Governing AI-Assisted Pressure Injury Risk Prediction in Viet Nam: Nursing Assessment, Risk-Based Prevention and Outcome Monitoring

Governing AI-Assisted Pressure Injury Risk Prediction in Viet Nam: Nursing Assessment, Risk-Based Prevention and Outcome Monitoring

An older patient is admitted to a hospital with pneumonia. Initially, the patient can change position in bed with some assistance. Three days later, weakness has increased, nutritional intake has declined, and the patient is spending longer periods lying in the same position.

During the morning shift, an artificial intelligence (AI) system identifies an increased risk of pressure injury using information from the electronic health record.

The alert deserves attention, but it cannot tell the nurse whether tissue damage has already begun, whether the patient needs a different support surface, or whether the existing prevention plan remains appropriate. Those questions still require bedside assessment.

For the nursing team, the immediate priority is to understand what has changed and how the patient should be supported. The response may involve examining the patient's skin, reviewing mobility and nutritional needs, adjusting positioning assistance, checking medical devices, or arranging additional clinical support.

Identifying risk and preventing harm are different tasks. A prediction may draw attention to a developing problem, but it does not ensure that appropriate preventive care is delivered.

As healthcare organizations in Viet Nam explore AI-assisted nursing applications, pressure injury prevention illustrates both the potential value of predictive technology and the importance of understanding its limitations.

The challenge is not simply to identify which patients are more likely to develop pressure injuries. It is to connect changing risk with timely nursing assessment, individualized prevention, and reliable monitoring of what happens afterward.

Pressure Injury Prevention Depends on Recognizing Changing Risk

A pressure injury is localized damage to the skin or underlying soft tissue, usually occurring over a bony prominence or in association with a medical or other device. It can result from intense or prolonged pressure, or pressure combined with shear. Whether tissue damage develops also depends on the patient's underlying condition and the ability of the tissues to tolerate mechanical loading.

Immobility is a major concern, but it is not the only one. Impaired sensation, reduced tissue perfusion, nutritional problems, serious illness, and pressure associated with medical devices can all increase vulnerability.

Some patients deteriorate gradually. Others experience sudden changes following surgery, acute infection, sedation, circulatory instability, or transfer to a higher level of care. Someone who could reposition independently yesterday may require extensive nursing assistance today.

Pressure injuries can cause pain, infection, delayed recovery, and considerable distress. They can also increase the complexity and cost of hospital care. Although many injuries can be prevented or their severity reduced, not every pressure injury is avoidable, particularly in patients with severe physiological compromise.

Prevention therefore requires attention to both clinical risk and the circumstances in which care is delivered.

Nurses need to recognize changes in patient condition, but they also need sufficient time, staffing, equipment, and multidisciplinary support to respond. A documented assessment offers limited protection if the interventions identified as necessary cannot be carried out.

Traditional tools such as the Braden, Norton, and Waterlow scales can help structure pressure injury risk assessment. Their predictive performance, however, varies between patient populations and care settings. A total score does not fully explain why an individual patient is vulnerable or which preventive measures are most appropriate.

AI offers another way to examine risk. By analyzing routinely collected clinical information, a model may identify patterns that are difficult to recognize when relevant observations are distributed across different records.

Whether that additional information improves prevention depends on its reliability, clinical relevance, and integration into everyday nursing practice.

What the 2026 International Guideline Means for AI-Assisted Prediction

The fourth edition of the Prevention and Treatment of Pressure Ulcers/Injuries: Clinical Practice Guideline, developed by the National Pressure Injury Advisory Panel (NPIAP), European Pressure Ulcer Advisory Panel (EPUAP), and Pan Pacific Pressure Injury Alliance (PPPIA), was released in 2026.

For healthcare organizations considering AI-assisted pressure injury prediction, the guideline provides an important reference for clinical practice.

It describes a structured approach connecting early risk screening, comprehensive risk assessment, and individualized prevention planning.

Risk screening should occur as soon as possible after admission or initial contact with a healthcare service, with reassessment at appropriate intervals and when the patient's condition worsens. Patients identified as likely to be at risk should receive a comprehensive assessment incorporating relevant clinical factors, skin and tissue findings, and professional judgment.

There is an important qualification. In populations where substantial pressure injury risk is already evident, such as critically ill patients or individuals with certain spinal cord injuries, preliminary screening may offer little additional value. These patients may be considered at risk from the outset and proceed directly to comprehensive assessment.

The guideline also discusses AI and machine learning approaches to risk prediction.

Although these technologies are being developed and evaluated, the guideline states that AI-based pressure injury risk assessment systems are not yet established for routine clinical use as methods of comprehensive risk assessment. No AI-based program has been accepted as a replacement for the full assessment process described in the guideline.

