, level of professional supervision and consequences for patient care.

Safe Oral Intake Requires an Individualized Clinical Plan

Following screening and appropriate assessment, the clinical team must determine how the patient can receive nutrition, hydration and medication safely.

Some patients may begin oral intake without modification. Others require adjustments to food texture, liquid consistency, positioning, assistance or swallowing strategies.

Patients who cannot swallow safely may temporarily require alternative routes for nutrition, hydration and medication.

The appropriate plan depends on swallowing function, neurological status, nutritional needs, medical conditions and treatment goals.

A screening result alone should not be treated as a complete dietary prescription.

The International Dysphagia Diet Standardisation Initiative (IDDSI) provides a framework for describing food textures and drink thickness using standardized terminology.

Consistent terminology can support communication among clinicians, speech-language therapy personnel, nurses, dietitians, food service staff, patients and caregivers.

However, IDDSI describes texture and consistency levels. It does not independently determine which level is appropriate for a particular patient.

That decision requires professional assessment.

Thickened liquids should not be introduced automatically simply because dysphagia is suspected or identified.

They may improve selected measures of swallowing safety for some patients, but can also affect acceptance, fluid intake, hydration and quality of life. Reducing aspiration observed during an assessment does not necessarily establish that a modified liquid consistency will prevent pneumonia.

Similarly, an unnecessarily restrictive diet may interfere with nutrition and recovery.

The care team should aim for the least restrictive clinically appropriate plan, balancing swallowing safety with nutrition, hydration and patient preferences.

Where oral intake remains unsafe, the hospital needs a timely plan for alternative nutrition and hydration.

The AHA/ASA 2026 acute stroke guideline addresses enteral feeding, including the use of nasogastric feeding during the early period and consideration of gastrostomy when a prolonged inability to swallow safely is anticipated.

These decisions must remain individualized.

Feeding tubes do not eliminate every risk of aspiration, including aspiration of saliva or other secretions.

Oral hygiene, clinical monitoring and appropriate rehabilitation remain important regardless of the feeding route.

Medication Administration Requires a Separate Safety Decision

Medication administration is an important aspect of dysphagia management that can be overlooked when attention focuses primarily on food and liquids.

A patient may tolerate certain food textures but still have difficulty swallowing tablets or capsules.

Likewise, someone who cannot safely drink thin liquids may require another approach to medication administration.

Patients recovering from stroke frequently need time-sensitive medicines for cardiovascular disease, secondary prevention and other medical conditions.

A swallowing restriction should therefore trigger an appropriate medication plan rather than simply delay essential treatment.

Crushing tablets or opening capsules is not automatically safe.

Some medicines have modified-release formulations, protective coatings or other characteristics that make alteration inappropriate.

Changing the formulation may affect absorption, effectiveness or toxicity.

The prescriber and pharmacist should be involved whenever the formulation or route of administration requires modification.

Depending on the medicine and clinical circumstances, suitable alternatives may include different oral formulations, appropriate non-oral routes or administration through an enteral feeding tube when compatible.

Medication should not automatically be mixed into thickened liquids or modified foods without assessing the specific product and patient's needs.

An AI system should not independently recommend crushing medication merely because it has identified a swallowing problem.

Medication safety requires verification of the drug, formulation, dose, route and clinical consequences of any proposed change.

This reinforces the importance of swallowing screening before oral medication administration during acute stroke care.

The safety of eating, drinking and taking medication are related, but they require distinct clinical considerations.

Aspiration Prevention Requires More Than Withholding Food and Drink

Reducing aspiration-related complications requires attention to the patient's overall condition, not simply food texture or temporary restrictions on oral intake.

Appropriate oral hygiene is one important component.

Oral bacteria and secretions can contribute to pneumonia risk when aspirated, particularly in vulnerable patients.

The AHA/ASA 2026 acute ischemic stroke guideline recognizes that an oral hygiene protocol may be reasonable for reducing pneumonia risk.

However, no single oral care method has been shown to eliminate aspiration-related pneumonia.

