Governing AI-Assisted Preoperative Risk Assessment in Viet Nam: Clinical Review, Patient Communication and Safe Surgical Planning
IMPLEMENTATION INSIGHTS
Governing AI-Assisted Preoperative Risk Assessment in Viet Nam: Clinical Review, Patient Communication and Safe Surgical Planning
IMPLEMENTATION INSIGHTS
Last reviewed: 23 October 2026
A patient is preparing for planned abdominal surgery. During the preoperative assessment, an artificial intelligence system reviews the patient's medical history, laboratory results and information about the proposed procedure. It estimates an elevated probability of postoperative complications.
The result raises concerns. The patient has anemia and several chronic medical conditions, but the operation is intended to treat a disease that could become more difficult to manage if surgery is delayed.
The surgeon needs to consider whether the predicted risk changes the balance between the expected benefits and potential harms of proceeding. The anesthesiologist must assess the patient's condition, identify anesthesia-related concerns and determine whether further evaluation or preparation is appropriate. The patient wants to understand what the prediction means and whether surgery remains a reasonable choice.
A numerical risk estimate cannot answer these questions on its own.
The prediction may identify concerns that deserve closer attention, but it does not establish whether surgery should proceed, which additional investigations are necessary or whether the patient will need intensive care afterward.
Those decisions depend on the patient's actual clinical condition, the proposed procedure, the expected benefits of treatment, available alternatives and the consequences of delaying or not performing surgery.
For hospitals in Viet Nam, AI-assisted preoperative risk assessment offers an opportunity to strengthen clinical evaluation and surgical planning. Its value, however, depends on how the information is interpreted, communicated and used by the healthcare professionals responsible for the patient's care.
Predicting postoperative complications is not the same as deciding whether surgery is appropriate.
A model may identify patients with an increased probability of adverse outcomes without establishing which interventions will reduce that risk. A higher predicted probability also does not automatically justify postponing surgery, refusing an operation or reserving an intensive care bed.
The challenge for healthcare organizations is to ensure that AI contributes useful information without turning uncertain predictions into automatic treatment decisions.
What preoperative risk assessment should achieve
Preoperative assessment helps clinicians understand the patient's condition, identify problems that may influence surgical and anesthesia care, and prepare an appropriate management plan.
The process involves more than estimating the probability of complications. Clinicians must consider the disease being treated, the nature of the proposed operation, existing medical conditions, functional capacity, relevant investigations and the patient's preferences.
For some patients, the assessment may identify a previously unrecognized condition that warrants attention before surgery. For others, it may confirm that the planned procedure can reasonably proceed without additional investigations or changes to the care pathway.
The extent of evaluation should reflect the patient's circumstances and the proposed surgery. More testing does not automatically mean better preparation, particularly when the results are unlikely to influence clinical management.
The European Society of Anaesthesiology and Intensive Care (ESAIC) published updated guidance on preoperative assessment of adults undergoing elective noncardiac surgery in January 2025 [1]. The guidance emphasizes structured clinical evaluation, appropriate investigations, optimization and organizational preparation, including the importance of identifying relevant concerns early enough to address them.
The American Heart Association and American College of Cardiology provide a related approach to perioperative cardiovascular management.
On September 29, 2026, the AHA/ACC published a reaffirmed guideline for cardiovascular management of patients undergoing noncardiac surgery [2]. This publication replaces the 2024 edition following a review of evidence available through March 2026. Importantly, it retains the recommendations, supporting text, figures, tables and references of the earlier guideline rather than introducing a new set of clinical recommendations.
The guideline supports a stepwise approach to cardiovascular evaluation, helping clinicians distinguish situations where surgery can proceed from those where further assessment may be warranted. It also emphasizes team-based care and patient engagement.
The 2022 European Society of Cardiology guidelines similarly recommend considering patient-related risk factors, surgical risk, functional capacity and relevant clinical findings when planning noncardiac surgery [3].
Together, these recommendations demonstrate why risk estimation is only one part of preoperative care. The clinical question is not simply how likely a complication is to occur, but what that information means for the patient's treatment and preparation.
This article focuses on adults preparing for planned, non-emergency, predominantly noncardiac surgery. It addresses assessment and decisions made before the operation rather than autonomous anesthesia management, AI-assisted surgical procedures or prediction systems that rely on information collected during surgery.
Traditional risk assessment tools are not necessarily AI
Preoperative risk prediction has been part of surgical practice for decades. Several established instruments use clinical information and statistical methods to estimate the likelihood of complications.
These tools remain relevant, but their purposes and limitations differ.
The Revised Cardiac Risk Index (RCRI), developed by Lee and colleagues in 1999, is widely used to estimate the risk of major perioperative cardiac complications in patients undergoing noncardiac surgery [4]. It uses six clinical predictors involving selected cardiovascular conditions, insulin-treated diabetes, renal dysfunction and the type of operation.
Its relative simplicity makes it practical in many settings. However, RCRI primarily addresses cardiovascular risk and should not be treated as a comprehensive measure of all postoperative complications.
The American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP) Surgical Risk Calculator serves a broader purpose.
The original calculator, developed by Bilimoria and colleagues, uses preoperative patient characteristics and procedure-specific information to estimate the likelihood of several postoperative outcomes, including complications and mortality [5].
The calculator continues to be maintained and updated. As of October 2026, the published ACS NSQIP Surgical Risk Calculator is version 4.0.4, with its most recent parameter update in April 2026 [6].
Its estimates are derived from data involving patients who underwent comparable procedures. They can support discussions about possible outcomes, but they remain estimates rather than guarantees for an individual patient.
The American Society of Anesthesiologists Physical Status Classification System (ASA-PS) has a different function. It provides a standardized description of a patient's pre-anesthesia medical status.