This reflects the current state of clinical evidence and guidance. It should not be interpreted as a legal prohibition against AI research or appropriately governed evaluation.

For hospitals, the implication is that AI-generated risk estimates may provide supplementary clinical information, but they should not determine whether essential nursing assessments are performed or whether indicated preventive care is provided.

A patient with significant mobility limitations, impaired perfusion, or early signs of tissue damage requires appropriate clinical attention even when an algorithm produces a low-risk result.

The guideline reinforces a broader principle: risk assessment should help clinicians identify the factors that can be addressed through prevention, rather than simply assign patients to numerical risk categories.

Understanding What the AI System Actually Predicts

AI-assisted pressure injury prediction typically uses clinical information to estimate the likelihood of a future pressure injury over a defined period.

Depending on the system, the information analyzed may include age, mobility, functional dependence, comorbidities, previous pressure injuries, nutritional indicators, physiological observations, medical devices, and other variables recorded during hospitalization.

Some models generate predictions at admission. Others update risk estimates as new clinical information becomes available.

Before introducing such a system, clinicians need to understand the population for which it was developed, the outcome being predicted, the prediction period, and the information used to generate the result.

These distinctions have practical consequences.

A model estimating pressure injury risk during hospitalization is not necessarily suitable for predicting whether an injury will develop within the next 24 hours. Similarly, a model developed for adult medical and surgical wards cannot be assumed to perform reliably in intensive care, pediatric services, or long-term care settings.

Risk prediction must also be distinguished from detecting or classifying an injury that already exists.

An AI model analyzing electronic health records may identify increased future risk without recognizing that early tissue damage is already present. Conversely, an image-based wound assessment system may help characterize an existing injury without providing a validated estimate of future pressure injury risk.

These are different clinical functions and require different forms of evaluation.

A high-risk prediction does not mean an injury is inevitable, just as a low-risk prediction does not establish that an injury cannot occur.

The reliability of a prediction depends partly on the underlying information. A model may not recognize changes in mobility that have not been documented, skin abnormalities that were overlooked, or preventive interventions recorded inconsistently.

Differences in documentation and clinical practice between hospitals may further affect performance.

The prediction should therefore be interpreted as one source of clinical information, not as a complete assessment of the patient's condition.

Nursing Assessment Remains the Foundation of Prevention

Consider two patients who receive similar AI-generated risk estimates.

One has reduced mobility following abdominal surgery but remains physiologically stable. The other is recovering from severe illness, has impaired peripheral perfusion, and requires several medical devices.

Their overall risk estimates may be similar, but their prevention needs could differ substantially.

The first patient may benefit from assisted repositioning, appropriate pressure redistribution, and early mobilization when clinically suitable. The second may need closer tissue monitoring, careful coordination of positioning with medical treatment, and particular attention to device-related pressure.

The value of nursing assessment lies in identifying these differences.

Mobility, activity, sensory function, moisture exposure, nutrition, tissue perfusion, general medical condition, and treatment-related restrictions all contribute to the clinical picture.

The assessment must also identify injuries that are already present and distinguish pressure-related tissue damage from other conditions, including moisture-associated skin damage.

Comprehensive skin and tissue examination is particularly important because early injury may already exist when an AI prediction appears.

The 2026 international guideline recommends comprehensive skin and tissue assessment as soon as possible after admission or transfer, as part of pressure injury risk assessment, periodically according to the patient's condition, when clinical deterioration occurs, at relevant transitions in care, and before discharge.

Examination should include the sacrum, heels, and other vulnerable areas, together with skin exposed to pressure from masks, oxygen interfaces, tubing, splints, braces, and other devices.

Pressure-related changes may occur in locations that receive less attention during routine assessment.

Skin tone is another consideration. Early pressure-related damage may be harder to recognize when assessment relies primarily on visible redness. Changes in skin color relative to the patient's usual appearance, temperature, tissue consistency, swelling, and localized pain may provide additional evidence of injury.

The guideline emphasizes visual and tactile assessment techniques appropriate to the individual's skin characteristics and clinical condition.

An AI system cannot reliably account for bedside findings that were never recognized or documented.

Nurses must therefore retain the ability to identify clinically important risk independently of the algorithm. When bedside findings and the AI prediction disagree, the patient's clinical condition should guide the immediate response.

The disagreement may subsequently provide useful information about the model's limitations, the quality of available data, or the need for reassessment.

Turning Risk Information Into Individualized Prevention

An AI alert becomes clinically useful when it helps the care team determine whether prevention needs to change.