Positioning, assistance during meals, appropriate feeding pace and adherence to an individualized swallowing plan may also contribute to safer care.

Patients need support appropriate to their ability to maintain posture, follow instructions and coordinate swallowing.

Swallowing function can change during hospitalization.

A patient who previously tolerated oral intake may develop new difficulties after neurological deterioration, respiratory illness, reduced alertness or another acute event.

The same concern applies after extubation and during transfers between care settings.

An earlier screening result should not be treated as permanently valid when relevant clinical circumstances have changed.

Withholding oral intake also does not eliminate aspiration pneumonia risk. Some patients continue to aspirate secretions, while the likelihood of pneumonia depends on several additional factors.

Prolonged nil-by-mouth status should therefore not be presented as a complete preventive strategy.

The objective is to manage aspiration risk while maintaining appropriate nutrition, hydration, essential treatment and rehabilitation.

Professional Responsibilities Across the Stroke Care Team

Post-stroke dysphagia management requires coordinated contributions from several professional groups.

Stroke physicians and other authorized clinicians assess neurological and medical conditions, determine the need for further investigation and approve appropriate treatment plans.

Nurses play an important role in recognizing swallowing concerns, performing screening within their training and scope of practice, observing patients during meals and medication administration, and communicating changes promptly.

Speech-language therapy personnel contribute specialized swallowing assessment, rehabilitation planning and recommendations concerning swallowing strategies and oral intake.

Rehabilitation physicians, dietitians, pharmacists and other professionals may be involved according to the patient's needs.

The hospital must coordinate these responsibilities.

Some facilities may have trained nurses performing initial screening, with specialized assessment provided by rehabilitation or speech-language therapy services.

Others may organize the process differently because of staffing, expertise and available resources.

The essential requirement is that responsibilities are clearly defined and consistent with applicable professional standards and local procedures.

The organization also needs a pathway for abnormal, inconclusive or incomplete screening results.

A patient should not remain indefinitely without a nutrition, hydration or medication plan because specialist assessment is delayed.

At the same time, a failed screen should not automatically lead to an invasive procedure without the necessary clinical evaluation.

Interim care arrangements, referral responsibilities and reassessment plans should be documented.

Family members and caregivers can contribute information about previous swallowing difficulties, food preferences and assistance needs.

Their observations should complement qualified assessment rather than replace it.

Clear communication is especially important when caregivers participate in feeding or medication support.

Where AI May Improve the Clinical Workflow

The value of AI depends on the problem the hospital is trying to solve.

One organization may struggle to ensure that every eligible stroke patient receives swallowing screening before oral intake. Another may perform screening consistently but experience delays in obtaining specialist assessment after abnormal findings. A third may have FEES or VFSS available but struggle to coordinate reassessment when patients transfer to rehabilitation services.

These problems require different responses.

Missing screening documentation may be addressed through structured electronic records or rule-based reminders without a complex AI model.

AI may offer additional value when relevant information is difficult to retrieve from narrative records or when validated models identify clinically important patterns not readily recognized through simpler methods.

Acoustic-based analysis may eventually support selected screening tasks in defined populations, while computer vision may assist with measurements from instrumental swallowing studies.

However, these applications require independent validation and appropriate clinical oversight.

A prediction model may offer limited additional benefit if trained staff already screen patients promptly and follow a reliable referral pathway.

Conversely, adding more risk alerts may worsen delays if the specialist service cannot accommodate additional assessments.

The implementation plan should therefore account for downstream clinical capacity.

Technology should be evaluated for its incremental contribution beyond established screening methods, existing electronic systems and improvements in professional coordination.

Local Validation Must Address Clinical Consequences

An AI-assisted dysphagia system should undergo appropriate evaluation before routine clinical use.

The intended patient population must be defined.

A model developed using adults with various neurological or head and neck conditions should not automatically be applied to acute stroke patients.

Similarly, a model validated in stable rehabilitation patients may behave differently among people with severe acute neurological impairment.

The clinical task must also be clear.

Predicting dysphagia, identifying aspiration, analyzing voice characteristics and interpreting instrumental studies require different evidence and reference standards.

