Governing AI-Assisted Chronic Kidney Disease Case Finding in Viet Nam: Confirming Chronicity, Risk Stratification and Referral Follow-Up

Governing AI-Assisted Chronic Kidney Disease Case Finding in Viet Nam: Confirming Chronicity, Risk Stratification and Referral Follow-Up

An outpatient physician reviewing a patient's medical history notices that kidney function has been reduced on several occasions. The results were obtained months apart, sometimes during appointments with different clinicians. No urine albumin-to-creatinine ratio (UACR) has been recorded, and there is no documented diagnosis of chronic kidney disease (CKD). Although the information has been available in the medical record, no clear plan for further kidney assessment appears to have been established.

Artificial intelligence (AI) could help healthcare organizations recognize situations like this. By examining laboratory results, clinical histories and repeated healthcare encounters, AI-assisted systems may identify patients who need further kidney assessment, highlight missing investigations or bring previously overlooked abnormalities to clinical attention.

Finding an abnormal result, however, is not the same as diagnosing CKD. A reduced estimated glomerular filtration rate (eGFR) may reflect an acute illness or a temporary change in kidney function. Persistent albuminuria may indicate kidney damage even when eGFR remains relatively preserved. Once CKD is confirmed, management depends on its cause, severity, risk of progression and the patient's broader clinical circumstances.

For healthcare organizations in Viet Nam, the challenge is therefore not simply whether AI can identify more patients who may have CKD. It is whether those findings lead to accurate clinical assessment, appropriate risk stratification and reliable follow-up without creating unnecessary diagnoses, investigations or referrals.

Recognizing Patients at Risk Without Overdiagnosing CKD

Chronic kidney disease can develop without noticeable symptoms, particularly during its earlier stages. Patients may continue receiving treatment for diabetes, hypertension or cardiovascular disease without realizing that kidney abnormalities are also present. Routine clinical encounters therefore provide opportunities to identify people who would benefit from further kidney assessment.

Viet Nam's Ministry of Health addressed this issue in Decision No. 2388/QD-BYT, dated 12 August 2024, which issued the professional guidance Diagnosis and Treatment of Chronic Kidney Disease and Certain Kidney Diseases. The guidance recommends active assessment of people with recognized CKD risk factors using measurements of kidney function and urinary albumin, generally at least annually for relevant risk groups.

The KDIGO 2024 Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease similarly recommends assessing people at risk using both glomerular filtration rate and urine albumin measurements. The American Diabetes Association's 2026 Standards of Care recommend assessing UACR and eGFR at least annually in all people with type 2 diabetes, regardless of treatment.

These recommendations provide a clinical foundation for targeted case finding. They do not imply that every patient requires identical testing or that AI is necessary to implement appropriate kidney assessment.

A patient may have an eGFR within the usual reference range while persistent albuminuria indicates kidney damage. Another patient may have a reduced eGFR during an acute illness without sufficient evidence of chronic disease. Measuring serum creatinine alone may therefore provide an incomplete picture.

Electronic health records can support recognition by bringing together historical laboratory measurements, diagnoses and previous healthcare encounters. A conventional clinical decision-support system may identify repeated reductions in eGFR or prompt UACR testing when an eligible patient has not undergone the recommended assessment.

More advanced systems may use machine learning to identify longitudinal patterns associated with previously unrecognized CKD. Large language models may also interpret clinical documentation alongside structured laboratory data. However, these approaches introduce additional uncertainty because the systems may misunderstand clinical context, overlook missing information or generate conclusions that the available evidence does not support.

Not every digital system that identifies possible CKD is an AI system. Rule-based alerts, statistical risk calculators, machine-learning models and large language models perform different functions and require different forms of evaluation.

For example, the Kidney Failure Risk Equation (KFRE) is a statistical prediction tool. Incorporating it into electronic software does not automatically make that software an AI system. Similarly, identifying two abnormal laboratory results separated by a predefined interval can be accomplished through conventional rules without machine learning.