For a patient whose mobility has declined, the response may involve reviewing assisted repositioning, pressure-redistributing support surfaces, heel protection, and the frequency of skin inspection.

For someone receiving oxygen therapy, attention may be needed around the ears, nose, or face. Another patient may be vulnerable to pressure from a brace, catheter, or other therapeutic device.

Poor nutritional intake may indicate a need for further nutritional assessment and support. Moisture exposure, impaired circulation, skin fragility, or treatment-related restrictions may require different interventions.

The 2026 guideline emphasizes addressing the individual factors contributing to risk rather than selecting preventive measures solely from a total assessment score.

Repositioning provides a practical example.

Regular movement and pressure relief remain fundamental preventive measures, but the interval between repositioning events should reflect the patient's condition, ability to move, tissue tolerance, support surface, comfort, and goals of care.

The guideline conditionally suggests that repositioning at two-hourly or three-hourly intervals may be appropriate for many individuals at risk when a suitable pressure-redistributing full-body support surface is available.

It also suggests avoiding routine extension to four-, five-, or six-hour intervals. Longer intervals may be appropriate for selected individuals whose risk has decreased, whose ability to reposition effectively has improved, and whose skin and tissue condition remains satisfactory.

These recommendations are supported by very low-certainty evidence and should not be interpreted as fixed schedules suitable for every patient.

Critically ill patients with shock or systemic hypoperfusion may require more frequent, incremental repositioning or smaller assisted shifts in body position. When physiological instability prevents regular full-body repositioning, the care team may need to adapt the regimen according to hemodynamic and respiratory tolerance, while maintaining appropriate pressure relief and tissue monitoring.

In palliative or end-of-life care, comfort, patient preferences, and goals of care may justify a different approach. The patient and caregivers should be informed about the potential consequences of less frequent repositioning.

No support surface completely eliminates the need to consider repositioning and pressure relief.

Preventive care also needs to be proportionate. Unnecessary interventions, inappropriate equipment, excessive movement, and sleep disruption may introduce discomfort or other risks without providing meaningful additional benefit.

Patients and family caregivers can contribute to prevention when circumstances allow, particularly by reporting discomfort, changes in mobility, or new skin concerns.

An AI-generated high-risk alert should therefore prompt a review of the patient's actual prevention needs, not automatic application of every available intervention.

The team should also recognize when appropriate preventive measures are already in place. Confirming the adequacy of an existing plan may be the correct response, even if no additional intervention is introduced.

Integrating AI Into Daily Nursing Practice

The way risk information reaches nurses can be as important as the predictive model itself.

An alert appearing after the relevant nursing assessment may provide little benefit. A dashboard requiring staff to open another application may be overlooked, while repeated notifications about risks already addressed can increase workload without improving care.

AI-assisted prediction should therefore be evaluated within established nursing routines.

Daily handover, ward safety discussions, and structured care reviews may provide opportunities to consider changes in predicted risk alongside the patient's condition, existing assessments, and outstanding care needs.

An increased risk estimate during morning handover, for example, could prompt the team to review whether the patient has become less mobile, whether repositioning assistance remains adequate, and whether new pressure-related skin changes have been observed.

The alert may also reveal an operational problem. A patient may require a pressure-redistributing mattress that is not immediately available, or additional assistance with positioning that the current staffing arrangement cannot reliably provide.

Such situations require clinical coordination and resource decisions, not simply another notification.

The hospital should establish clear responsibilities for reviewing predictions, completing assessments, updating prevention plans, and escalating unresolved concerns.

Depending on the patient's needs and the hospital's approved model of care, these responsibilities may involve the bedside nurse, nurse in charge, wound care specialist, attending clinician, nutrition professional, physiotherapist, or other members of the multidisciplinary team.

Documentation should make it possible to understand whether the patient was assessed, what concerns were identified, which interventions were planned, and whether the necessary care was delivered.

An alert acknowledgment alone is not evidence of successful prevention.

At the same time, hospitals should avoid introducing excessive documentation requirements. Repeatedly entering information already available in the clinical record can increase nursing workload without contributing to patient safety.

The workflow must also remain functional when AI is unavailable. A technical failure, delayed data update, or temporary interruption of system access should not prevent essential nursing assessments or interventions.

Technology should support the continuity of care rather than create a dependency that makes prevention less reliable.

What the Latest Research Shows

Research published in 2026 provides encouraging findings about AI-assisted pressure injury risk prediction, although important questions about clinical effectiveness remain unanswered.

In October 2026, Greup and colleagues published a study in the Journal of Clinical Nursing describing the development and validation of a dynamic pressure injury risk prediction model using electronic health record data from Erasmus University Medical Center in the Netherlands.