The ASA statement revised in October 2025 and published in Anesthesiology Open in April 2026 explains that ASA-PS helps communicate medical comorbidities and contributes to perioperative risk assessment when considered alongside factors such as surgery type, frailty and functional condition [7].
The final ASA physical status classification is a clinical determination made by the anesthesiologist on the day of anesthesia care, following evaluation of the patient. It is not an AI-generated prediction and should not be regarded as a comprehensive risk model on its own.
These distinctions matter when hospitals compare established tools with newer machine learning systems.
RCRI focuses on selected cardiac complications. ACS NSQIP estimates several postoperative outcomes. ASA-PS describes physical status. The information they provide is related but not interchangeable.
A calculator should not automatically be described as artificial intelligence simply because it produces a numerical probability.
Newer machine learning models may use more complex relationships between variables and larger datasets. However, greater computational complexity does not necessarily mean greater clinical value.
A useful comparison should establish whether the newer model provides reliable information beyond existing approaches, using comparable patient populations and outcomes, and whether that additional information helps clinicians make more appropriate decisions.
What newer AI models may contribute
Machine learning models can analyze combinations of clinical variables and identify patterns that may be difficult to represent through simpler prediction methods.
Depending on their design, they may use demographic information, medical history, functional status, laboratory results and details of the proposed procedure.
Some models estimate postoperative cardiac complications, while others focus on mortality, respiratory complications, acute kidney injury, unplanned ICU admission or composite measures of postoperative morbidity.
Their potential contribution is to provide clinicians with additional information when evaluating an individual patient's risk.
However, the quality and clinical applicability of the evidence vary.
In June 2026, Dhaliwal and colleagues published a systematic review and Bayesian network meta-analysis comparing machine learning approaches with traditional methods for predicting postoperative cardiac complications following noncardiac surgery [8].
The review included 13 studies evaluating 54 models and 927,113 patients. Machine learning approaches generally demonstrated better discrimination than traditional risk scores, with automated machine learning and gradient boosting methods performing particularly well in the analyses.
These findings suggest that newer models may improve the prediction of selected cardiovascular outcomes.
However, none of the included studies externally validated its machine learning models, and only six studies were considered to have a low risk of bias. Questions also remained about calibration and clinical applicability.
The findings therefore support further evaluation rather than an assumption that better statistical performance is sufficient for routine clinical adoption.
Another study published in Scientific Reports in June 2026 examined a model designed for late-preoperative risk assessment using electronic health record data [9].
Yang and colleagues evaluated the model in a retrospective cohort of 58,630 adults undergoing major or complex surgery across six clinical sites in China.
The researchers defined a specific prediction point after final surgical scheduling and assignment of the surgical team, but before the operation began. The model used information available at that time, including patient, disease, procedure and scheduling characteristics.
It demonstrated encouraging predictive performance in a geographically separate validation cohort. However, the validation site remained affiliated with the broader development network, and the study did not establish that using the model improved clinical decisions or patient outcomes.
This distinction matters. A model can predict complications accurately within the datasets used for evaluation without demonstrating that its introduction will make surgical care safer.
For hospitals, the question is not only whether the model produces a better estimate. It is whether that estimate provides information that clinicians can use appropriately within the actual preoperative pathway.
The timing of a prediction matters
One of the most important issues in preoperative AI assessment is identifying when the prediction will be generated.
A patient may undergo several assessments before surgery, including the initial surgical consultation, preoperative investigations, anesthesia assessment, specialist review and final evaluation before the operation.
The information available at these stages can differ substantially.
A model intended to support discussion when surgery is first proposed must use information available at that time. It cannot legitimately depend on laboratory results that have not yet been obtained or on clinical decisions made later in the pathway.
A model used after surgery has been scheduled may have access to updated investigations, a confirmed surgical approach and other information relevant to planning.
This distinction becomes particularly important when models are developed from electronic health records.
Retrospective datasets often contain information collected throughout a hospital admission. Without careful attention to timing, a model may use variables that were not available when the intended clinical decision occurred.
This can create information leakage and make the model appear more accurate than it would be during actual preoperative use.
A study published in JMIR Perioperative Medicine on October 21, 2026, illustrates why prediction timing matters [10].
Nishino and colleagues evaluated models using data from 35,702 adults undergoing noncardiac surgery under general anesthesia at three Japanese hospitals. Models incorporating intraoperative and early postoperative information performed better than those using preoperative information alone.
This finding demonstrates the predictive value of additional information. However, variables collected during or after surgery cannot be used to support a decision that must be made beforehand.
The study also reported variation in calibration between facilities, highlighting the difficulty of transferring absolute risk estimates across clinical settings.
For an application intended to support preoperative decisions, evaluation must be restricted to information genuinely available before the decision being made.
Hospitals should therefore define the intended prediction point, permitted inputs and approach to missing information.
The organization should also be able to distinguish when clinical information was recorded from when it actually became available to the team. Data entered retrospectively into an electronic record may not have been accessible at the time of the original decision.
Predictions should be reviewed when important clinical information changes.
A risk estimate generated several weeks before surgery may no longer be appropriate if the patient develops a new illness, requires a different procedure or experiences a significant change in medical condition.
The prediction must remain connected to the patient's actual circumstances rather than becoming a fixed number carried through the entire surgical pathway.
A higher predicted risk does not automatically justify postponing surgery
An elevated probability of postoperative complications may justify closer clinical review. It does not, by itself, establish that surgery should be delayed or avoided.
The decision must consider why surgery is needed, what benefits are expected and what could happen if treatment is postponed.
For a patient undergoing surgery for a serious disease, delay may have consequences of its own. The condition may progress, symptoms may worsen or an opportunity for effective treatment may become more difficult to achieve.
