Governing AI-Assisted Inpatient Delirium Risk Identification in Viet Nam: Targeted Bedside Assessment, Multicomponent Prevention and Safe Clinical Response

Governing AI-Assisted Inpatient Delirium Risk Identification in Viet Nam: Targeted Bedside Assessment, Multicomponent Prevention and Safe Clinical Response

During a morning handover, a nurse notices that an older patient recovering from hip fracture surgery is unusually quiet. Before surgery, the patient recognized family members and answered questions normally. Now, the patient struggles to follow a conversation, responds slowly and shows little interest in eating. There is no agitation or disruptive behavior. The change could easily be attributed to fatigue, pain medication or the effects of surgery.

The hospital's artificial intelligence (AI) system has identified several factors associated with delirium, including advanced age, recent surgery, acute illness and medication exposure. It generates an alert recommending further assessment.

The alert may be useful, but it cannot determine what has happened to the patient. The nurse still needs to establish whether there has been an acute change from the patient's usual mental state. The clinical team must consider possible causes, including medication effects, hypoxemia, infection, metabolic disturbances and other complications. Appropriate assessment cannot depend on whether AI has generated a warning.

This situation illustrates a central challenge in AI-assisted delirium prevention. Predicting who may develop delirium is different from recognizing an acute cognitive change, and neither task guarantees that the patient will receive appropriate preventive or clinical care.

For healthcare organizations in Viet Nam, AI may help professionals recognize vulnerability, identify overlooked information and coordinate timely assessment. Its implementation must not weaken established medical and nursing responsibilities, encourage unnecessary sedation or turn a risk prediction into an automatic diagnosis.

The objective is to improve the reliability of delirium prevention and clinical response, not simply to classify more patients as high-risk.

Why Delirium Matters for Hospital Patient Safety

Delirium is an acute disturbance of attention and awareness, accompanied by changes in cognition, that develops over a relatively short period and typically fluctuates over time.

Patients may have difficulty concentrating, following conversations or understanding their surroundings. They may experience disorganized thinking, altered perception, disturbances of the sleep–wake cycle or changes in psychomotor activity.

Some become restless, frightened or agitated. Others become unusually quiet, withdrawn or less responsive. Mixed presentations may occur, with symptoms changing during the same hospital admission.

These differences contribute to underrecognition. A patient repeatedly trying to leave bed or remove medical equipment is likely to attract attention. Someone who becomes less communicative, moves slowly and sleeps through much of the day may simply be described as tired.

The hypoactive presentation deserves particular attention because it can be mistaken for depression, fatigue, dementia or the expected effects of aging.

Delirium is associated with longer hospital stays, functional decline, complications, increased care needs and higher mortality. These associations do not mean that delirium independently causes every adverse outcome, since it may also reflect the severity of the underlying illness.

Nevertheless, its consequences can be substantial. Patients experiencing delirium may have difficulty communicating symptoms, understanding treatment or participating in rehabilitation. Families may become distressed by sudden changes in cognition or behavior. Some patients continue to experience cognitive or functional difficulties after the acute illness has improved.

Not every episode is preventable, particularly among patients with severe illness or multiple contributing conditions. There are, however, opportunities to address modifiable risks, recognize changes earlier and investigate potentially reversible causes.

These responsibilities belong within routine hospital quality and patient-safety management, regardless of whether AI is available.

Predicting Delirium Is Not the Same as Diagnosing It

One of the most important distinctions in AI-assisted delirium care is the difference between predicting future risk and identifying a clinical syndrome that is already present.

A prediction model may use age, previous cognitive impairment, medical diagnoses, laboratory results, medication exposure and other clinical information to estimate the likelihood of delirium developing during hospitalization.

A high predicted risk does not establish that delirium is present. Some patients with multiple risk factors remain cognitively stable throughout admission, while others develop delirium after unexpected deterioration despite initially being classified as low-risk.

When a patient's mental state has already changed, the immediate question is no longer simply whether delirium might develop. It is whether an acute disturbance is present and what clinical conditions may be contributing.

Delirium must also be distinguished from dementia.

Dementia generally involves chronic or progressive cognitive impairment, whereas delirium develops acutely and commonly fluctuates. The two conditions may coexist, and people living with dementia are particularly vulnerable to delirium.

Knowledge of the patient's usual cognitive and functional abilities is therefore essential.

A change that seems minor to a clinician unfamiliar with the patient may be obvious to a family member or caregiver. Someone who normally recognizes relatives, participates in conversations and follows instructions may suddenly struggle to maintain attention.

Neither advanced age nor dementia establishes a delirium diagnosis. Equally, the absence of agitation does not exclude delirium.

Hospitals considering AI-assisted risk identification need to preserve these distinctions throughout their clinical workflows. Technology may identify patients who deserve attention, but diagnosis and management remain professional clinical responsibilities.

What International Guidelines Recommend

Several international guidelines provide a foundation for delirium prevention, assessment and management.

The National Institute for Health and Care Excellence (NICE) guideline CG103, Delirium: Prevention, Diagnosis and Management in Hospital and Long-Term Care, was updated in January 2023.