Healthcare organizations should establish the intended purpose of each system before considering implementation. Identifying patients who require further assessment, recognizing established CKD and predicting future kidney disease are different clinical functions. Evidence supporting one function cannot automatically be used to justify another.

This distinction matters for clinical validation, professional responsibilities, patient communication and regulatory assessment. It also determines how the organization should evaluate whether the technology improves care.

Confirming Chronicity Before Establishing a Diagnosis

The defining feature of chronic kidney disease is the persistence of abnormalities of kidney structure or function, with implications for health, for at least three months.

Under KDIGO 2024, CKD may be established through evidence of a persistently reduced GFR below 60 mL/min/1.73 m² or recognized markers of kidney damage, including persistent albuminuria with a UACR of at least 30 mg/g, approximately 3 mg/mmol. Other qualifying findings, such as relevant structural abnormalities, may also establish CKD when the criteria for chronicity and clinical significance are satisfied.

An eGFR of 60 mL/min/1.73 m² or higher does not exclude CKD when other qualifying evidence of kidney damage is present. Conversely, an eGFR between 60 and 89 mL/min/1.73 m², without another recognized marker of kidney damage, is insufficient by itself to establish CKD.

Viet Nam's 2024 Ministry of Health guidance is consistent with the need to establish chronicity rather than assigning a diagnosis solely from an isolated abnormal result.

This creates a particular safety challenge for AI-assisted case finding. A system may recognize an eGFR below a predefined threshold, but it should not automatically interpret that result as evidence of established chronic disease.

Consider a patient whose laboratory-reported eGFR falls to 52 mL/min/1.73 m² during an acute infection. Without adequate historical information, the finding could represent acute kidney injury, previously unrecognized CKD or acute deterioration superimposed on chronic kidney disease.

The clinician needs to assess the patient's current condition, previous kidney function, medication exposure, volume status and other relevant findings before determining the appropriate next step. In acute kidney dysfunction, creatinine concentrations may not be in a steady state, making conventional creatinine-based eGFR estimates less reliable for interpreting current kidney function.

Evidence of chronicity may come from previous laboratory measurements, imaging findings, documented clinical history or subsequent testing. Two abnormal blood tests obtained only a few days apart do not establish that the abnormality has persisted for three months. Conversely, when reliable historical evidence already demonstrates chronicity, there may be no need to wait another three months to establish the diagnosis.

The three-month diagnostic criterion must not become a reason to delay necessary clinical care.

A newly identified or unexpected reduction in kidney function may require reassessment within days or weeks, depending on the circumstances. Suspected acute kidney injury, rapidly worsening kidney function, oliguria, significant electrolyte disturbances or other urgent findings require timely evaluation rather than placement in a routine CKD confirmation process.

The 2026 European Society of Cardiology (ESC) guidelines on cardiovascular disease and CKD, developed in collaboration with the European Renal Association, recommend repeating newly detected unexpected reductions in eGFR or elevations in albuminuria within days to weeks to support timely diagnosis and management of potentially acute kidney problems.

Albuminuria also requires appropriate confirmation. KDIGO recommends confirming an elevated UACR obtained from a random urine sample with a subsequent first-morning sample where appropriate. Results should be interpreted in the context of factors that may temporarily increase urinary albumin excretion, including urinary tract infection, recent strenuous exercise, menstruation and certain acute clinical conditions.

An AI system that treats every elevated UACR as permanent kidney damage may contribute to overdiagnosis. Conversely, a system that dismisses an isolated abnormality without appropriate reassessment may contribute to delayed recognition of clinically important disease.

A safer approach distinguishes patients with confirmed CKD from those requiring further diagnostic assessment, those with potentially acute kidney abnormalities and those whose available records are insufficient to determine their status.

These distinctions need to remain visible to the clinician reviewing the finding. An algorithmic classification should not conceal uncertainty that could materially change patient management.

When Data Quality Changes the Clinical Interpretation

AI-assisted CKD case finding depends on the quality and continuity of clinical information.