At the same time, certain medical conditions may warrant additional assessment or treatment before a planned procedure.
The appropriate decision depends on the patient and the clinical circumstances.
The reaffirmed 2026 AHA/ACC guideline supports distinguishing situations where surgery can proceed from those where a pause for further cardiovascular evaluation is warranted [2].
It also emphasizes that cardiovascular investigations and treatment should follow appropriate clinical indications rather than being performed solely because an operation is planned.
An AI-generated score does not establish that additional testing will improve outcomes. Nor does it demonstrate that delaying surgery until a particular numerical threshold is reached will benefit the patient.
Most prediction models estimate the probability of an outcome based on relationships observed in their development data. They do not necessarily estimate what would happen if surgery were postponed, the treatment plan changed or an additional intervention introduced.
These are different clinical questions.
For example, a model may identify patients with a higher probability of postoperative complications because of anemia or cardiovascular disease. That finding does not, by itself, demonstrate that postponing surgery will reduce their risk.
A clinically justified postponement should therefore be based on the patient's condition, the expected benefits of additional care, the risks of delay and relevant professional guidance.
AI may contribute to the discussion, but it should not automatically determine the outcome.
This is especially important when software recommendations are connected to surgical scheduling. A system that labels patients as unsuitable for surgery solely because their predicted risk exceeds a threshold could turn an uncertain estimate into a restrictive clinical decision.
Hospitals should ensure that such decisions remain subject to appropriate professional assessment and review.
There is also a distinction between declining an operation solely because of an AI-generated score and making a clinically justified decision not to proceed because safe treatment cannot be provided under the circumstances.
A surgeon or anesthesiologist may have legitimate reasons to recommend postponement, additional assessment or an alternative treatment approach. The presence of AI should not weaken that professional responsibility.
The important requirement is that the decision rests on an appropriate clinical assessment rather than an automated prediction alone.
The surgeon and anesthesiologist have different but connected responsibilities
Preoperative decision-making involves professional responsibilities that should not be transferred to an AI model.
The surgeon evaluates the condition requiring treatment, the proposed procedure, its expected benefits, relevant alternatives and procedure-related risks.
The anesthesiologist assesses the patient's condition in relation to anesthesia and perioperative care, identifies relevant concerns and develops an appropriate anesthesia plan.
Other specialists may contribute when existing medical conditions require further evaluation or management.
A patient with significant cardiovascular disease may benefit from cardiology input. Another may require assessment of respiratory disease, renal function, anemia or other clinically relevant conditions.
However, additional consultations should be based on appropriate indications. A high AI score alone should not automatically generate referrals to multiple specialties or an extended sequence of investigations.
The ASA's Basic Standards for Preanesthesia Care, last amended in October 2025, affirm the anesthesiologist's responsibility for determining the patient's medical status and developing the anesthesia care plan [11].
The standards include reviewing medical records, interviewing and examining the patient, assessing relevant investigations and consultations, and ensuring that appropriate anesthesia consent has been obtained.
An AI prediction may support this process, but it does not replace these clinical responsibilities.
The same principle applies when the AI result differs from the clinician's assessment.
If the model indicates unexpectedly high risk, the team should consider whether the input information is correct, whether the predicted outcome is relevant and whether the result identifies a concern that has not been adequately addressed.
If the model produces a reassuring estimate despite significant symptoms or clinical findings, those findings should not be disregarded.
The objective is not automatic agreement between clinicians and software. It is to ensure that useful information is considered without weakening professional judgment.
Clinically important disagreement should lead to review
A discrepancy between an AI prediction and a clinician's assessment may have several explanations.
The model may be missing relevant information, an existing diagnosis may have been coded incorrectly, a laboratory result may be outdated or the planned procedure may have changed.
The system may also have been developed in a population that differs from the patients receiving care at the hospital.
Conversely, the model may identify a pattern that deserves attention even when the patient's condition initially appears reassuring.
These situations require interpretation rather than automatic acceptance or rejection.
Clinicians should be able to review the information used to generate the prediction and correct inaccurate inputs when necessary.
The organization should also establish how clinically significant disagreements are documented and discussed.
For example, if AI indicates a high probability of complications but the team decides to proceed, the relevant clinical reasoning may be documented within the ordinary preoperative record.
Similarly, when clinicians identify important risks that the system has not captured, those findings should remain central to the management plan.
The purpose is not to create extensive documentation for every difference between human and AI assessment.
It is to preserve a clear basis for clinically important decisions and identify recurring problems that may indicate limitations in the system.
Disagreement between the surgeon and anesthesiologist requires particular attention because their professional assessments address different but connected aspects of the operation.
An AI score should not be used to settle such a disagreement automatically.
Where the clinical team cannot reach agreement about an important safety concern, the matter should be addressed through the institution's appropriate professional consultation and decision-making arrangements.
A useful AI application should make clinical review easier, not create additional barriers to appropriate care.
Risk prediction should support appropriate preoperative optimization
An important purpose of preoperative assessment is identifying clinical problems that can reasonably be addressed before surgery.
For selected patients, this may involve reviewing cardiovascular or respiratory conditions, investigating anemia, assessing functional capacity, evaluating relevant medications or addressing nutritional concerns.
The appropriate measures depend on the patient's condition, the proposed operation and the available evidence.
Not every patient requires extensive testing or prolonged optimization.
NICE guideline NG45 on routine preoperative tests for elective surgery provides an example of a selective approach [12]. It recommends investigations according to the type of procedure, relevant comorbidities and clinical circumstances rather than routinely performing every available test.
The ESAIC guidance similarly emphasizes structured evaluation and appropriate preparation [1].