NICE recommends assessing delirium risk when people first present to hospital or long-term care. Relevant risk factors include age 65 years or older, cognitive impairment or dementia, current hip fracture and severe illness.

The guideline recommends observing patients for changes in cognition, perception, physical function and social behavior, with particular attention to hypoactive delirium, including withdrawal, slower responses, reduced movement and diminished engagement.

When indicators of delirium are identified, NICE recommends assessment using the 4AT by a practitioner competent to administer it.

In critical care or the recovery room after surgery, NICE recommends using the Confusion Assessment Method for the Intensive Care Unit (CAM-ICU) or the Intensive Care Delirium Screening Checklist (ICDSC) instead.

Where structured assessment indicates delirium, a healthcare professional with relevant expertise should establish the final diagnosis.

The American Psychiatric Association (APA) published its updated Practice Guideline for the Prevention and Treatment of Delirium in September 2025.

The guideline recommends regular structured assessment using valid and reliable measures, establishment of baseline cognitive status, review of contributing factors and individualized treatment planning. It also recommends multicomponent nonpharmacological interventions for patients with delirium or at risk of developing it.

These recommendations reinforce the importance of connecting risk identification with bedside assessment and preventive care. Neither guideline treats an AI-generated risk score as a substitute for clinical evaluation.

For hospitals in Viet Nam, international guidance can inform locally approved procedures, taking account of applicable Vietnamese professional requirements, workforce competencies and available resources.

Reliable Bedside Assessment Starts With the Patient's Baseline

Recognizing delirium often depends on knowing what is normal for the individual patient.

Without an established baseline, clinicians may find it difficult to distinguish newly developed cognitive changes from longstanding impairment.

Family members and caregivers can provide valuable information about how the patient usually communicates, recognizes people, manages daily activities and responds to familiar situations.

Nursing observations during hospitalization are equally important. A nurse may recognize that a patient who was communicating normally yesterday now struggles to maintain attention or becomes unusually withdrawn.

When delirium is suspected, the clinical team needs an appropriate structured assessment rather than relying solely on general impressions.

The 4AT evaluates alertness, orientation, attention and evidence of acute change or fluctuation. It supports rapid identification of possible delirium but is not an AI model, and its result does not independently establish the final diagnosis.

Findings must be interpreted alongside the patient's clinical condition, baseline cognitive status and relevant medical information.

For critically ill patients, CAM-ICU and ICDSC provide approaches suited to the intensive care environment.

Assessment in intensive care requires additional attention to sedation and level of arousal. A patient who cannot be assessed reliably because of deep sedation or unresponsiveness should not simply be recorded as delirium-negative.

Hospitals should establish how these circumstances are documented and when reassessment is appropriate.

Language and communication also affect assessment reliability. A validated instrument may not perform identically when translated informally or used with patients who have significant hearing impairment, aphasia or other communication difficulties.

Where Vietnamese-language versions of assessment tools are used, healthcare organizations should consider the available validation evidence, intended population and training necessary for reliable administration.

The usefulness of an assessment depends on both the instrument and the competence of the person applying it.

An AI alert may prompt clinical review, but staff should not wait for a positive prediction before assessing a patient who has become newly confused, less responsive or otherwise different from baseline.

Multicomponent Prevention Remains the Foundation of Care

Delirium prevention is best understood as coordinated care addressing several contributing factors rather than relying on a single intervention.

NICE recommends a tailored multicomponent prevention package for people at risk, including assessment of contributing factors within 24 hours of admission and interventions adapted to individual needs.

Relevant measures include supporting orientation, promoting appropriate sleep, maintaining hydration and nutrition, assisting mobility, managing pain, reviewing medications and ensuring that patients can use necessary hearing or visual aids.

These interventions must reflect the patient's condition and care setting.

An older patient who becomes disoriented in an unfamiliar ward may benefit from clear explanations, appropriate lighting, familiar objects and opportunities for family involvement.

Another patient may need attention to poorly controlled pain, dehydration, constipation or medication-related cognitive effects. A patient recovering from surgery may require an individualized mobilization plan that considers physiological stability, functional ability and fall risk.

These measures are not simply routine comforts. They are components of evidence-informed delirium prevention.

The APA 2025 guideline recommends multicomponent nonpharmacological care for both prevention and management.

Evidence supporting this approach has continued to develop.

A systematic review and meta-analysis published in Age and Ageing in March 2026 included 87 randomized trials involving 19,289 older adults across relevant healthcare settings, excluding ICU-only studies.

Within that review, an analysis of 17 comparisons involving multicomponent nonpharmacological interventions found lower delirium occurrence than usual care, with a pooled risk ratio of 0.56 and a 95% confidence interval of 0.45–0.72.

However, the certainty of evidence for this comparison was rated very low, and the interventions and patient populations varied.

Another systematic review and meta-analysis published in the Journal of Advanced Nursing in August 2026 examined 21 randomized and quasi-experimental studies involving 6,354 hospitalized older adults.

Nurse-led or nurse-coordinated multicomponent interventions were associated with reduced delirium incidence, with a pooled risk ratio of 0.64 and a 95% confidence interval of 0.58–0.70.