For dysphagia screening, an appropriate reference assessment may involve qualified clinical evaluation and instrumental investigations where indicated.

For models intended to identify aspiration, the reference assessment must be suitable for determining whether aspiration occurs. FEES or VFSS may provide appropriate reference evidence according to the clinical question and study design.

The absence of coughing or an apparently normal voice should not serve as the sole reference standard.

Local validation should consider silent aspiration, dysarthria, aphasia, cognitive impairment and other characteristics that may affect performance.

Acoustic systems require additional technical evaluation.

Background noise, microphone characteristics, recording distance, voice quality and language differences may influence results.

Vietnamese-language implementation deserves specific consideration when models analyze speech or phonation.

A system trained using English-language speech recordings should not automatically be assumed to perform equally well with Vietnamese speakers.

Hospitals should examine whether changes in recording conditions alter the classification of otherwise similar clinical information.

A system that produces inconsistent results because of microphone placement or environmental noise may not be suitable for clinically consequential decisions.

Performance evaluation should examine sensitivity, specificity, predictive values, calibration where appropriate and the clinical consequences of errors.

False-negative findings are particularly important because they may lead to missed assessment or inappropriate reassurance.

False-positive results may create unnecessary restrictions, investigations and clinical workload.

Decision thresholds should be determined through clinical evaluation rather than selected solely because they produced favorable results in the development dataset.

The quality and consistency of reference assessments also matter.

A hospital should examine whether relevant patient groups received appropriate verification and whether the reference outcome was established through methods suitable for the intended clinical task.

Technical performance cannot be interpreted independently of the data and reference standards used to generate it.

Measuring Whether Screening Improves Patient Care

Successful AI-assisted dysphagia implementation cannot be demonstrated merely by counting the number of patients assigned a risk score.

The first question is whether eligible stroke patients receive appropriate swallowing screening before oral intake.

The hospital should define the population included in this measure and distinguish completed screening from attempts that were stopped or could not be performed.

Patients unable to participate safely require an appropriate alternative clinical pathway rather than being treated as simply having missing screening data.

The organization should also evaluate whether abnormal or inconclusive results lead to timely professional assessment.

Where specialist referral is needed, the hospital should monitor whether the assessment is completed and whether the findings result in an appropriate care plan.

The next consideration is whether swallowing safety decisions are implemented reliably.

Patients should have documented nutrition and hydration plans, appropriate medication arrangements and clear communication of relevant instructions to nurses, pharmacists, dietitians and caregivers.

An appropriate temporary oral intake restriction should be distinguished from an unnecessary delay that leaves a patient without adequate nutrition, hydration or essential medication.

Clinical outcomes may include aspiration-related respiratory complications, pneumonia, nutritional and hydration status, duration of non-oral feeding, swallowing recovery and appropriate return to oral intake.

These outcomes require careful interpretation.

Pneumonia after stroke can occur through several mechanisms and should not automatically be attributed to a missed swallowing problem.

Similarly, the absence of pneumonia does not prove that swallowing assessment or aspiration prevention was adequate.

Changes in diagnostic testing, patient characteristics and documentation may influence recorded outcomes.

Balancing measures should include unnecessary nil-by-mouth duration, dehydration, inadequate nutrition, inappropriate diet restrictions, delayed medication administration, avoidable referrals and additional workload.

For AI-assisted systems, evaluation should also examine false-positive and false-negative classifications, professional overrides and instances in which AI outputs influenced care inappropriately.

The most meaningful assessment follows the patient from initial screening through professional evaluation, care planning, implementation and subsequent outcomes.

The objective is to determine whether the complete intervention produces safer and more effective care than the hospital's previous process.

Reassessment During Recovery and Transitions of Care

Swallowing function can change considerably during recovery from stroke.

Some patients improve rapidly over the first several days, while others require prolonged rehabilitation or experience persistent dysphagia.

An early decision about feeding may therefore need revision.

A patient who initially requires non-oral nutrition may later tolerate selected foods or liquids. Another may need further assessment before progressing to a less restrictive diet.