For AI-assisted care, a risk estimate may identify patients who deserve closer attention. However, the model does not automatically establish which intervention is indicated.
A high predicted cardiovascular risk should not independently trigger stress testing if the patient's clinical circumstances do not justify it.
Likewise, predicting an increased probability of complications does not demonstrate that a particular preoperative intervention will reduce that probability.
It is useful to distinguish risk factors that may be modifiable from those that are not.
Age, for example, may contribute to risk prediction but cannot be changed before surgery. Other findings may identify opportunities for appropriate treatment or adjustments to the care plan.
The goal of optimization is not simply to make the AI score lower. It is to improve the patient's condition or preparedness when there is a reasonable clinical expectation of benefit.
Hospitals should also consider the consequences of unnecessary optimization requirements.
Repeated investigations, additional appointments and avoidable postponements may increase cost, anxiety and patient burden without improving outcomes.
A responsible approach therefore evaluates not only what further preparation is possible, but whether it is appropriate and proportionate for the individual patient.
Postoperative care planning is not the same as predicting ICU admission
Preoperative risk assessment can help teams anticipate the level of care a patient may require after surgery.
Some patients may need closer monitoring, specialized nursing care or access to higher-acuity services because of their medical complexity or the physiological demands of the proposed operation.
However, predicting ICU admission is not equivalent to determining that intensive care will be clinically necessary.
An AI model trained on historical ICU admissions may partly reflect local admission practices, bed availability and institutional criteria for postoperative monitoring.
A high predicted probability may therefore indicate that similar patients have frequently been admitted to intensive care in the past. It does not establish that the individual patient requires an ICU bed.
In a January 2026 editorial in the British Journal of Anaesthesia, Argalious emphasized the importance of considering the actual care requirements that justify postoperative ICU admission and whether those needs can be met safely in other hospital settings [13].
This distinction is relevant to surgical planning.
The appropriate postoperative care location depends on anticipated organ support requirements, monitoring intensity, clinical staffing, the patient's condition and the capabilities of the available care areas.
Some patients may need enhanced postoperative monitoring without requiring ICU admission. Others may develop unexpected complications despite a relatively reassuring preoperative assessment.
Risk estimates may therefore help inform resource planning, but they should not automatically determine ICU admission or reservation.
The same principle applies to other resources, including blood products, rehabilitation services and anticipated length of stay.
Prediction may support preparation. It should not replace individualized clinical planning.
Patient communication should make risk understandable
Patients preparing for surgery need to understand why the operation is proposed, what benefits are expected, what significant risks are involved and what alternatives may be available.
AI-generated estimates may contribute to these discussions, particularly when they help clinicians explain the probability of a specific complication.
However, numerical estimates can easily be misunderstood.
A predicted risk of 10% does not mean that exactly one in every ten patients treated at the hospital will experience the outcome. It is a model-based estimate derived from information about a particular population.
Its usefulness depends on the outcome being predicted, the model's calibration, its applicability to the patient and the uncertainty surrounding the result.
Clinicians should therefore avoid presenting AI-generated percentages as precise or guaranteed individual outcomes.
Where appropriate, probabilities may be communicated using understandable frequencies. A hypothetical 10% estimate, for example, can be explained as approximately ten out of one hundred comparable patients experiencing the specified outcome.
The outcome itself should be clearly identified.
A prediction of any postoperative complication is not the same as a prediction of death, permanent disability or an event requiring additional treatment. Similarly, a model estimating 30-day mortality cannot automatically provide meaningful information about long-term function or the expected benefit of surgery.
Patients also need to understand the consequences of not proceeding.
A discussion focused only on the probability of complications may create unnecessary fear or imply that a higher score makes surgery inappropriate.
The reaffirmed AHA/ACC guideline emphasizes patient engagement and shared decision-making [2]. The ESC guidelines likewise recommend communicating perioperative cardiovascular risk in understandable absolute terms and involving patients in treatment decisions [3].
For hospitals, the important question is not simply whether a numerical risk estimate appears in the medical record.
It is whether the patient receives an understandable explanation of the options, potential benefits, significant risks and remaining uncertainties.
AI may support that conversation, but it should not independently deliver definitive treatment recommendations without appropriate professional review.
Informed consent involves more than a risk percentage
Informed consent is not simply obtaining a signature.
For planned surgery, the patient should receive appropriate information about the proposed procedure, expected benefits, significant risks and relevant alternatives.
The surgical and anesthesia teams should distinguish between risks associated with the operation, anesthesia-related concerns, risks arising from existing medical conditions and uncertainties that cannot be quantified reliably.
An AI estimate may help explain one or more of these risks, but it does not replace communication by qualified clinicians.
The model's intended outcome must also be respected.
A system designed to predict a composite measure of postoperative morbidity may not provide reliable estimates for each individual complication included in that composite.
Similarly, a model that estimates mortality should not be used to imply that it accurately predicts quality of life, recovery time or long-term surgical benefit.
Communication should reflect the patient's preferences and ability to understand the information.
Some patients may find numerical estimates helpful. Others may benefit from a discussion comparing treatment choices and their likely consequences in simpler terms.
If AI contributes materially to the risk information being discussed, clinicians should be able to explain its supporting role and relevant limitations accurately. This does not mean that every form of AI assistance necessarily creates an identical additional consent requirement.
The organization's communication and consent arrangements should be assessed according to the actual use of the technology, applicable legal requirements and institutional procedures.
The objective is to support informed participation, not to persuade the patient through an apparently precise AI score.
Local validation matters for Vietnamese surgical services
A model that performs well in one healthcare system may not perform equally well in another.
Differences in surgical procedures, patient characteristics, clinical documentation and postoperative care can affect prediction performance.
This is particularly relevant when models are developed using datasets from other countries.