The authors assessed the overall evidence as moderate certainty and highlighted the potential value of structured prevention programs, including approaches based on the Hospital Elder Life Program.

A 2021 Cochrane review also found moderate-certainty evidence that multicomponent nonpharmacological interventions probably reduce delirium incidence in hospitalized adults outside intensive care.

Taken together, these studies support organized clinical prevention programs, although their different methodologies and levels of evidence certainty should be considered.

They do not demonstrate that AI independently improves delirium outcomes.

A hospital cannot attribute the benefits of multicomponent care to AI merely because a prediction system helps identify patients for the program. The additional contribution of the technology must be evaluated separately.

What AI May Contribute to Delirium Prevention

AI-assisted delirium risk identification commonly relies on information already recorded in hospital electronic systems.

Depending on the model, relevant variables may include age, previous cognitive impairment, medical diagnoses, severity of illness, laboratory findings, medication exposure and clinical observations.

Machine-learning systems may analyze combinations of these variables to estimate the probability of delirium developing within a specified period.

Natural language processing may also help identify relevant information contained in clinical notes but missing from structured fields.

For example, a previous history of cognitive impairment may appear in a referral letter without being entered into the coded diagnosis list. An AI system could identify that information and prompt the clinical team to verify it.

This may be useful for deciding which patients could benefit from additional assessment or a more structured prevention plan.

However, the intended function must be explicit.

A model developed to predict postoperative delirium should not automatically be used to diagnose delirium in a general medical ward. An ICU prediction model may not perform appropriately among patients undergoing routine elective surgery.

Likewise, a system estimating delirium risk over the next several days performs a different function from one identifying an acute change that may already represent delirium.

The hospital should define the patient population, prediction period, target outcome and clinical action expected from the system.

Without a clear action, an AI-generated risk score may provide little practical benefit.

A low-risk classification must also not become a justification for withholding basic preventive care. Good nursing assessment, appropriate communication, sleep support and management of modifiable clinical problems should not be reserved only for patients identified by an algorithm.

What the 2026 Evidence Shows About AI Prediction

A systematic review published in JMIR Medical Informatics on 16 September 2026 provides an important assessment of electronic health record-based delirium prediction.

Huang and colleagues examined 29 studies involving prediction models developed, validated or evaluated using routinely collected hospital data.

The studies addressed several prediction tasks, including early-admission risk assessment, postoperative delirium, dynamic ICU prediction and external validation of existing models.

Twenty studies used retrospective cohort designs, while 18 incorporated machine-learning or hybrid approaches.

Many models demonstrated promising discrimination. Among studies reporting internal discrimination, the area under the receiver operating characteristic curve ranged from 0.77 to 0.97.

These findings suggest that routinely collected hospital information can contain useful signals associated with delirium risk.

However, performance in retrospective datasets does not establish that a model is ready for routine clinical implementation.

Only 12 of the 29 studies reported external discrimination, 15 reported calibration, and three included decision curve analysis.

Although some studies included prospective evaluation or workflow integration, evidence concerning the effects of implementation on clinical decisions, preventive care and patient outcomes remained limited.

Risk-of-bias assessment classified eight studies as low risk, ten as unclear risk and eleven as high risk.

The review also found that more complex algorithms did not consistently outperform conventional statistical or rule-based approaches.

For healthcare organizations, these findings illustrate the difference between technical prediction and clinical usefulness.

A model may distinguish patients who will and will not develop delirium in a retrospective dataset without demonstrating that clinicians can use its predictions to improve care.

Performance may also change when a system is introduced into another hospital with different patient populations, documentation practices or methods of identifying delirium.

Calibration is particularly important.

A model that systematically overestimates absolute risk may generate excessive alerts and unnecessary workload. One that underestimates risk may leave vulnerable patients without appropriate attention.

False-negative findings are especially concerning when they reinforce existing failures to recognize hypoactive delirium.

The systematic review supports continued development and careful evaluation of delirium prediction systems, but it does not establish that any single algorithm is ready for unrestricted clinical adoption.

A Randomized Trial Highlights the Implementation Challenge

One of the most relevant studies was published in Anesthesia & Analgesia on 9 October 2026.

Scharp and colleagues conducted a pragmatic randomized controlled trial evaluating AI-driven clinical decision support intended to improve postoperative delirium prevention.

The trial took place at Mount Sinai Hospital in New York City.

Patients with evidence of preoperative cognitive impairment were identified using an approach involving natural language processing, cognitive assessment information and diagnostic codes.

The researchers randomized 7,294 patients to either a noninterruptive electronic clinical decision-support intervention or routine perioperative care.

The intervention was intended to encourage anesthesia teams to follow established practices associated with delirium prevention. It was not a trial demonstrating that an AI risk score could independently diagnose delirium.

The findings illustrated an important implementation problem.

Only 1,255 of the 7,294 randomized patients, or 17.2%, had documented postoperative delirium assessments.

Among those assessed, 426 developed delirium.

Delirium was documented in 225 of 641 assessed patients in the intervention group and 201 of 614 assessed patients in the control group, corresponding to 35.1% and 32.7%, respectively.