In Vietnamese healthcare settings, relevant data may be distributed across hospital information systems, laboratory platforms, outpatient records and documents from other healthcare providers. A patient may have several serum creatinine measurements recorded at one hospital but no available UACR results. Previous investigations performed elsewhere may not be accessible through the current electronic record.

These gaps can substantially alter the interpretation of an AI-generated finding. The absence of a UACR measurement must not be treated as evidence of normal albumin excretion. Likewise, the absence of a recorded CKD diagnosis does not establish that the patient has never been assessed or diagnosed at another facility.

Laboratory units and measurement methods introduce additional risks. Serum creatinine may be reported in mg/dL or µmol/L, while UACR may be expressed in mg/g or mg/mmol. Incorrect conversion, inconsistent data mapping or errors in matching results to the correct patient can lead to inappropriate classification and clinical recommendations.

The equation used to calculate eGFR also matters. Viet Nam's 2024 Ministry of Health guidance includes the CKD-EPI 2021 creatinine equation and discusses other estimating approaches in specified circumstances. KDIGO recommends using appropriately validated estimating equations and supports race-free estimation of GFR.

Healthcare organizations therefore need to understand which equations their laboratories use, whether historical results were calculated differently and how changes in laboratory methods or reporting conventions may influence apparent trends.

Creatinine-based eGFR may be less reliable in people with unusually low or high muscle mass or other conditions affecting creatinine generation or measurement. Where a more accurate estimate would influence an important clinical decision, combined creatinine–cystatin C estimation or another appropriate method may be considered when available.

These considerations are especially relevant when algorithms examine longitudinal changes. An apparent deterioration may represent genuine disease progression, but it may also reflect an acute illness, measurement variability, a change in laboratory methodology or incomplete records.

Clinicians should be able to review the laboratory values, dates, source information and relevant clinical findings supporting an AI-generated result. A system that produces a convincing conclusion without making its underlying evidence reasonably accessible may be difficult to use safely, even when its reported predictive performance is strong.

Using Risk Stratification to Guide Appropriate Care

Once CKD has been confirmed, the clinical question changes. The objective is no longer simply to establish whether disease is present, but to understand the patient's risk and determine the most appropriate management pathway.

KDIGO recommends classifying CKD according to cause, GFR category and albuminuria category, commonly referred to as the CGA framework. This recognizes that patients with similar eGFR values can face substantially different risks depending on albuminuria, underlying disease, rate of progression and other clinical factors.

A patient with stable G3a CKD and minimal albuminuria may require a different monitoring and management plan from someone with a similar eGFR but severe albuminuria, persistent hematuria or rapidly declining kidney function.

Viet Nam's 2024 Ministry of Health guidance incorporates risk-based assessment and follow-up approaches, including specialist management for patients with more advanced CKD and markedly increased albuminuria.

For adults with established CKD categories G3–G5, KDIGO additionally recommends using an externally validated equation to estimate the absolute risk of kidney failure.

The Kidney Failure Risk Equation is a well-established example. Its commonly used four-variable version incorporates age, sex, eGFR and UACR to estimate the probability of kidney failure requiring kidney replacement therapy over two or five years.

KDIGO identifies a five-year kidney failure risk of approximately 3%–5% as a threshold that may help inform nephrology referral, alongside eGFR, albuminuria and other clinical considerations. A two-year risk above 10% may help determine when multidisciplinary care is appropriate.

The 2026 ESC guidelines, developed with the European Renal Association, also support validated kidney failure risk assessment in patients with CKD G3–G5 to help guide management and timely specialist referral.

These recommendations offer a more individualized approach than treating every reduced eGFR as an automatic indication for specialist referral. However, the thresholds should not be converted into inflexible automated rules.

A prediction equation must be appropriate for the population and clinical circumstances in which it is applied. An equation developed for CKD G3–G5 should not automatically be extended to G1–G2 or to other populations outside its validated scope.

For Vietnamese healthcare organizations, assessing the local performance and calibration of a risk equation is particularly important when its results will influence referral decisions or resource allocation. Good predictive performance in another healthcare system does not guarantee equally accurate risk estimates in every local population.