The AHA/ASA 2026 rehabilitation guideline supports reassessment at relevant transitions and during recovery.

For patients with persistent post-stroke dysphagia, comprehensive swallowing reassessment within three months after discharge from acute care is considered reasonable.

Reassessment should reflect neurological recovery, clinical condition and functional progress.

It should not consist simply of repeating the same screening test without considering whether more detailed clinical or instrumental evaluation is required.

Communication between care settings is essential.

When a patient transfers from an acute stroke unit to a rehabilitation facility, the receiving team should have access to swallowing assessment findings, current dietary recommendations, medication arrangements and outstanding assessment needs.

Information about previous FEES or VFSS investigations should be available when relevant.

Patients returning home also require understandable instructions.

Caregivers should know the recommended food and drink consistencies, assistance requirements, medication arrangements and changes that warrant further medical review.

AI may eventually help identify overdue follow-up assessments or incomplete handovers, but this would represent a separate application requiring evaluation.

Generating a reminder does not establish that the appropriate clinical action has occurred.

Protecting Patient Information and Voice Recordings

AI-assisted dysphagia screening may involve sensitive clinical information, including stroke diagnoses, neurological findings, swallowing assessments, images, videos and voice recordings.

These records may identify patients directly or indirectly.

Voice recordings deserve particular attention because they can contain identifying characteristics and information that may be reused for purposes unrelated to the original clinical assessment.

However, not every voice recording should automatically be classified as biometric identification data. Its legal characterization depends on the information processed, the purpose of processing and applicable statutory definitions.

Healthcare organizations should determine whether an AI application genuinely requires identifiable audio or video and whether its purpose can be achieved with less intrusive information.

Where recordings are necessary, their collection purpose, access restrictions, security measures and retention arrangements should be defined.

Patients should receive appropriate information about the technology, including applicable disclosure and consent requirements.

Clinical assessment must also be distinguished from secondary uses of patient information.

A voice recording collected to support swallowing evaluation should not automatically become available for unrelated model training or commercial development.

Where technology suppliers process or store patient information, responsibilities for hosting, access, maintenance, subprocessors, security and incident management should be addressed.

External cloud services, remote access and cross-border processing may create additional legal obligations.

These arrangements should be assessed before identifiable patient information is transferred to an external AI system.

Legal and Professional Responsibilities in Viet Nam

AI-assisted post-stroke dysphagia screening must operate within Viet Nam's applicable healthcare, artificial intelligence, medical device and personal data protection frameworks.

The Law on Medical Examination and Treatment No. 15/2023/QH15, effective from 1 January 2024, as amended, establishes the broader framework for healthcare practice, patient rights, medical records and organizational responsibilities.

Ministry of Health Decision No. 3312/QD-BYT, dated 5 November 2024, provides relevant national professional guidance concerning stroke diagnosis, treatment and rehabilitation.

Decision No. 3554/QD-BYT, dated 14 November 2025 and effective from 1 July 2026, provides rehabilitation technical procedures, including swallowing assessment.

These documents should inform locally approved care pathways, taking account of the qualifications, scope of practice and professional responsibilities of the healthcare personnel involved.

Ministry of Health Circular No. 31/2021/TT-BYT on nursing activities in hospitals is relevant to assessment, individualized care, monitoring and clinical communication.

Circular No. 43/2018/TT-BYT addresses prevention and management of medical incidents, while Circular No. 19/2013/TT-BYT provides a framework for hospital quality management.

These instruments support existing clinical and organizational responsibilities. They do not establish a mandatory AI-assisted dysphagia screening procedure.

Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on 1 March 2026.

Decree No. 142/2026/ND-CP, effective from 1 May 2026, provides implementing provisions, while Decision No. 33/2026/QD-TTg, effective from 15 August 2026, establishes the list of high-risk AI systems under the relevant framework.

The legal classification of an AI-assisted dysphagia application should not be determined solely by its use in a healthcare facility.

Assessment requires consideration of the intended purpose, actual functions, degree of automation, deployment circumstances and applicable legal criteria.