The difference was not statistically significant.

The intervention also did not demonstrate significant improvements in adherence to the twelve perioperative practices evaluated.

However, the low proportion of patients with documented delirium assessments substantially limited interpretation of the primary outcome.

The researchers did not conclude that AI-driven clinical decision support had definitively been proven ineffective. Instead, they highlighted inconsistent postoperative delirium assessment and the difficulty of translating electronic recommendations into changes in clinical practice.

For hospitals considering similar technology, this finding is important.

An AI system may identify patients at risk and provide useful recommendations, but its impact cannot be assessed reliably when the organization does not consistently perform the assessments needed to establish patient outcomes.

Likewise, guidance delivered electronically may have limited value when it does not fit clinical routines or change professional behavior.

The practical challenge is not simply making AI available. It is ensuring that assessment, preventive actions and outcome measurement are consistently performed.

Clinical Response Must Focus on Underlying Causes

When delirium is suspected or confirmed, clinical assessment should identify and address contributing conditions.

Delirium is a syndrome with multiple possible causes, including infection, hypoxemia, metabolic disturbances, dehydration, pain, urinary retention, constipation, medication effects and withdrawal states.

Acute neurological conditions, including stroke and seizures, may also present with changes in attention, awareness or behavior.

The urgency of assessment depends on the patient's condition.

Sudden focal neurological deficits, seizures, severe hypoxemia, marked physiological instability or another potentially life-threatening condition require appropriate emergency evaluation.

An AI system should not delay that response by requiring completion of a routine delirium screening workflow.

When delirium is identified, clinicians should review recent clinical changes, medication exposure, physiological abnormalities and potentially reversible contributors.

The management plan must reflect the underlying problem.

For example, delirium associated with an acute infection requires appropriate investigation and treatment of the infection alongside supportive care.

Medication-related cognitive disturbance may require review of the implicated medicine, its indication, dose and available alternatives.

Withdrawal from alcohol or certain medications may require a specific clinical management pathway.

These situations should not be treated identically.

An AI-generated risk classification does not determine the underlying cause, and a positive screening result does not establish the appropriate treatment.

The technology may support recognition and communication, while diagnosis, investigation and management remain the responsibility of qualified healthcare professionals.

Medication and Restraint Decisions Require Particular Caution

Delirium can create difficult clinical situations when patients become frightened, distressed, disoriented or behave in ways that present an immediate safety risk.

The response must balance safety, dignity and the patient's clinical needs.

An AI system should not automatically recommend sedation, antipsychotic medicines or physical restraints simply because a patient is classified as high-risk.

The APA 2025 guideline recommends multicomponent nonpharmacological interventions as the foundation of prevention and management.

It advises against using antipsychotic medications to prevent delirium or hasten its resolution.

When medication is considered for severe neuropsychiatric disturbances, clinicians should assess contributing causes, attempt appropriate de-escalation strategies and consider whether the patient is experiencing significant distress or poses a risk of physical harm.

Benzodiazepines should generally not be used for delirium unless a specific clinical indication exists, such as certain withdrawal syndromes.

These distinctions matter because sedation may reduce visible agitation without resolving delirium or its underlying causes. A quieter patient is not necessarily a patient whose delirium has improved.

Physical restraints require equally careful consideration.

Restrictive interventions may cause distress, injury, immobility and other harms. They should not be treated as routine responses to confusion, wandering or an elevated AI-generated risk score.

The APA guideline recommends avoiding physical restraints except in circumstances involving imminent injury to the patient or others, with appropriate monitoring and repeated reassessment.

Any restrictive intervention must also comply with applicable Vietnamese legal and professional requirements.

For AI governance, a predictive score must not become an automated justification for potentially harmful interventions.

The clinical team must understand why the patient's behavior has changed and select the least restrictive appropriate response.

Delirium in Intensive Care Requires a Distinct Approach

ICU delirium presents additional challenges because patients may be critically ill, mechanically ventilated or receiving sedative and analgesic medications.

Assessment must account for sedation depth and level of arousal.

CAM-ICU and ICDSC are commonly used structured assessment tools in this setting, while sedation assessment helps determine whether delirium evaluation can be performed and interpreted reliably.

The Society of Critical Care Medicine (SCCM) published a focused update to its PADIS guidelines in February 2025.

The update addresses selected issues involving anxiety, sedation, delirium, mobility and sleep in critically ill adults.

It conditionally suggests dexmedetomidine over propofol for sedation in mechanically ventilated adults when light sedation or delirium reduction is a high priority.

It also conditionally supports enhanced mobilization and rehabilitation compared with usual care.

For antipsychotic treatment of ICU delirium, the panel found insufficient evidence to recommend either for or against use over usual care.

These recommendations require interpretation within the relevant ICU population and clinical circumstances.

The recommendation concerning dexmedetomidine is not an instruction to administer it to every patient at risk of delirium, nor should it become an automatic medication order generated by a prediction system.

ICU prediction also raises additional methodological questions.

Sedation depth, mechanical ventilation, medication exposure and physiological measurements may reflect illness severity, treatment decisions or both.