The reliability of the input information also matters. A system should not silently assign a normal UACR value when albuminuria has not been measured. Incomplete information may justify further assessment, but it should not be converted into an apparently precise risk estimate without adequate justification.

A low predicted kidney failure risk does not exclude the need for nephrology assessment. Rapid deterioration, severe albuminuria, unexplained hematuria, suspected glomerular disease, possible urinary obstruction or difficult management problems may require specialist attention regardless of the calculated score.

Risk stratification is most useful when it helps clinicians individualize follow-up and match available services to clinical needs. It should support professional judgment rather than replace it.

Closing the Gap Between Identification and Follow-Up

One of the most important weaknesses in chronic disease detection is the gap between recognizing a problem and completing the next clinical action.

An AI system may identify patients with repeated reductions in eGFR and place them on an electronic worklist. Yet the clinical problem remains unresolved if no one reviews the findings, necessary investigations are not arranged or the patient does not return.

A reliable case-finding program therefore requires a process that assigns responsibility and documents what happens after identification.

The initial clinical review should determine whether CKD is already confirmed, whether additional investigations are necessary, whether a potentially acute problem requires escalation and whether existing follow-up arrangements are adequate.

When further testing is indicated, the care plan should identify who is responsible for arranging the investigation, reviewing the result and determining the subsequent clinical action. Issuing a laboratory order is not equivalent to completing the diagnostic process.

For patients with established CKD, ongoing management may include monitoring kidney function and albuminuria, evaluating cardiovascular and metabolic risks, reviewing medication safety and considering kidney-protective treatment according to applicable clinical guidance.

The ADA 2026 Standards of Care and the 2026 ESC guidelines provide updated recommendations relevant to kidney and cardiovascular risk management. Treatment decisions nevertheless require individualized assessment of clinical indications, contraindications, patient preferences and other relevant circumstances. A case-finding system should not independently generate treatment instructions for implementation without appropriate professional review.

Specialist referral introduces another potential gap. An electronic referral may have been issued, but the patient may not have attended an appointment. The receiving facility may require additional information, or the patient may face difficulties with transportation, cost or access to specialist services.

For non-urgent referrals, a reasonable process includes explaining the purpose of referral, communicating relevant clinical information, arranging an appropriate appointment or follow-up plan and checking whether the next stage of care has occurred. Urgent clinical concerns require an accelerated pathway consistent with the patient's condition.

The referring service should also have an appropriate mechanism for receiving specialist recommendations and incorporating them into ongoing care. This is particularly important when patients undergo laboratory testing, outpatient treatment and nephrology assessment at different healthcare organizations.

No individual facility can control every action taken by another provider. Nevertheless, responsibilities for communication, reasonable follow-up, documentation and transfer of care should be clearly defined.

Patient communication deserves equal attention. Being told that an AI system has identified possible kidney disease may create anxiety if the result has not been clinically confirmed.

Patients should understand that an abnormal measurement or AI-generated finding may indicate the need for further assessment rather than an established diagnosis. Explanations should focus on what has been found, why additional assessment may be necessary and what the next step will involve.

A predicted kidney failure risk should not be presented as an inevitable outcome. The purpose of communicating risk is to support informed decisions and appropriate care, not to create unnecessary concern.

The objective is to help patients understand their care pathway and participate meaningfully in decisions affecting their health.

What the 2026 Evidence Tells Us

Recent research suggests that AI and electronic clinical decision support can contribute to CKD identification and management. However, the evidence also demonstrates why technical performance, improvements in clinical processes and meaningful patient outcomes must be evaluated separately.

In September 2026, Wu and colleagues published a study in BMJ Health & Care Informatics examining the use of a large language model to analyze longitudinal electronic health records for CKD identification in Chinese primary care.

The researchers selected 2,300 individuals from an initial dataset of 29,963 eligible adults and generated clinical vignettes for model evaluation. Using one month of electronic health record data, the model achieved an area under the receiver operating characteristic curve of 0.921, sensitivity of 89.1%, specificity of 86.8% and an F1 score of 0.579.