Hospitals in Viet Nam may use different surgical coding systems, electronic medical records and laboratory reporting arrangements from those used during model development.
A system relying on detailed procedure codes may produce unreliable estimates if the local hospital cannot provide equivalent information.
Patient populations also differ. Patterns of comorbidity, functional status, disease severity and access to preoperative preparation may vary between institutions.
A model developed mainly among patients undergoing complex inpatient surgery should not automatically be applied to minor ambulatory procedures.
Similarly, a system evaluated within one surgical specialty may require further assessment before use in another.
Local evaluation should determine whether the necessary information can be obtained reliably, whether predictions are generated at the intended clinical time point and whether the model performs adequately in the population where it will be used.
Discrimination is important, but it is not sufficient.
Calibration also matters. If a model consistently predicts substantially higher complication rates than those observed among comparable patients, the resulting percentages may mislead clinical discussions and planning.
Systematic underestimation can create inappropriate reassurance.
Hospitals should also consider whether performance differs across relevant patient groups or procedure types.
The international FUTURE-AI consensus guideline, published in The BMJ in 2025, identifies fairness, universality, traceability, usability, robustness and explainability as principles for trustworthy healthcare AI [14].
TRIPOD+AI provides recommendations for transparent reporting of clinical prediction model studies, including models developed using machine learning [15].
These documents can inform evaluation, but they are not clinical approvals or guarantees that a model is safe for use in any particular hospital.
For Vietnamese surgical services, the practical implication is that a model should be assessed against the intended patient population, available data, clinical workflow and decisions it is expected to support.
What should hospitals evaluate after introducing AI?
A hospital considering AI-assisted preoperative risk assessment should define the improvement it expects before introducing the technology.
A higher AUC than an existing risk tool may be encouraging, but it is not sufficient evidence that surgical care has improved.
The organization should examine whether AI provides useful information and whether that information contributes to more appropriate decisions.
One area is the quality of clinical assessment.
Does the system help identify important concerns that might otherwise be missed? Does it support more consistent review of relevant medical conditions? Does it reduce unnecessary duplication without weakening the assessment?
Another area is the appropriateness of subsequent decisions.
AI may influence specialist referrals, additional investigations, preoperative treatment, modifications to the surgical plan or arrangements for postoperative care.
However, an increase in investigations or consultations is not automatically an improvement.
The organization should examine whether these actions were clinically justified and whether they contributed to a more suitable care plan.
Unnecessary postponement deserves particular attention.
If AI recommendations lead to more delayed or cancelled procedures, the hospital should consider whether those decisions were warranted and whether the expected benefits of postponement outweighed the potential consequences.
For some patients, delay may cause disease progression, additional symptoms, more hospital visits, financial burden or psychological distress.
Hospitals should also recognize a less obvious evaluation problem. If the introduction of AI changes which patients are offered surgery, postoperative outcomes measured only among those who undergo an operation may provide an incomplete picture.
For example, a reduction in observed complication rates could partly reflect the exclusion or postponement of higher-risk patients rather than safer care for comparable patients.
Evaluation should therefore consider the consequences for patients whose procedures are delayed, changed or not performed, as well as outcomes among those who undergo surgery.
Postoperative outcomes remain important, although their selection should reflect the model's intended function.
Relevant measures may include major complications, unplanned ICU admission, unplanned reoperation, mortality, prolonged hospitalization or other clinically meaningful outcomes.
Changes in these outcomes need careful interpretation. Complication rates can be influenced by patient selection, surgical technique, anesthesia management, staffing and other quality-improvement activities.
A reduction in complications after AI installation does not, by itself, prove that the technology caused the improvement.
Patient communication should also be evaluated.
Hospitals may consider whether patients understand their principal risks, expected benefits and available treatment options, and whether discussions support meaningful participation in decisions.
A system that generates more detailed percentages without improving patient understanding may offer limited practical value.
Unexpected consequences should be monitored as well, including inappropriate postponements, unnecessary investigations, overreliance on AI recommendations and additional administrative workload.
Comparisons before and after implementation should take account of changes in patient characteristics, procedures and other factors that influence clinical outcomes.
These measures are proposed organizational evaluation considerations. They are not a mandatory Vietnamese AI preoperative assessment scorecard or an independently validated instrument.
The success of AI-assisted preoperative assessment should be judged by the quality of clinical decisions and care planning, not simply by improvements in statistical prediction performance.
Integrating AI into existing surgical and anesthesia workflows
Before introducing AI, a hospital should understand how its current preoperative pathway operates.
The organization needs to identify when risk assessment begins, how relevant clinical information is collected and how findings are communicated between surgical and anesthesia teams.
The intended role of AI can then be defined.
One application might support early identification of patients who could benefit from additional review. Another might provide updated estimates after the proposed procedure and preoperative investigations are confirmed.
These functions occur at different stages and should not be treated as interchangeable.
Clinical leaders and frontline professionals should participate in defining the intended use, selecting the technology and identifying what improvement is expected.
The surgeon and anesthesiologist should remain able to review the underlying clinical information and determine whether the output is relevant to the patient.
The hospital should establish how inaccurate inputs are corrected and how clinically important disagreements are handled.
The process should be practical. A system that makes it difficult to correct an incorrect procedure code or requires excessive administrative work before a recommendation can be questioned may interfere with clinical care.
Predictions should also be reviewed when the patient's condition or surgical plan changes.
A fallback process is essential. Staff must remain capable of conducting appropriate preoperative assessment when AI is unavailable or unsuitable.
Training should address the system's purpose, limitations and relationship to established professional responsibilities.
Realistic scenarios may help clinicians understand how to respond when a patient has a high predicted risk but a strong indication for surgery, or when a low prediction conflicts with concerning clinical findings.