Models trained on these variables may behave differently when sedation practices or clinical protocols change.

Hospitals should therefore avoid transferring prediction thresholds developed in general wards directly into intensive care without appropriate validation.

The Vietnamese Healthcare Context

Delirium prevention must fit the professional responsibilities and resources of Vietnamese healthcare organizations.

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

The Circular emphasizes comprehensive, continuous, safe and individualized nursing care, including clinical assessment, identification of patient needs and risks, nursing care planning, monitoring and evaluation.

These responsibilities are relevant to recognizing changes in mental status, providing supportive care and communicating clinical concerns. The Circular does not establish a national AI delirium prediction standard.

Ministry of Health Circular No. 43/2018/TT-BYT on prevention of medical incidents and Circular No. 19/2013/TT-BYT on hospital quality management provide additional organizational foundations for patient safety and continuous quality improvement.

These regulations should not be presented as mandating AI-based delirium prediction or a specific delirium assessment instrument.

Local research provides useful context.

A study published in Y Dược lâm sàng 108 in 2022 examined nurses' knowledge and practice concerning delirium assessment in intensive care units across the Vinmec hospital system.

The researchers conducted a cross-sectional survey involving 95 nurses between 15 February and 15 April 2022.

The study reported that 61% of respondents assessed for delirium when caring for critically ill patients, while 70% had never heard of or had not used a delirium assessment tool.

Thirty-six percent considered delirium assessment tools too complicated for everyday practice.

The researchers concluded that although awareness of delirium was relatively satisfactory, routine assessment practices and use of standardized tools remained inadequate.

These findings highlight a possible gap between professional awareness and routine bedside assessment.

However, the study represents a particular healthcare system and period. Its findings should not be interpreted as national estimates of delirium assessment quality or AI readiness in Vietnamese hospitals.

More broadly, Vietnamese healthcare organizations vary in staffing, digital infrastructure, specialty services and documentation practices.

Some hospitals have integrated electronic records and standardized nursing assessments. Others rely on combinations of digital systems, narrative clinical notes and paper documentation.

Information about a patient's usual cognitive status may be incomplete, particularly when previous care occurred elsewhere.

Family members and caregivers may therefore contribute valuable information about baseline cognition and recent changes. Their observations should complement professional assessment rather than replace it.

These practical conditions matter when introducing AI.

A model developed using comprehensive structured nursing observations may not perform reliably in a hospital where such assessments are inconsistently recorded.

If the principal problem is poor bedside recognition, staff education, appropriate assessment tools and clearer clinical responsibilities may provide greater immediate benefit than purchasing a prediction model.

Technology should address an identified care problem rather than substitute for the underlying care process.

Designing a Reliable AI-Assisted Clinical Workflow

A hospital considering AI-assisted delirium risk identification should begin by defining the patient population and the clinical action expected from the system.

A reasonable initial setting may be adult medical or surgical inpatient care, where staff already assess patients with recognized delirium risk factors.

The system could support identification of patients who may benefit from targeted assessment or review of preventive care.

However, it should not replace routine observation for acute mental status changes or discourage assessment in patients classified as low-risk.

Professional responsibilities need to remain explicit.

Nursing staff may identify changes in attention, behavior, sleep or responsiveness during routine care.

The responsible medical team assesses possible causes and determines the diagnostic and treatment plan.

Pharmacists may contribute to medication review, while rehabilitation professionals and other team members support mobility and functional recovery.

Families and caregivers may provide essential information about baseline cognition and recent changes.

The organization needs a process that coordinates these contributions.

An AI alert should reach a professional who can review it and arrange an appropriate response.

If notifications appear only on a separate dashboard that staff rarely use, the system may offer little benefit. If alerts are frequent but rarely relevant, they may contribute to alert fatigue.

Likewise, if the technology generates more assessments without improving preventive care or clinical action, it may increase workload without meaningful benefit.

An effective workflow should connect risk identification to a defined care process.

For patients at increased risk, this may include review of contributing factors, individualized prevention planning and ongoing observation.

For patients with new signs of delirium, the pathway shifts toward timely structured assessment, clinical evaluation and management of underlying causes.

These pathways must remain distinct.

A high predicted risk does not establish delirium, and a low predicted risk must not delay assessment of a new clinical problem.

Local Validation Must Reflect Real Clinical Conditions

Before deployment, healthcare organizations need to determine whether an AI model performs reliably using representative local data.

The first question is whether the intended outcome matches the clinical task.

A model developed to predict postoperative delirium may have limited relevance to general medical inpatients. A model predicting delirium during ICU admission may be unsuitable for patients receiving routine ward care.

The hospital should also examine how delirium outcomes were established in the model's development data.

Retrospective studies sometimes rely on diagnostic codes or clinical notes to identify delirium.

These sources may be incomplete because delirium is frequently underrecognized or inadequately documented.

A model trained on such outcomes may learn patterns associated with documented delirium rather than the full range of clinically occurring cases.

This creates a risk that technical performance estimates will be misleading.

The timing of predictor information also requires attention.

A model intended to provide early warning should not use information recorded after delirium has already developed.