These findings suggest that large language models may recognize clinically relevant patterns in longitudinal records. However, the comparatively modest F1 score illustrates why strong discrimination or sensitivity alone does not fully describe a model's usefulness in case finding.

An important limitation is that the primary reference classification relied on recorded diagnostic codes. The researchers also examined individuals with laboratory abnormalities who lacked a formal CKD diagnosis, but these analyses did not replace a prospective, independently adjudicated clinical assessment of chronicity.

Consequently, the study should not be interpreted as demonstrating that the model can independently establish CKD according to all accepted diagnostic criteria. It also did not show that routine deployment improved referral completion, treatment or long-term kidney outcomes.

Evidence from randomized clinical decision-support studies provides a complementary perspective.

In May 2026, Zheng and colleagues published a cluster randomized clinical trial in JAMA Network Open involving 3,390 patients across 30 primary care centers in China. The study evaluated an electronic CKD clinical decision-support system in an environment where both intervention and control groups received education and policy support.

At six months, the intervention did not demonstrate significant additional improvements across the principal CKD management measures compared with the control approach. Selected medication safety alerts showed potential benefit, but the overall results did not establish an independent improvement in CKD management attributable to the electronic decision-support system.

The published six-month findings represented an initial assessment of a longer trial and should not be interpreted as definitive evidence about long-term kidney or cardiovascular outcomes.

Another randomized study, CKD-DETECT, published in The American Journal of Medicine in 2026, examined whether electronic alerts could improve UACR testing among patients with type 2 diabetes who had not undergone UACR assessment during the preceding 12 months and had no known CKD.

The trial included 400 patients. UACR testing was ordered for 36.0% of patients in the alert group compared with 11.5% in the control group.

Despite this substantial increase in test ordering, new CKD diagnoses were recorded in one patient in the alert group and two patients in the control group. No nephrology referrals or new prescriptions of CKD-related medications were recorded during the study.

The intervention improved a defined testing process and increased identification of elevated albuminuria, but it did not demonstrate a corresponding improvement in new CKD diagnoses, specialist referral or treatment within the observation period.

The distinction between ordering a test and completing the subsequent clinical pathway is important. An additional order does not necessarily mean that the investigation was performed, that an abnormal result was reviewed or that appropriate care followed.

The two randomized decision-support studies should also not be described as trials of machine-learning AI merely because their interventions used electronic technology.

Together, these studies illustrate a central challenge for digital kidney care. Identifying an abnormality or prompting an investigation may be useful, but neither guarantees that the finding will be clinically confirmed or followed by appropriate management.

For Vietnamese healthcare organizations, the evidence supports careful evaluation of AI-assisted case finding rather than assuming that stronger prediction performance will automatically translate into better patient outcomes.

Evaluating AI-Assisted CKD Case Finding in Vietnamese Healthcare Settings

Before implementing AI-assisted CKD case finding, healthcare organizations should identify the specific clinical problem they intend to address and the population in which the system will operate.

An outpatient diabetes service may discover that most eligible patients undergo serum creatinine testing but relatively few have documented UACR results. Another organization may find that abnormal kidney function results are recognized but repeat investigations or specialist referrals are frequently delayed.

These situations require different solutions. In some settings, improving laboratory ordering, clarifying professional responsibilities or strengthening result follow-up may offer greater immediate value than introducing a complex predictive model.

Where AI is considered, its performance should be compared with existing clinical processes and, when relevant, simpler rule-based decision support. Local validation should include representative patient records rather than relying exclusively on performance figures supplied by the technology developer.

The reference standard used for evaluation must reflect the intended purpose of the system.

A tool designed to identify patients requiring further kidney assessment should be evaluated against clinically appropriate review criteria. A system claiming to identify confirmed CKD requires a reference assessment that adequately addresses accepted diagnostic criteria, including chronicity.

Previous diagnostic codes alone may be an inadequate reference standard when the purpose is to identify disease that has not yet been recognized or documented.