Software updates and changes to clinical data or workflow may affect performance. The organization should therefore consider whether further evaluation is needed after significant modifications.
Where appropriate and permitted by applicable requirements, an initial evaluation may involve using AI alongside existing practice before allowing its recommendations to actively influence routine decisions.
The aim is to establish whether the technology adds useful information without creating new risks or weakening existing safeguards.
Vietnamese professional and legal responsibilities
AI-assisted preoperative risk assessment must be considered within Viet Nam's legal framework for medical examination and treatment, anesthesia and surgical care, artificial intelligence, medical devices and personal data protection.
These frameworks address different responsibilities. Compliance with one set of requirements does not automatically establish compliance with the others.
Medical examination, treatment and patient rights
The Law on Medical Examination and Treatment No. 15/2023/QH15, effective January 1, 2024, as subsequently amended, provides an important legal foundation for healthcare services and professional practice in Viet Nam [16].
The Law establishes rights relevant to preoperative decision-making, including the right to receive information about the patient's condition, treatment options and potential risks.
Article 11 addresses the patient's right to choose examination and treatment methods after receiving appropriate information, explanation and consultation, subject to the applicable legal provisions.
Article 13 addresses the right to refuse examination or treatment under the conditions established by law, including the relevant written acknowledgment of responsibility where required.
These provisions are relevant when AI-generated estimates form part of discussions about planned surgery.
A patient should not be led to believe that a software-generated probability independently determines whether an operation is permitted or prohibited.
At the same time, shared decision-making does not require clinicians to provide an intervention that is professionally inappropriate or beyond the healthcare facility's authorized capabilities.
Clinical decisions must remain within the applicable professional and institutional framework.
AI does not independently authorize a practitioner to make decisions outside their permitted scope or remove responsibilities for clinical judgment, patient communication and documentation.
Pre-anesthesia assessment and professional coordination
Circular No. 13/2012/TT-BYT, issued by the Ministry of Health on August 20, 2012, provides specific guidance on anesthesia and resuscitation activities in healthcare facilities authorized to perform relevant procedures [17].
The Circular remains in effect and addresses pre-anesthesia examination, patient preparation and coordination between anesthesia services and surgical departments.
For non-emergency procedures within its scope, it specifies that the pre-anesthesia examination is performed within one to seven days before surgery or the relevant procedure.
This formal pre-anesthesia examination should be distinguished from the broader process of identifying and addressing preoperative risks, which may need to begin earlier.
The Circular provides for anesthesiologists to assess the patient's condition, request additional investigations or consultations when necessary and document relevant findings.
Article 10 addresses the professional responsibilities of anesthesiologists during the pre-anesthesia examination. Where conditions are not sufficiently safe, the anesthesiologist may request that the surgeon temporarily postpone the operation, with the request documented in the medical record. If agreement cannot be reached, the matter must be reported to the department head for resolution under the applicable arrangements.
The Circular also requires communication with the professionals performing the operation and anesthesia care about possible anesthesia-related risks and complications.
The anesthesiologist must explain relevant risks and benefits of anesthesia to the patient or legally authorized representative before the applicable consent process.
Article 15 further establishes responsibilities for coordination between surgical departments and anesthesia services, including patient preparation, communication of the surgical plan and arrangements for anesthesia care.
Importantly, Article 15 addresses the authority of the anesthesiologist responsible for the anesthesia team to refuse anesthesia where specified requirements are not met or where anesthesia safety is threatened. The physician remains responsible for that professional decision.
Where the surgeon and anesthesiologist cannot reach agreement before or during surgery, the Circular provides for professional consultation in accordance with the applicable procedures.
These provisions are directly relevant to AI-assisted preoperative assessment.
A model may help identify concerns requiring review, but it does not replace the professional pre-anesthesia examination or the coordination required between surgeons, anesthesiologists and other relevant personnel.
Nor should an AI recommendation independently determine that a patient is suitable or unsuitable for anesthesia.
The distinction is between an automated risk score and a clinical decision made by an authorized professional who has assessed the patient and is responsible for the relevant aspects of care.
Hospitals should ensure that AI supports, rather than bypasses, established assessment, consultation, communication and documentation requirements.
Artificial intelligence regulation
Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on March 1, 2026 [18].
Decree No. 142/2026/NĐ-CP, effective May 1, 2026, provides implementing provisions, including those concerning AI risk classification and management [19].
Decision No. 33/2026/QĐ-TTg, effective August 15, 2026, establishes the list of high-risk AI systems [20].
The healthcare-specific entries identified in that list concern defined AI-assisted surgical and robotic surgical applications. General preoperative risk prediction software is not separately identified among those healthcare entries.
However, the absence of an explicit entry should not be interpreted as a definitive legal determination that every preoperative AI application is low-risk.
The classification and applicable obligations depend on the system's actual functions, intended purpose, deployment arrangements and relevant legal criteria.
A model that provides additional risk information may have a different regulatory profile from a system designed to automate or control clinical decisions.
Statutory AI risk classification should also be distinguished from an organization's clinical risk assessment.
Even where a particular application does not fall within a specified high-risk category, inappropriate use could contribute to unnecessary postponement, inadequate preparation or other patient-safety problems.
Healthcare organizations should therefore examine both the legal status of the product and the clinical consequences of its use.
The introduction of AI does not independently transfer the legally established authority or responsibilities of healthcare professionals and organizations.
Medical device regulation
Medical device requirements may apply to AI software depending on its intended medical purpose and functions.
Decree No. 98/2021/NĐ-CP on medical device management, as amended by subsequent regulations including Decrees No. 07/2023/NĐ-CP and No. 04/2025/NĐ-CP, provides part of the relevant regulatory framework [21].