Such information leakage can make retrospective results appear better than performance during actual clinical use.

Local validation should use an appropriate reference process involving qualified clinical assessment and clearly defined outcome criteria.

Evaluation should examine discrimination, calibration, clinically relevant thresholds and the proportion of alerts requiring professional review.

False-negative cases deserve particular attention.

A model that systematically misses hypoactive delirium or patients with incomplete documentation may reinforce existing recognition gaps.

Performance should also be examined across relevant patient groups, including older adults, people with dementia, patients with different communication needs and those treated in different departments.

Validation should not rely solely on an overall accuracy statistic.

The organization needs to determine whether the system produces information that can be acted upon safely and whether its use improves existing clinical practice.

An initial supervised or silent evaluation should preserve standard care while model performance, limitations and possible unintended effects are investigated.

Measuring Whether Delirium Care Actually Improves

The effectiveness of delirium prevention should be assessed through measures reflecting both clinical practice and patient outcomes.

Reliable assessment is the starting point.

Hospitals should determine which patients are eligible for structured delirium assessment, when reassessment is clinically indicated and how completion is documented.

They should distinguish between patients assessed without evidence of delirium, patients with positive findings and those for whom assessment could not be completed reliably.

Missing assessment information should not be interpreted as absence of delirium.

This is particularly important in light of the 2026 randomized trial, where incomplete postoperative assessments substantially limited interpretation of intervention effectiveness.

Process measures may examine whether baseline cognition is documented, whether changes from baseline are recognized and whether suspected delirium leads to timely clinical evaluation.

Prevention measures may include implementation of relevant multicomponent interventions, medication review, sleep support, mobilization and attention to sensory needs.

However, completion should reflect clinically appropriate care rather than indiscriminate application of every intervention to every patient.

Outcome measures may include newly identified delirium during hospitalization, episode duration, relevant complications, functional recovery and other clinically important outcomes where reliable information is available.

Balancing measures are also necessary.

These may include unnecessary antipsychotic use, excessive sedation, avoidable physical restraints, additional workload, inappropriate alerts and delays in evaluating other acute conditions.

An increase in documented delirium after introducing structured assessment may reflect better recognition rather than a true increase in clinical incidence.

Conversely, a reduction in recorded delirium may be misleading if assessment becomes less consistent.

Outcome changes should therefore be interpreted alongside assessment coverage, documentation quality, patient characteristics and clinical practice.

For AI-assisted workflows, evaluation should additionally examine whether risk predictions are reviewed appropriately, whether alerts influence prevention planning and whether important cases are missed.

A system that predicts delirium accurately but does not improve assessment, prevention or clinical response may offer limited practical value.

The objective is to determine whether the complete intervention improves care compared with the hospital's previous approach.

Supporting Recovery and Safe Transitions of Care

Delirium management does not necessarily end when the patient leaves the hospital.

Some patients recover quickly, while others continue to experience cognitive or functional difficulties after the acute illness has improved.

The APA 2025 guideline recommends attention to continuing assessment, medication review, possible cognitive consequences and education for patients and caregivers during transitions of care.

This is particularly important when patients return home or move to another healthcare facility.

Discharge documentation should communicate relevant information about the delirium episode, possible contributing causes, treatment changes and any unresolved concerns.

If the patient has not fully recovered, the receiving clinician or healthcare service needs an appropriate follow-up plan.

Depending on clinical circumstances, this may involve reassessment of cognition, medication review, rehabilitation or additional specialist evaluation.

Families should receive understandable information about what occurred and which subsequent changes warrant medical attention.

Persistent cognitive impairment after hospitalization should not automatically be diagnosed as dementia without appropriate assessment.

Delirium, pre-existing cognitive impairment and subsequent cognitive difficulties require careful clinical interpretation.

AI may eventually support identification of unresolved follow-up needs, but this would represent a separate function requiring appropriate evaluation.

The immediate priority is to ensure that relevant information and clinical responsibility are not lost during transitions of care.

Legal and Professional Responsibilities in Viet Nam

AI-assisted delirium risk identification must operate within Viet Nam's applicable healthcare, AI, 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 professional medical practice, patient rights, clinical records and healthcare organizations' responsibilities.

For hospital nursing, Circular No. 31/2021/TT-BYT provides relevant requirements concerning assessment, individualized care, coordination, monitoring and documentation.

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

These regulations support organizational responsibilities for safe care and continuous quality improvement.

They should not be presented as establishing a specific AI-based delirium prediction requirement or mandating the use of a particular assessment instrument.

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.

An AI application intended to predict delirium risk or assist clinical documentation should not automatically be assigned a particular legal classification solely because it is used in a hospital.

The applicable classification and obligations depend on the system's intended purpose, actual functions, degree of automation, deployment circumstances and relevant legal criteria.

Clinical risk must also be considered independently.

A system that misses delirium, produces misleading classifications or encourages inappropriate sedation may create substantial patient-safety concerns regardless of its formal AI classification.

Healthcare organizations must also determine whether a particular product falls within the 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 concerning medical device risk determination and related regulatory requirements.