Evaluation should examine sensitivity, specificity, positive predictive value, false-positive findings and clinically important missed cases. It should also assess how the system handles missing UACR measurements, potentially acute changes in kidney function and records containing inconsistent or incomplete historical information.

Cases not flagged by the system also require appropriate independent review. Evaluating only AI-positive patients can conceal missed abnormalities and produce a misleading impression of performance.

Particular attention should be given to patients whose clinical circumstances may affect model performance, including older adults, people with low muscle mass and those with incomplete longitudinal records.

Healthcare organizations should also consider differences between facilities. Laboratory methods, clinical documentation, patient populations and referral arrangements may vary substantially, even within the same healthcare network. A model that performs well in one hospital may require further validation before being used elsewhere.

Technical validation is necessary, but it is not sufficient.

Organizations also need to determine whether identified patients actually receive appropriate clinical assessment. A system that generates substantially more alerts may increase workload without improving care if clinicians cannot review the findings or arrange necessary investigations.

The evaluation should therefore extend across identification, diagnostic assessment, risk stratification and follow-up.

Relevant measures include the proportion of eligible patients receiving appropriate kidney assessment, the proportion of flagged cases reviewed by a qualified professional, completion of indicated repeat investigations and the proportion of suspected cases whose diagnostic status is resolved.

For patients with confirmed CKD, organizations can examine whether appropriate risk classification, monitoring arrangements and management plans are documented. Where nephrology referral is indicated, evaluation should distinguish referrals issued from specialist assessments actually completed and recommendations communicated back to the referring clinician.

These measures require carefully defined denominators. Not every flagged patient needs a nephrology referral, and an increase in the number of CKD diagnoses is not automatically evidence of improved care.

Balancing measures are equally important. Unnecessary repeat testing, inappropriate diagnostic labels, avoidable referrals, additional clinical workload, delayed assessment of urgent abnormalities and financial costs to patients should be considered when evaluating the intervention.

Access to care also deserves attention. Patients receiving treatment across several facilities, living far from specialist services or facing financial barriers may have difficulty completing investigations and follow-up. These problems may not be visible in a dashboard that measures only laboratory orders or electronic referrals.

Where reliable longitudinal data are available, healthcare organizations may examine kidney function trajectories, clinically important kidney events, hospitalizations and cardiovascular outcomes. Such measures require appropriate follow-up periods and careful interpretation.

Changes observed after implementation cannot automatically be attributed to AI, particularly when laboratory policies, staffing, treatment availability or other elements of the care pathway have changed simultaneously.

Ongoing monitoring is necessary because performance may change as patient populations, data sources, clinical practice or model versions evolve. Organizations should establish procedures for reviewing errors, investigating safety concerns and modifying or suspending system use when necessary.

The most meaningful evidence of value will come from demonstrating that the complete AI-supported clinical pathway performs better than previous practice, rather than relying on the model's technical accuracy alone.

Legal, Professional and Data Governance in Viet Nam

AI-assisted CKD case finding operates within several areas of Vietnamese law, including medical practice, artificial intelligence, medical device management and personal data protection.

The Law on Medical Examination and Treatment No. 15/2023/QH15, as amended and reflected in Consolidated Document No. 26/VBHN-VPQH dated 26 February 2026, provides the broader legal framework for medical practice, professional responsibilities and patient care.

For CKD specifically, Ministry of Health Decision No. 2388/QD-BYT dated 12 August 2024 remains an important national professional reference. Healthcare organizations should ensure that their clinical pathways are consistent with applicable Vietnamese requirements and relevant professional guidance.

Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on 1 March 2026. Decree No. 142/2026/ND-CP, effective from 1 May 2026, provides detailed implementing provisions, including requirements relevant to AI risk classification.

Decision No. 33/2026/QD-TTg, effective from 15 August 2026, establishes the list of high-risk AI systems. The healthcare-specific entries in that list concern certain AI-assisted surgical and robotic surgical systems. AI-assisted CKD case finding is not expressly identified as a separate healthcare category in those entries.