A system organizing clinical information for professional review may raise different regulatory considerations from software intended to classify aspiration risk or influence oral intake decisions.

Clinical risk must also be assessed independently.

An AI system that incorrectly reassures staff about swallowing safety may create significant patient-safety concerns regardless of whether it falls within a particular statutory high-risk category.

Healthcare organizations should separately determine whether a product falls within Viet Nam's applicable medical device regulatory framework.

Relevant provisions may include Decree No. 98/2021/ND-CP on medical device management, as amended, and Ministry of Health Circular No. 24/2026/TT-BYT, effective from 1 July 2026, concerning medical device risk determination and related regulatory requirements.

The intended medical purpose and actual functions of the product are relevant to that assessment.

Personal data protection must also be considered.

The Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/ND-CP, both effective from 1 January 2026, provide relevant requirements concerning processing and protection of personal information.

Healthcare organizations should establish the applicable lawful basis and defined purposes for processing, appropriate access controls, security measures, retention arrangements and responsibilities for external technology providers.

Secondary use of clinical records or voice recordings for AI development, validation, retraining or commercial improvement requires assessment of the applicable legal grounds and safeguards.

Cross-border processing and external hosting may also require additional legal review and documentation.

International clinical guidelines, including AHA/ASA and European Stroke Organisation–European Society for Swallowing Disorders recommendations, can inform professional practice but do not automatically establish binding Vietnamese legal obligations.

Similarly, the organizational AI governance measures discussed in this article should be distinguished from specific statutory requirements unless an applicable provision expressly establishes them.

A Practical Starting Point for Vietnamese Hospitals

A hospital considering AI-assisted post-stroke dysphagia screening should begin by reviewing the reliability of its existing swallowing safety process.

A focused examination of recent stroke admissions may reveal whether patients received swallowing screening before oral intake, whether abnormal findings led to professional assessment and whether appropriate nutrition and medication plans were established.

The review should include patients whose screening could not be completed.

These patients may be especially vulnerable to delays because they require an alternative pathway while essential treatment, hydration and nutrition are addressed.

Stroke physicians, nurses, rehabilitation professionals, speech-language therapy personnel, dietitians, pharmacists and quality management staff should contribute according to the hospital's organization.

The findings may reveal that the main problems involve inconsistent screening, limited specialist availability or poor communication between departments.

These problems may be addressed through training, locally approved procedures, structured documentation and better coordination without introducing a complex AI model.

Where AI is expected to add value, the hospital should define a narrow use case.

One option may be evaluating whether AI-assisted review of clinical records improves the identification of patients with incomplete swallowing assessment documentation.

Another may involve acoustic-based screening in a clearly defined adult stroke population.

The latter requires more demanding clinical validation because its results could influence decisions about further assessment or oral intake.

An initial silent evaluation can compare AI outputs with appropriate professional reference assessments without allowing unvalidated predictions to direct patient care.

The hospital should examine false-negative results, false-positive findings, data quality and performance among patients with dysarthria, aphasia, cognitive impairment or other neurological deficits.

Recording conditions and language-related performance should be evaluated when acoustic models are used.

If the results justify further investigation, a limited supervised clinical pilot may be considered.

The pilot should preserve established swallowing screening requirements, professional assessment pathways and appropriate clinical oversight.

It should also determine whether AI-generated information reaches the relevant staff at a useful point in the care process.

Downstream capacity is important.

A system that increases referrals without improving access to specialist assessment may lengthen waiting times rather than improve care.

The hospital should therefore evaluate whether the intervention improves timely assessment, appropriate oral intake decisions and continuity of care without introducing unnecessary restrictions or workload.

Expansion should depend on demonstrated safety, clinical usefulness and available resources.

AI should remain a supporting component of swallowing safety rather than a prerequisite for providing essential clinical care.

From AI Screening to Safer Stroke Recovery

Post-stroke dysphagia illustrates why responsible AI implementation must begin with the clinical care pathway rather than the technology itself.

AI may help healthcare organizations identify patterns associated with swallowing impairment, retrieve relevant information or support interpretation of selected clinical signals.