A tool used solely to organize information for professional review may raise different regulatory considerations from software intended to provide medical predictions or treatment recommendations.

The presence of AI alone does not establish medical device status. Conversely, describing software as an administrative assistant does not eliminate the need to assess its actual medical purpose.

A product-specific regulatory assessment should therefore be conducted before procurement, commercialization or routine clinical deployment.

International guidelines may inform local clinical practice, but they do not automatically establish binding obligations under Vietnamese law.

Similarly, organizational safeguards proposed for AI-assisted delirium care should be distinguished from specific statutory requirements unless an applicable legal provision establishes them.

Protecting Patient Information, Rights and Dignity

AI-assisted delirium risk identification may involve identifiable information about diagnoses, cognitive assessments, medications, physiological observations and clinical documentation.

These data require appropriate protection.

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 personal data processing and protection.

Healthcare organizations should determine the applicable lawful basis and purposes for processing, restrict access to authorized personnel and implement appropriate security and retention arrangements.

Where external AI providers are involved, the organization should assess contractual responsibilities, data access, hosting, secondary use and cross-border processing requirements.

Patient records should not automatically be reused for unrelated commercial model training or development without assessing the applicable legal basis and required safeguards.

Delirium also raises additional concerns about a patient's ability to understand information or participate in decisions during periods of impaired cognition.

Healthcare professionals must continue to respect patient rights and apply relevant legal and professional requirements concerning decision-making, communication and the involvement of legally authorized representatives where appropriate.

The existence of an AI-generated risk score does not justify unnecessary restrictions on liberty, routine intrusive surveillance or disregard for patient preferences and dignity.

Technology should support individualized care while preserving professional accountability and patients' rights.

A Practical Starting Point for Vietnamese Hospitals

A hospital considering AI-assisted delirium risk identification should begin by examining its current assessment and prevention practices.

An initial review may focus on older adults admitted to a medical or surgical department.

The organization can examine whether baseline cognitive status is documented, whether acute changes are recognized and whether appropriate structured assessment occurs when delirium is suspected.

It should also review whether relevant preventive measures are routinely available and whether staff know how to respond to new cognitive or behavioral changes.

The findings may reveal opportunities that do not require AI.

Inconsistent delirium recognition may respond to staff education, standardized assessment procedures and greater involvement of families in establishing baseline cognition.

A ward with poor sleep conditions, limited mobilization support or incomplete medication review may benefit from strengthening those practices before introducing predictive technology.

These improvements should not be postponed while an AI system is evaluated.

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

One possible approach is evaluating whether an electronic health record-based model can identify adult inpatients who may benefit from targeted delirium prevention review.

The intended patient population, prediction timing, available data and expected clinical actions should be specified before testing begins.

An initial silent evaluation can assess model performance without allowing predictions to influence live clinical decisions.

This provides an opportunity to examine missed cases, unnecessary alerts, calibration and data quality while established care continues.

If the results justify further evaluation, a supervised pilot may be considered.

The pilot should preserve routine clinical observation, structured delirium assessment when indicated and established treatment responsibilities.

It should also define who reviews AI-generated information, how important findings are communicated and when urgent clinical evaluation is required.

Nursing leadership, medical teams, geriatrics, psychiatry or other relevant specialists, pharmacy, information technology and quality management may contribute according to the hospital's organization and capabilities.

Implementation should be reviewed regularly.

If the system generates excessive alerts, misses clinically important cases or fails to improve preventive care, the organization should have authority to modify, restrict or discontinue its use.

Expansion to additional departments should depend on appropriate evidence and local readiness.

A model evaluated in general medical wards should not automatically be used in intensive care, postoperative recovery, pediatric care or other populations without considering further validation.

Throughout implementation, clinical assessment and treatment must remain available when the AI system is unavailable.

The objective is to strengthen the hospital's delirium prevention and management process, not to make essential care dependent on a prediction model.

From Predicting Delirium to Protecting Cognitive Safety

Delirium illustrates why responsible AI implementation must remain grounded in clinical care rather than technological capability alone.

AI may help healthcare organizations recognize patients with increased vulnerability, identify overlooked information and direct professional attention toward opportunities for prevention.

Research published through 2026 demonstrates that electronic health record-based delirium prediction is technically feasible.

However, the available evidence does not establish that such models consistently reduce delirium incidence or improve patient outcomes when deployed in routine hospital practice.

At the same time, international guidelines and systematic reviews support structured assessment and multicomponent nonpharmacological prevention as important elements of safer care.

The 2026 randomized trial provides an additional lesson: a clinical decision-support system may fail to demonstrate benefit when routine assessment and implementation processes are incomplete.

For hospitals in Viet Nam, the most appropriate approach is therefore to strengthen existing nursing and medical care, establish clear responsibilities and introduce AI only where it offers a plausible and measurable additional benefit.

A patient who becomes quieter, less responsive or newly confused must receive appropriate clinical attention regardless of the AI prediction.

A patient identified as high-risk needs preventive care tailored to modifiable factors, not an automatic diagnosis or medication order.

When delirium develops, the clinical team must investigate contributing causes, provide appropriate treatment and protect the patient's safety and dignity.