Circular No. 24/2026/TT-BYT, effective July 1, 2026, further addresses risk determination and management measures for products classified as medical devices [22].
Before deployment, hospitals should determine whether the proposed software is regulated as a medical device and which requirements apply to its classification, documentation and lawful use.
Regulatory authorization obtained in another country may provide useful supporting information, but it should not automatically be treated as evidence of compliance with all Vietnamese requirements.
The assessment should relate to the actual product and its functions rather than solely to its commercial description as an AI risk calculator or clinical decision-support system.
Personal data protection
AI-assisted preoperative assessment may involve processing medical history, diagnoses, laboratory results, planned surgical procedures and other information associated with identifiable patients.
The Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP took effect on January 1, 2026 [23, 24].
Healthcare organizations should assess the applicable legal basis and purposes of data processing, consent requirements or exceptions, access controls, information security, retention and arrangements with technology providers.
Particular attention is needed when patient information is transferred to external systems or proposed for secondary purposes such as algorithm development, model training or commercial product improvement.
The applicable obligations depend on the nature of the data, processing purposes, technical architecture and responsibilities of the parties involved.
Hospitals should not assume that using clinical information to generate a risk estimate automatically authorizes its use for unrelated purposes.
Where an AI system is being evaluated through clinical research rather than solely introduced for ordinary clinical service, the applicable research ethics and participant protection requirements should be assessed separately.
These legal assessments should inform technology selection and deployment. They do not replace appropriate clinical evaluation or professional judgment.
A practical approach for surgical services in Viet Nam
For hospitals, surgical centers and anesthesia departments in Viet Nam, responsible AI-assisted preoperative assessment should begin with a clearly defined clinical need.
An organization may wish to improve the consistency of risk review, identify patients who need additional preoperative assessment or strengthen discussions about potential complications.
Each objective requires an appropriate evaluation approach.
The hospital should first consider whether its existing preoperative process is reliable.
Relevant information should be available to the responsible clinicians, and significant medical concerns should be recognized early enough for appropriate action.
Surgical and anesthesia teams should have effective arrangements for communicating findings, discussing disagreements and developing a coordinated plan.
Patient information and consent processes should also support meaningful discussion of benefits, risks and alternatives.
Where these arrangements are weak, improvements in clinical workflow and professional coordination may provide important benefits independently of AI.
If a suitable AI system is available, its introduction should be linked to a specific decision point and measurable quality objective.
Clinical leaders should review the supporting evidence, intended patient population, required inputs and known limitations.
A supervised implementation can then examine whether the model contributes useful information, influences decisions appropriately and creates any unexpected consequences.
Particular attention should be given to unnecessary postponement, investigations without clear clinical indications and decisions about postoperative care based primarily on predicted probabilities.
The hospital should also evaluate whether patients receive clearer explanations and whether clinicians can use the technology without excessive administrative burden.
Responsible implementation does not mean making every preoperative assessment more technologically complex.
It means helping clinical teams make appropriate decisions, prepare patients effectively and avoid preventable harm.
From predicting complications to better surgical decisions
AI-assisted preoperative risk assessment has the potential to provide surgical and anesthesia teams with additional information about postoperative complications.
Recent studies have reported encouraging predictive performance, including improvements over some traditional risk tools.
However, the clinical importance of these findings depends on more than the ability to distinguish patients who will and will not experience adverse outcomes.
A useful prediction must be available at the correct time, relevant to the proposed procedure and sufficiently reliable for the patient population in which it will be used.
Even then, it does not independently determine whether surgery should proceed.
Clinical decisions must consider the patient's condition, the expected benefit of treatment, available alternatives, opportunities for appropriate preparation and the consequences of delay.
The surgeon, anesthesiologist and relevant specialists retain their respective professional responsibilities. Patients need understandable information and an opportunity to participate meaningfully in decisions about their care.
For healthcare organizations in Viet Nam, the objective is to use AI to strengthen these established processes, not to allow numerical predictions to replace them.
Evaluation should therefore examine whether the technology improves clinical review, patient communication and surgical planning while avoiding unnecessary investigations, postponements and restrictions on care.
The value of AI-assisted preoperative risk assessment should not be judged simply by how accurately it predicts complications. It should be judged by whether the information helps clinicians and patients make better-informed decisions and prepare a safer, more appropriate plan of care.
Professional Education Notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized medical, surgical, anesthesia, cardiology, intensive care, legal, regulatory or patient-safety advice.
The opening clinical scenario is illustrative and does not represent a documented patient case.
This article focuses on preoperative risk assessment in adults preparing for non-emergency, predominantly noncardiac surgery. It does not establish the safety or effectiveness of AI systems for autonomous anesthesia management, AI-assisted surgical intervention, intraoperative decision-making or postoperative prediction based on information unavailable before surgery.
The organizational approaches discussed, including clinical review of AI predictions, preoperative optimization, patient communication, surgical planning and performance monitoring, are proposed considerations for healthcare quality improvement and responsible AI implementation. They are not mandatory Vietnamese Ministry of Health AI preoperative assessment protocols or independently validated clinical assessment instruments.
Traditional tools, including RCRI, ACS NSQIP and ASA Physical Status, have different purposes and limitations. Their use should follow applicable professional guidance and should not be assumed to establish the safety or effectiveness of newer AI models.
A predicted probability of postoperative complications should not independently determine whether surgery is postponed, refused or performed, which additional investigations are required or whether a patient receives postoperative ICU care.
International guidance from ESAIC, AHA/ACC, ESC, ASA, NICE and other organizations reflects specific clinical populations, settings and evidence assessments. These recommendations should not automatically be interpreted as binding requirements for healthcare facilities in Viet Nam.