The researchers worked with nursing experts to select predictors and consider how the model could be used in clinical practice.

Separate models were developed for the first 72 hours of hospitalization and for subsequent periods.

In the validation datasets, the reported areas under the receiver operating characteristic curve (AUROC) were 0.790 and 0.795, respectively.

In the subsets of patients with recorded Waterlow assessments, the model achieved better discrimination than the Waterlow score. The corresponding AUROC comparisons were 0.816 versus 0.702 during the first 72 hours and 0.800 versus 0.677 afterward.

These findings suggest that information already available in electronic hospital records can support regularly updated risk estimation.

However, the study was conducted at a single tertiary hospital, with intensive care and psychiatric wards excluded. The validation used data from a later period at the same institution rather than an independent evaluation across multiple healthcare systems.

Its findings cannot automatically be generalized to Vietnamese hospitals or other patient populations.

Importantly, the study assessed predictive performance. It did not demonstrate that using the model in routine clinical care reduced pressure injury incidence.

A related study by Spoon and colleagues examined the acceptability and feasibility of integrating the Decubitus Risk Prediction Alerts based on Artificial Intelligence (DRAAI) system into nursing practice.

The pilot was conducted in three general hospital wards, with 55 nurses participating in the evaluation.

The system generated predictions for 428 hospitalized patients. Among them, 128 received at least one prediction indicating increased risk.

A nursing care plan was documented for 101 of these 128 patients, approximately 79%.

The timing of those care plans matters. For 29 patients, a plan had already been created before the first at-risk AI prediction. Forty patients had a plan created within 24 hours after the first alert, another 14 within 24 to 48 hours, and 18 after more than 48 hours. For 27 patients, no corresponding nursing care plan was documented during the evaluation period.

These findings describe the relationship between AI alerts and documented nursing care plans. They do not establish that AI caused the plans to be created or that every planned intervention was delivered.

The nurses generally considered the system feasible to incorporate into clinical practice. Nevertheless, some initially found it difficult to distinguish between an alert predicting future pressure injury risk and the identification of an injury that already existed.

The study highlighted the importance of education, ongoing communication, and nurse involvement in implementation.

It provided useful evidence about workflow feasibility and user acceptance, but it did not establish that AI-assisted prediction reduced pressure injuries.

A systematic review by Sengul and colleagues, published in the International Wound Journal on 22 October 2026, examined digital, mobile, and AI-supported interventions for pressure injury prevention and management.

The review included 12 studies published between 2017 and 2025.

The interventions were diverse, including wearable monitoring systems, repositioning support technologies, mobile applications, telenursing, digital documentation, and AI-assisted wound assessment.

Only one included study specifically examined AI-based image recognition for pressure injury assessment. The review should therefore not be treated as evidence focused exclusively on AI risk prediction models.

Across the included studies, improvements were reported in several care processes, including repositioning adherence, monitoring, documentation, and caregiver or staff engagement.

Evidence for sustained reductions in pressure injury incidence, improved healing, cost-effectiveness, and long-term implementation was less consistent. Differences between interventions, settings, populations, and study designs limited the extent to which findings could be combined or generalized.

Taken together, these studies support further investigation of AI and digital tools as potential contributors to pressure injury prevention.

They do not yet establish that a hospital will achieve better patient outcomes simply by introducing an AI prediction system with good retrospective performance.

For healthcare organizations, that distinction should guide both investment decisions and expectations about clinical benefit.

Why Vietnamese Hospitals Need Local Validation

A prediction model developed in another country may perform differently when introduced into a Vietnamese hospital.

Hospitals vary in patient populations, clinical documentation, staffing, electronic record systems, nursing practices, and the availability of preventive equipment.

Some facilities document skin assessments through structured electronic forms. Others may rely more heavily on narrative clinical notes, paper records, or separate wound care documentation.

These differences affect the information that a model can analyze. They also influence how pressure injuries are recorded and identified for research and quality monitoring.

If an injury already present on admission is documented only several days later, it may be incorrectly classified as hospital-acquired. If pressure injuries are recorded mainly when specialist wound care is requested, the resulting data may reflect referral practices as well as actual injury occurrence.

A model trained on incomplete documentation may therefore learn to predict which injuries are recorded rather than which patients develop tissue damage.

Before evaluating an AI system, hospitals should establish a reliable process for identifying pressure injury outcomes, distinguishing existing injuries from new injuries, and reviewing uncertain cases.

Local validation should use data representative of the intended patient population and clinical setting.