Research published through 2026 demonstrates promising technical capabilities across screening, instrumental assessment and rehabilitation.

However, the available evidence does not establish that routine AI-assisted dysphagia screening consistently reduces aspiration pneumonia, dehydration, malnutrition, mortality or other adverse outcomes in Vietnamese healthcare settings.

Established clinical guidance provides a clearer foundation for action.

Patients with acute stroke require timely swallowing screening before oral intake, professional assessment when indicated, individualized nutrition and medication plans, and reassessment as their condition changes.

The national guidance issued through Decisions No. 3312/QD-BYT and No. 3554/QD-BYT provides relevant professional foundations for Vietnamese hospitals. International recommendations from AHA/ASA and European professional organizations offer additional evidence for strengthening these processes.

AI should be introduced where it can demonstrate an improvement beyond established care.

A patient who appears alert may still have impaired swallowing. A negative AI prediction cannot exclude silent aspiration. An abnormal screening result does not automatically determine the appropriate diet, medication route or rehabilitation plan.

Each decision requires relevant clinical evidence, professional judgment and consideration of the patient's condition and preferences.

Hospitals must also prevent unnecessary delays in hydration, nutrition and essential medication while supporting reassessment and appropriate progression toward oral intake during recovery.

Success should not be measured simply by how accurately a system classifies swallowing sounds or how many patients it identifies as high-risk.

It should be judged by whether patients receive timely assessment, appropriate nutrition and medication, effective rehabilitation and safer, more coordinated care throughout stroke recovery.

That is where AI-assisted post-stroke dysphagia screening may make a meaningful contribution to healthcare quality and patient safety.

Selected References and Regulatory Sources

Vietnamese Clinical and Rehabilitation Guidance

Ministry of Health of Viet Nam. Decision No. 3312/QD-BYT, dated 5 November 2024, issuing Diagnosis and Treatment of Stroke (Hướng dẫn chẩn đoán và điều trị đột quỵ não).

Ministry of Health of Viet Nam. Decision No. 3554/QD-BYT, dated 14 November 2025, issuing Guidance on Rehabilitation Technical Procedures, Volume 2 (Hướng dẫn quy trình kỹ thuật về Phục hồi chức năng, Tập 2). Effective 1 July 2026.

International Clinical Guidelines

Prabhakaran S, Gonzalez NR, Zachrison KS, et al. 2026 Guideline for the Early Management of Patients With Acute Ischemic Stroke: A Guideline From the American Heart Association/American Stroke Association. Stroke. 2026;57(8)–e436. DOI: 10.1161/STR.0000000000000513.

Prabhakaran S, Gonzalez NR, Zachrison KS, et al. Correction to: 2026 Guideline for the Early Management of Patients With Acute Ischemic Stroke. Stroke. 2026;57(8)–e467. DOI: 10.1161/STR.0000000000000530.

Richards LG, Ifejika NL, Stein J, et al. 2026 Guideline for Adult Stroke Rehabilitation and Recovery: A Guideline From the American Heart Association/American Stroke Association. Stroke. 2026;57(10)–e658. DOI: 10.1161/STR.0000000000000536.

Dziewas R, Michou E, Trapl-Grundschober M, et al. European Stroke Organisation and European Society for Swallowing Disorders Guideline for the Diagnosis and Treatment of Post-Stroke Dysphagia. European Stroke Journal. 2021;6(3)–CXV. DOI: 10.1177/23969873211039721.

International Dysphagia Diet Standardisation Initiative. IDDSI Framework and Descriptors. International framework for standardized food texture and drink thickness terminology.

Clinical AI and Dysphagia Research

De Andrade JBC, de Carvalho Leite SN, de Morais ABCG, Carneiro TS, Fernandes PC, Alves DI. Screening and Evaluation of Post-Stroke Dysphagia: Insights From Neurology, Artificial Intelligence and Data Science—A Scoping Review. Annals of Biomedical Engineering. Published July 2026. DOI: 10.1007/s10439-026-04267-7.