Evidence showing improved predictive discrimination or calibration does not automatically establish that an AI system improves surgical decisions, reduces postoperative complications or produces better patient outcomes.
Results from retrospective datasets, selected surgical populations and healthcare systems outside Viet Nam should not be generalized to local clinical practice without appropriate assessment of intended use, data availability, validation, clinical workflow and implementation conditions.
The legal status and obligations associated with individual AI systems, medical device software, professional practice, informed consent, personal data processing and clinical research require separate assessment under the applicable Vietnamese legal framework.
References
Lamperti M, Romero CS, Guarracino F, et al. Preoperative Assessment of Adults Undergoing Elective Noncardiac Surgery: Updated Guidelines From the European Society of Anaesthesiology and Intensive Care. European Journal of Anaesthesiology. 2025;42(1):1–35. DOI: 10.1097/EJA.0000000000002069.
Thompson A, Fleischmann KE, Smilowitz NR, et al. 2026 AHA/ACC/ACS/ASNC/HRS/SCA/SCCT/SCMR/SVM Guideline for Perioperative Cardiovascular Management for Noncardiac Surgery: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Journal of the American College of Cardiology. 2026;88(13):1543–1643. DOI: 10.1016/j.jacc.2026.06.017.
Halvorsen S, Mehilli J, Cassese S, et al. 2022 ESC Guidelines on Cardiovascular Assessment and Management of Patients Undergoing Non-Cardiac Surgery. European Heart Journal. 2022;43(39):3826–3924. DOI: 10.1093/eurheartj/ehac270.
Lee TH, Marcantonio ER, Mangione CM, et al. Derivation and Prospective Validation of a Simple Index for Prediction of Cardiac Risk of Major Noncardiac Surgery. Circulation. 1999;100(10):1043–1049. DOI: 10.1161/01.CIR.100.10.1043.
Bilimoria KY, Liu Y, Paruch JL, et al. Development and Evaluation of the Universal ACS NSQIP Surgical Risk Calculator: A Decision Aid and Informed Consent Tool for Patients and Surgeons. Journal of the American College of Surgeons. 2013;217(5):833–842.e1–3. DOI: 10.1016/j.jamcollsurg.2013.07.385.
American College of Surgeons. ACS NSQIP Surgical Risk Calculator. Version 4.0.4. Last parameter update: April 2026.
American Society of Anesthesiologists. Statement on ASA Physical Status Classification System. Revised October 15, 2025. Anesthesiology Open. 2026;1(1). DOI: 10.1097/AO9.0000000000000002.
Dhaliwal S, Chen S, Papas C, Hughes I, Cavalucci D, O'Rourke N. Machine Learning vs. Traditional Methods for Predicting Postoperative Cardiac Complications After Non-Cardiac Surgery: A Systematic Review and Bayesian Network Meta-Analysis. Anaesthesia. 2026;81(10):1415–1427. DOI: 10.1111/anae.70237.
Yang J, Dai S, Yu H, Cai X, Zhang Z. Development and External Validation of a Lightweight, Explainable Clinical Decision Support System for Personalized Perioperative Risk Stratification Using Electronic Health Records. Scientific Reports. 2026;16:29834. DOI: 10.1038/s41598-026-60202-9.
Nishino T, Mase H, Kim C, Sugita S, Kondo Y, Ishikawa M. Machine Learning–Based Prediction of Major Postoperative Complications After Noncardiac Surgery Under General Anesthesia Using Routinely Available Electronic Health Record Data: Retrospective Multicenter Development and Validation Study. JMIR Perioperative Medicine. 2026;9. Published October 21, 2026. DOI: 10.2196/91315.
American Society of Anesthesiologists. Basic Standards for Preanesthesia Care. Committee on Practice Parameters. Last amended October 15, 2025.
National Institute for Health and Care Excellence. Routine Preoperative Tests for Elective Surgery. NICE Guideline NG45. Published April 5, 2016.
Argalious MY. Postoperative Intensive Care Unit Admissions: Leveraging Science to Strike the Right Balance. British Journal of Anaesthesia. 2026;136(1):5–8. DOI: 10.1016/j.bja.2025.10.034.
Lekadir K, Frangi AF, Porras AR, et al. FUTURE-AI: International Consensus Guideline for Trustworthy and Deployable Artificial Intelligence in Healthcare. BMJ. 2025;388. DOI: 10.1136/bmj-2024-081554.
Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI Statement: Updated Guidance for Reporting Clinical Prediction Models That Use Regression or Machine Learning Methods. BMJ. 2024;385. DOI: 10.1136/bmj-2023-078378.
National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment. Adopted January 9, 2023; effective January 1, 2024, as subsequently amended.
Ministry of Health of Viet Nam. Circular No. 13/2012/TT-BYT on Guidelines for Anesthesia and Resuscitation Activities. Issued August 20, 2012; effective October 10, 2012.
National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted December 10, 2025; effective March 1, 2026.
Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing Certain Provisions and Measures for Implementing the Law on Artificial Intelligence. Issued April 30, 2026; effective May 1, 2026.
Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. Issued June 30, 2026; effective August 15, 2026.
Government of Viet Nam. Decree No. 98/2021/NĐ-CP on Medical Device Management, as amended, including by Decree No. 07/2023/NĐ-CP and Decree No. 04/2025/NĐ-CP.
Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Risk Determination and Management Measures for Products Classified as Medical Devices and Amendments to Circular No. 05/2022/TT-BYT. Issued June 30, 2026; effective July 1, 2026.
National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted June 26, 2025; effective January 1, 2026.
Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing Certain Provisions and Measures for Implementing the Law on Personal Data Protection. Issued December 31, 2025; effective January 1, 2026.