The evaluation should examine discrimination, calibration, and performance at clinically meaningful thresholds.

Discrimination describes how well a model separates patients who experience the defined outcome from those who do not. Calibration examines whether predicted probabilities correspond reasonably with observed event frequencies.

A model may demonstrate acceptable discrimination but still produce poorly calibrated risk estimates.

The practical consequences of selecting an alert threshold also matter.

Hospitals need to understand how many patients would be classified as high-risk, how frequently clinically important cases might be missed, and what additional workload would be created.

A threshold that performs satisfactorily in retrospective analysis may generate too many alerts for a busy ward to review effectively.

There is also a difference between statistical performance and clinical usefulness. A model may improve prediction compared with an existing assessment tool but contribute little additional value if it identifies patients whose risks have already been recognized and appropriately addressed.

Hospitals should therefore evaluate not only predictive accuracy but also whether the system identifies clinically meaningful opportunities for prevention.

Performance should be examined across relevant patient groups, including different wards, admission types, age groups, and clinical conditions.

A model validated for adult general wards should not automatically be extended to intensive care, pediatric services, or other populations without supporting evidence.

Another concern is information leakage. Data used to generate a prediction must correspond to information that would actually have been available at the intended prediction time. Including observations or outcomes recorded afterward can produce misleading estimates of model performance.

Preventive interventions introduce an additional complication.

Patients assessed as high-risk may receive effective prevention and never develop an injury. Their favorable outcome does not necessarily mean the original risk estimate was inappropriate.

Conversely, a patient may develop an injury despite appropriate assessment and preventive care.

The relationship between predicted risk and observed outcomes must therefore be interpreted in the context of the care patients actually received.

Local validation is not a one-time activity. Changes in patient populations, documentation practices, electronic record systems, clinical workflows, or model configuration may affect performance.

Hospitals need a process for recognizing when reassessment is necessary and deciding whether the system remains suitable for its intended use.

Measuring Whether Prevention Actually Improves

If a hospital introduces AI-assisted prediction, it needs to determine whether prevention has become more reliable and whether patient outcomes have improved.

A credible evaluation begins with understanding the existing situation.

Before introducing AI, the hospital should review its pressure injury surveillance methods, admission assessments, prevention practices, documentation quality, and relevant outcome measures.

Distinguishing injuries present on admission from those acquired during hospitalization is particularly important.

Accurate classification requires timely examination and documentation, including when patients are transferred between departments or levels of care.

Pressure injury prevalence and incidence should also be distinguished.

Prevalence describes the proportion of patients with pressure injuries at a specified point or during a defined period, including injuries that may have developed before admission.

Incidence concerns newly developing injuries within a defined population and observation period.

These measures answer different questions and should not be compared as though they represent the same outcome.

For hospital-acquired pressure injuries, organizations may monitor the proportion of eligible hospitalized patients who develop a new injury and, where appropriate, incidence per 1,000 patient-days.

The numerator, denominator, observation period, inclusion criteria, and surveillance methods must be clearly defined and applied consistently.

Separate monitoring of Category/Stage 2 or higher injuries may provide useful information about more substantial tissue damage. However, limiting surveillance to these injuries can overlook earlier damage and other clinically important categories, including deep tissue pressure injuries and unstageable injuries.

Medical device-related pressure injuries should also be identified. Injuries involving mucosal membranes require appropriate documentation but should not be assigned conventional skin pressure injury stages.

Outcome monitoring should extend beyond injury counts.

Hospitals need to understand whether patients receive timely risk assessments, whether skin and tissue examinations are completed, whether prevention plans address individual risk factors, and whether the selected interventions are actually delivered.

For AI-assisted workflows, additional measures may include the proportion of clinically relevant alerts reviewed, the timeliness of follow-up, the frequency of important risks identified despite negative AI predictions, and the number of alerts that do not lead to meaningful changes in care.

One particularly useful question is whether preventive action follows a clinically important change in the patient's condition.

A care plan recorded in the electronic system does not necessarily mean that repositioning, pressure relief, skin care, or other indicated interventions were completed.

Balancing measures are equally necessary. Increased nursing workload, duplicated documentation, unnecessary interventions, patient discomfort, disrupted sleep, and additional equipment or maintenance costs may offset potential benefits.

Changes in pressure injury incidence must also be interpreted carefully.

After a hospital improves skin assessment and documentation, more early-stage injuries may be identified. An initial increase in recorded cases may partly reflect better detection rather than worsening prevention.

Similarly, a reduction in documented injuries may reflect changes in case mix, surveillance methods, or reporting practices rather than a genuine reduction in harm.