Sreedevi EV, Iyer KS, Thankappan K, et al. Artificial Intelligence Tools for the Assessment and Management of Dysphagia: A Scoping Review. BMJ Open. 2026;16(7). DOI: 10.1136/bmjopen-2026-119582.

Rech RS, Lopes Becker CD, Bernieri Schiavon DE, Vieira DP, Hilgert JB, Hugo FN. Artificial Intelligence for Dysphagia Screening: A Machine Learning Approach. PLOS Digital Health. 2026;5(9). DOI: 10.1371/journal.pdig.0001755.

Yang W, Li S, Du Y, et al. Human–AI Collaboration for Dysphagia Rehabilitation From Effectiveness to Implementation Complexity: A Systematic Review. npj Digital Medicine. 2026;9:404. DOI: 10.1038/s41746-026-02729-9.

Relevant Vietnamese Legal and Regulatory Framework

National Assembly of Viet Nam. Law on Medical Examination and Treatment No. 15/2023/QH15, as amended.

Ministry of Health of Viet Nam. Circular No. 31/2021/TT-BYT on nursing activities in hospitals.

Ministry of Health of Viet Nam. Circular No. 43/2018/TT-BYT on prevention of medical incidents in healthcare facilities.

Ministry of Health of Viet Nam. Circular No. 19/2013/TT-BYT on hospital quality management.

National Assembly of Viet Nam. Law on Artificial Intelligence No. 134/2025/QH15, effective 1 March 2026.

Government of Viet Nam. Decree No. 142/2026/ND-CP, detailing certain provisions and implementation measures of the Law on Artificial Intelligence, effective 1 May 2026.

Prime Minister of Viet Nam. Decision No. 33/2026/QD-TTg, promulgating the List of High-Risk Artificial Intelligence Systems, effective 15 August 2026.

National Assembly of Viet Nam. Law on Personal Data Protection No. 91/2025/QH15, effective 1 January 2026.

Government of Viet Nam. Decree No. 356/2025/ND-CP, detailing certain provisions and implementation measures of the Law on Personal Data Protection, effective 1 January 2026.

Government of Viet Nam. Decree No. 98/2021/ND-CP on medical device management, as amended.

Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT concerning medical device risk determination and related regulatory provisions, effective 1 July 2026.

Editorial and Clinical Notice

This article is intended for professional education, evidence-informed discussion, responsible healthcare AI governance and healthcare quality improvement. It draws on Vietnamese professional guidance, international clinical recommendations, published research and relevant legal and regulatory sources.

The clinical scenario is illustrative and does not represent a documented individual patient case.

The approaches discussed for AI-assisted dysphagia screening, swallowing assessment, aspiration risk management, oral intake decisions and quality evaluation are organizational implementation considerations. They do not establish a mandatory Vietnamese Ministry of Health AI implementation standard or constitute an independently validated clinical decision instrument.

This article does not provide individualized medical, swallowing, nutrition or legal advice, authorize automated oral intake decisions, replace professional clinical judgment or determine the regulatory status of any particular AI product.

Dysphagia screening, clinical swallowing assessment, instrumental evaluation, aspiration risk management, nutrition planning, medication administration and rehabilitation remain the responsibility of appropriately qualified healthcare professionals, consistent with applicable Vietnamese requirements, current clinical guidance and locally approved procedures.

An AI-generated negative screening result must not independently authorize oral intake or be treated as proof that aspiration cannot occur. Clinical findings, patient readiness and appropriate professional assessment must remain central to swallowing safety decisions.

The research discussed does not establish that AI-assisted post-stroke dysphagia screening necessarily reduces aspiration pneumonia, dehydration, malnutrition, mortality or other adverse outcomes in Vietnamese healthcare settings. Appropriate local validation, supervised implementation, meaningful professional oversight and ongoing evaluation remain necessary before claims of clinical effectiveness can be made.

International guidance should be interpreted within its stated patient populations and clinical settings. Organizational safeguards proposed in this article should not automatically be regarded as specific mandatory requirements under Vietnamese law.

Digital Medicine Vietnam (DMVN) | Evidence and regulatory review: 1 November 2026.