Success should not be measured by the number of patients assigned a high AI risk score or the number of alerts generated.

It should be judged by whether delirium is recognized reliably, preventive care is delivered consistently, clinical deterioration is addressed promptly and patients receive safe, person-centered care throughout hospitalization and recovery.

That is where AI-assisted delirium risk identification may make a meaningful contribution to healthcare quality, cognitive safety and patient outcomes.

Selected References and Regulatory Sources

International Clinical Guidelines

National Institute for Health and Care Excellence. Delirium: Prevention, Diagnosis and Management in Hospital and Long-Term Care. NICE Clinical Guideline CG103. Published 2010; updated January 2023.

American Psychiatric Association. The American Psychiatric Association Practice Guideline for the Prevention and Treatment of Delirium. American Journal of Psychiatry. 2025;182(9):880–884. DOI: 10.1176/appi.ajp.25182013.

Lewis K, Balas MC, Stollings JL, et al. A Focused Update to the Clinical Practice Guidelines for the Prevention and Management of Pain, Anxiety, Agitation/Sedation, Delirium, Immobility, and Sleep Disruption in Adult Patients in the ICU. Critical Care Medicine. 2025;53(3)–e727.

Society of Critical Care Medicine. Clinical Practice Guidelines for the Prevention and Management of Pain, Agitation/Sedation, Delirium, Immobility, and Sleep Disruption in Adult Patients in the ICU. 2018, with focused update published in 2025.

Clinical AI and Implementation Evidence

Huang HM, Lu CS, Chang GW, Tien MH, Hsu YK. Prediction Models for In-Hospital Delirium Using Routinely Collected Electronic Health Record Data: Systematic Review. JMIR Medical Informatics. 2026;14. DOI: 10.2196/91618.

Scharp D, Meyers K, Nadkarni G, et al. Artificial Intelligence-Driven Clinical Decision Support for Postoperative Delirium Prevention: A Pragmatic Randomized Controlled Trial. Anesthesia & Analgesia. Published online 9 October 2026. DOI: 10.1213/ANE.0000000000008318.

Multicomponent Delirium Prevention Evidence

Sánchez A, Sawant-Uttekar P, Heinen D, et al. Effectiveness of Non-Pharmacological and Pharmacological Interventions in Preventing Delirium in Older Adults: A Systematic Review and Meta-Analysis of Randomised Controlled Trials. Age and Ageing. 2026;55(3). DOI: 10.1093/ageing/afag068.

Shaban M, Mohammed HH, Osman YM, Shaban MM. Nurse-Led Multicomponent Interventions for Delirium Prevention in Hospitalised Older Adults: A Systematic Review and Meta-Analysis. Journal of Advanced Nursing. Published online 4 August 2026. DOI: 10.1111/jan.70706.

Burton JK, Craig L, Yong SQ, et al. Non-Pharmacological Interventions for Preventing Delirium in Hospitalised Non-ICU Patients. Cochrane Database of Systematic Reviews. 2021;11. DOI: 10.1002/14651858.CD013307.pub3.

Vietnamese Clinical Evidence and Professional Guidance

Nguyen Thi Thuy, Nguyen Thi Hoa Huyen, Hoang Lan Van, Vu Thi Hong, Tran Thi Thuy Dung. Nhận thức và thực hành của điều dưỡng trong đánh giá sảng (Delirium) tại Đơn vị Chăm sóc tích cực của Bệnh viện Đa khoa quốc tế Vinmec. Y Dược lâm sàng 108. 2022;17(7). DOI: 10.52389/ydls.v17i7.1556.

Ministry of Health of Viet Nam. Circular No. 31/2021/TT-BYT, dated 28 December 2021, on nursing activities in hospitals.

Ministry of Health of Viet Nam. Circular No. 43/2018/TT-BYT, dated 26 December 2018, on prevention of medical incidents in healthcare facilities.

Ministry of Health of Viet Nam. Circular No. 19/2013/TT-BYT, dated 12 July 2013, on hospital quality management.

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

Relevant Vietnamese AI, Data Protection and Medical Device Framework

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 international clinical guidelines, published research, Vietnamese professional requirements and relevant legal and regulatory sources.

The clinical scenarios are illustrative and do not represent documented individual patient cases.

The approaches discussed for AI-assisted delirium risk identification, bedside assessment, multicomponent prevention, clinical response 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 or legal advice, authorize automated medication or restraint decisions, replace professional clinical judgment or determine the regulatory status of any particular AI product.

Delirium assessment, diagnosis, prevention, investigation of underlying causes, treatment and follow-up remain the responsibility of appropriately qualified healthcare professionals, consistent with applicable Vietnamese requirements, current clinical guidance and locally approved procedures.

AI-generated delirium risk predictions must not replace structured clinical assessment, delay evaluation of acute mental status changes or justify unnecessary sedation and restrictive interventions.

The research discussed does not establish that AI-assisted inpatient delirium prediction necessarily reduces delirium incidence, mortality, functional decline or healthcare costs in Vietnamese hospitals. Appropriate local validation, meaningful clinical oversight, supervised implementation 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 obligations under Vietnamese law.

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