Hospital leadership should examine outcome trends alongside patient severity, length of stay, ward characteristics, staffing, and changes in other preventive practices.

When feasible, evaluation designs should include appropriate comparison data and account for concurrent quality-improvement initiatives.

A decrease in recorded pressure injuries after AI implementation is not, by itself, evidence that AI caused the improvement.

Learning From Missed Risk and Unnecessary Alerts

Disagreement between an AI prediction and a nursing assessment can provide valuable opportunities for learning.

Consider a patient classified as low-risk whose mobility has deteriorated and whose heels show early tissue changes.

The clinical findings require attention regardless of the prediction.

The case may subsequently reveal that the model relied on outdated mobility information, did not capture relevant skin observations, or failed to recognize an important risk pattern.

The opposite situation also deserves consideration.

An AI system may generate a high-risk alert for a patient who is already receiving an appropriate prevention plan.

In that case, reviewing the existing plan and confirming that it remains suitable may be sufficient. The absence of an additional intervention should not automatically be regarded as a failure to respond.

Hospitals need to distinguish clinically justified decisions from genuine gaps in care.

When a hospital-acquired pressure injury develops, case review should consider the patient's condition, documented risk factors, bedside findings, preventive interventions, medical devices, equipment availability, and relevant clinical changes.

Where AI was involved, its predictions and the responses they prompted can contribute to the review.

The purpose is to identify missed opportunities for prevention and understand whether the system supported care or introduced confusion, delays, or additional workload.

An algorithmic warning does not prove that an injury was preventable, and the occurrence of an injury does not automatically establish individual professional error.

A constructive review process should encourage nurses to report concerns about inaccurate predictions, unclear alerts, and operational difficulties.

The resulting findings may guide changes in staff education, documentation, clinical pathways, alert presentation, or model configuration.

Such reviews are most valuable when they improve the reliability of care rather than assign blame based on retrospective algorithmic predictions.

Legal and Professional Responsibilities in Viet Nam

AI-assisted pressure injury risk prediction must be considered within Viet Nam's legal framework governing healthcare practice, nursing activities, artificial intelligence, medical devices, and personal data.

The Law on Medical Examination and Treatment No. 15/2023/QH15 provides the broader legal framework for professional healthcare practice, patient rights, and the responsibilities of healthcare organizations.

Ministry of Health Circular No. 31/2021/TT-BYT, effective from 27 February 2022, is particularly relevant because it regulates nursing activities in hospitals.

The Circular establishes requirements concerning clinical assessment, identification of patient needs and risks, nursing interventions, and evaluation of care outcomes. It emphasizes comprehensive, continuous, safe, and individualized nursing care.

These professional responsibilities cannot be considered fulfilled merely because an AI system generates a prediction or displays an alert.

Other relevant quality and safety instruments include Circular No. 19/2013/TT-BYT concerning hospital quality management and Circular No. 43/2018/TT-BYT concerning prevention of medical incidents, to the extent their provisions apply to the organization's activities.

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 detailed provisions for implementing the law, including matters related to AI risk classification and governance.

Decision No. 33/2026/QD-TTg, effective from 15 August 2026, establishes the current list of high-risk AI systems.

The healthcare entries in this list specifically concern certain AI-assisted surgical and robotic surgical systems. AI-assisted pressure injury risk prediction is not expressly listed as a separate healthcare category.

However, the absence of a specifically named category does not automatically establish that every such system is exempt from applicable AI obligations or presents minimal clinical risk.

The legal assessment must consider the system's intended purpose, actual functions, degree of automation, clinical consequences, and applicable statutory provisions.

Responsibilities may also differ between the AI provider and the healthcare organization deploying the system.

Medical device regulation requires a separate determination.

Decree No. 98/2021/ND-CP on medical device management, as amended, and relevant Ministry of Health instruments, including Circular No. 24/2026/TT-BYT, may apply depending on the intended medical purpose and characteristics of the product.

An AI application should not automatically be classified as a medical device simply because it analyzes healthcare information. Conversely, a product should not be assumed to fall outside medical device regulation merely because it is described as a nursing dashboard or administrative support system.

The intended use and actual capabilities require assessment before a particular product is introduced into clinical practice.

Patient data protection is another essential consideration.

AI-assisted prediction may process identifiable information about diagnoses, medications, mobility, nursing assessments, physiological observations, and other aspects of patient care.

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

Health status information falls within the scope of sensitive personal data and requires appropriate safeguards.

Decree No. 102/2025/ND-CP on health data management, effective from 1 July 2025, is also relevant to the governance, protection, processing, and use of digital health data.

Healthcare organizations should determine the applicable legal grounds for processing personal information, define the responsibilities of participating parties, and establish safeguards appropriate to the actual data processing activities.

These may include access restrictions, information security measures, data retention arrangements, vendor agreements, incident management procedures, and applicable assessment or documentation obligations.

When external technology providers are involved, contractual arrangements should clarify their access to clinical records, responsibilities for system maintenance, handling of security incidents, and permitted uses of patient information.

Secondary use of clinical records for model development, retraining, research, or commercial purposes requires separate consideration of the applicable legal and ethical requirements. Authorization to process information for patient care should not automatically be interpreted as authorization for every subsequent use.

Cross-border data transfers, overseas hosting, and access by external service providers must also be assessed under the relevant legal provisions, including any applicable impact assessment and documentation requirements.

These matters should be addressed before identifiable clinical information is made available to an external AI service.

International clinical guidelines may help hospitals develop evidence-informed protocols, but they do not automatically become binding Vietnamese legal requirements.

Similarly, proposed AI governance practices should be distinguished from statutory obligations unless a specific legal provision makes them mandatory.

A Practical Starting Point for Vietnamese Hospitals

A hospital considering AI-assisted pressure injury prediction should begin by examining how reliably its existing prevention process works.

The first question is whether patients are already receiving appropriate risk assessments and preventive care without AI.

Nursing leadership and quality management teams can review representative patient records to determine whether admission skin assessments are completed, whether changes in condition lead to reassessment, and whether prevention plans reflect individual risk factors.

They should also examine whether planned interventions are actually delivered.

A review may reveal that the main problems are incomplete documentation, inconsistent repositioning practices, limited access to suitable support surfaces, or insufficient nursing resources.

If nurses already recognize patients at risk but cannot obtain appropriate equipment or support, the immediate priority may be improving the prevention process rather than introducing a new prediction system.

Where documentation is incomplete or outcomes are inconsistently classified, improving data quality may be necessary before meaningful model validation can take place.

Once the baseline is understood, the hospital can define a limited clinical use case for evaluation.

For example, a supervised study might examine whether regularly updated AI predictions help nurses recognize changing pressure injury risk on selected adult medical or surgical wards.

The intended population, prediction timing, clinical response, staff responsibilities, and evaluation measures should be specified before implementation.

An initial retrospective or appropriately governed non-interventional evaluation may assess model performance without replacing existing nursing processes. The use of patient information must still comply with applicable authorization, privacy, research, and institutional requirements.

A subsequent supervised pilot can examine whether predictions reach nurses at useful times, whether alerts prompt appropriate reviews, whether prevention plans remain suitable, and whether the system introduces unnecessary workload or unintended clinical consequences.

The evaluation should include a process for responding to safety concerns as they arise. Clinically necessary care should never be delayed simply to preserve an evaluation protocol.

Nursing staff should participate throughout the process, including in selecting the use case, reviewing alert design, evaluating clinical relevance, and identifying workflow difficulties.

Their involvement is essential because the success of the system depends on how predictions are interpreted and acted upon in real clinical settings.

Decisions about expansion should consider clinical performance, workflow reliability, available resources, regulatory requirements, and the consequences of missed or unnecessary alerts.

Changes to the model, data sources, or intended patient population may require further evaluation before the system is used more broadly.

The hospital should also be prepared to modify, suspend, or discontinue use if evidence indicates that the system is unreliable, introduces unacceptable risks, or does not provide sufficient practical value.

The objective is not to create a prevention program dependent on AI. It is to determine whether AI makes an established patient-safety process more effective and reliable.

From Predicting Risk to Delivering Safer Nursing Care

AI-assisted pressure injury risk prediction offers a potential way to help healthcare organizations recognize changing patient needs and direct clinical attention toward patients who may require additional preventive care.

Recent research demonstrates that models using routinely collected hospital information can identify patterns associated with pressure injury risk. Early implementation studies also suggest that AI-generated alerts can be incorporated into nursing workflows when clinicians participate in their development and evaluation.

However, good predictive performance and positive user acceptance do not necessarily translate into improved patient outcomes.

For healthcare organizations in Viet Nam, responsible implementation requires a clear distinction between an algorithmic risk estimate, professional nursing assessment, and the actual delivery of prevention.

The 2026 international guideline reinforces the importance of comprehensive assessment, individualized interventions, and clinical judgment. AI should be considered within that framework rather than used to bypass it.

An alert needs to reach the appropriate professional at a useful time. The patient's condition must be assessed, the prevention plan must address relevant risk factors, and the organization must have sufficient capacity to deliver the care identified as necessary.

Outcome monitoring must then establish whether preventive care is being delivered more reliably and whether avoidable harm is being reduced.

The greatest practical value of AI may ultimately lie not in replacing an established risk assessment tool, but in drawing attention to a clinically meaningful change that might otherwise have been overlooked.

Even then, the prediction is only the beginning.

The real measure of success is whether patients receive timely, appropriate, and individualized preventive care—and whether hospitals can demonstrate that their approach to pressure injury prevention is becoming safer.

That is the standard against which AI-assisted pressure injury risk prediction should be evaluated.

Selected References and Regulatory Sources

International Clinical Guidelines

National Pressure Injury Advisory Panel, European Pressure Ulcer Advisory Panel, and Pan Pacific Pressure Injury Alliance. Prevention and Treatment of Pressure Ulcers/Injuries: Clinical Practice Guideline. Fourth Edition. Emily Haesler, Editor. 2026.

National Pressure Injury Advisory Panel, European Pressure Ulcer Advisory Panel, and Pan Pacific Pressure Injury Alliance. Pressure Injury Risk. In: Prevention and Treatment of Pressure Ulcers/Injuries: Clinical Practice Guideline. Fourth Edition. 2026.

National Pressure Injury Advisory Panel, European Pressure Ulcer Advisory Panel, and Pan Pacific Pressure Injury Alliance. Skin and Tissue Assessment. In: Prevention and Treatment of Pressure Ulcers/Injuries: Clinical Practice Guideline. Fourth Edition. 2026.

National Pressure Injury Advisory Panel, European Pressure Ulcer Advisory Panel, and Pan Pacific Pressure Injury Alliance. Repositioning and Mobilization. In: Prevention and Treatment of Pressure Ulcers/Injuries: Clinical Practice Guideline. Fourth Edition. 2026.

National Pressure Injury Advisory Panel, European Pressure Ulcer Advisory Panel, and Pan Pacific Pressure Injury Alliance. Classification. In: Prevention and Treatment of Pressure Ulcers/Injuries: Clinical Practice Guideline. Fourth Edition. 2026.

Clinical AI and Implementation Evidence

Greup S, Spoon D, de Vroed A, Werkhoven B, Timmerman E, van Dijk M, et al. Dynamic Pressure Ulcer Risk Predictions for Hospitalized Patients: Development and Validation of a Machine Learning Model With Nurses. Journal of Clinical Nursing. Published online 7 October 2026. DOI: 10.1111/jocn.70574.

Spoon D, de Vroed A, Greup S, van Dijk M, Ista E. Integrating Artificial Intelligence in Nursing Practice With Decubitus Risk Prediction Alerts: A Pilot Process Evaluation. Journal of Clinical Nursing. 2026;35(10):4184–4197. DOI: 10.1111/jocn.70212.

Sengul T, Koyuncu F, Sarıköse S, Yilmaz Akyaz D, Kozan Çikirikçi EH. Digital, Mobile and Artificial Intelligence-Supported Interventions for Pressure Injury Prevention and Management: A Systematic Review of Clinical, Workflow and Implementation Outcomes. International Wound Journal. 2026;23(10). DOI: 10.1111/iwj.71078.

Relevant Vietnamese Legal and Regulatory Framework

Law on Medical Examination and Treatment No. 15/2023/QH15.

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

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

Ministry of Health Circular No. 43/2018/TT-BYT on prevention of medical incidents.

Law on Artificial Intelligence No. 134/2025/QH15; Decree No. 142/2026/ND-CP; Decision No. 33/2026/QD-TTg.

Law on Personal Data Protection No. 91/2025/QH15; Decree No. 356/2025/ND-CP.

Decree No. 102/2025/ND-CP on health data management.

Decree No. 98/2021/ND-CP on medical device management, as amended; Ministry of Health Circular No. 24/2026/TT-BYT.

Editorial and Clinical Notice

This article is intended for professional education, healthcare quality improvement, and discussion of responsible AI implementation. It draws on international clinical guidelines, published research, and relevant Vietnamese regulatory sources.

It does not establish a mandatory AI implementation standard, replace professional nursing assessment, provide patient-specific medical advice, or constitute a legal determination of any particular AI product's regulatory status.

The clinical examples and proposed implementation approaches are illustrative. Individual care decisions and professional responsibilities must remain with appropriately qualified healthcare professionals, consistent with applicable Vietnamese law, current clinical guidance, and locally approved procedures.

The studies discussed do not establish that AI-assisted pressure injury risk prediction necessarily reduces pressure injury incidence or improves patient outcomes in Vietnamese healthcare settings. Claims of clinical effectiveness require appropriate supporting evidence, including evaluation relevant to the intended population and setting.

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