However, this observation should not be interpreted as a blanket legal classification of all CKD-related AI applications.

The regulatory position of a particular system depends on its functions, intended purpose, degree of automation, deployment arrangements and the applicable legal criteria. A system that prioritizes patients for professional review may have a different regulatory and clinical risk profile from software that independently generates diagnostic conclusions or directly influences treatment decisions.

Clinical risk and statutory AI risk classification are related but distinct considerations. A system that is not expressly included in a particular high-risk listing may still create significant patient-safety concerns if it misses acute kidney deterioration, incorrectly labels patients or contributes to delayed care.

Healthcare organizations must separately consider whether software performing diagnostic, predictive or clinical decision-support functions falls within Viet Nam's medical device regulatory framework.

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

Circular No. 24/2026/TT-BYT should not be understood as establishing a standalone approval pathway specifically for AI-assisted CKD software. Whether medical device requirements apply depends on the characteristics and intended medical purpose of the particular product, together with the relevant legal definitions and classification rules.

A product-specific legal and regulatory assessment is therefore necessary before procurement, commercialization or routine clinical deployment.

Personal data protection presents another important responsibility.

AI-assisted CKD case finding may involve laboratory results, diagnoses, medication histories, demographic information and other identifiable health data. 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 for personal data processing and protection.

Healthcare organizations should identify the applicable lawful basis and defined purposes for processing, establish appropriate access controls and security safeguards, and clarify responsibilities when external technology providers handle patient information.

The use of clinical records for AI development, validation, retraining or commercial model improvement should be assessed separately where those activities involve different purposes or legal grounds from the original provision of patient care.

The existence of a clinical record does not, by itself, authorize unrestricted secondary use of identifiable patient information.

External hosting, outsourcing and cross-border processing arrangements also require assessment under applicable Vietnamese requirements. Relevant documentation and impact-assessment obligations should be considered according to the nature of the processing and the legal provisions that apply.

These issues should be addressed before identifiable clinical information is transferred to an external AI platform.

Healthcare organizations should also establish appropriate clinical governance arrangements covering system access, professional review of outputs, error reporting, change management and continuity of care when the technology is unavailable.

International clinical guidelines and AI governance principles can inform these arrangements, but they should not automatically be presented as binding obligations under Vietnamese law. Similarly, organizational safeguards proposed for AI-assisted CKD case finding should be distinguished from specific statutory requirements.

Building a Practical Implementation Pathway

A reasonable starting point for a Vietnamese healthcare organization may be a defined outpatient population, such as adults receiving ongoing care for type 2 diabetes or hypertension.

Rather than attempting to identify CKD across an entire hospital network immediately, the organization could first examine whether patients meeting established clinical criteria are receiving appropriate kidney assessment and follow-up.

A multidisciplinary team involving internal medicine, nephrology, laboratory medicine, nursing, information technology and quality management can review the existing pathway and identify where patients are most frequently missed.

The review may reveal that eGFR results are readily available but albuminuria assessment is incomplete. It may show that historical test results are difficult to retrieve or that no clear responsibility exists for reviewing abnormal findings after the patient leaves the facility.

These findings should guide the initial intervention.

The organization can define which patients are eligible for case finding, what information the system may use, which findings require clinical verification and who is responsible for reviewing the outputs. Urgent abnormalities should remain subject to established escalation procedures and should not be delayed by routine case-finding processes.

Patients with known CKD, acute kidney injury, kidney transplants, ongoing dialysis or other conditions requiring distinct management pathways should be handled according to their clinical circumstances rather than automatically entering an undifferentiated case-finding workflow.

During initial implementation, AI outputs can be evaluated alongside existing clinical records and standard assessment procedures without automatically changing patient management. Disagreements should be investigated to determine whether they reflect missing information, incorrect data interpretation, inappropriate thresholds or genuine model errors.

Routine clinical use should begin only after the organization has addressed the relevant clinical, operational, safety and regulatory requirements, with appropriate professional oversight and monitoring.

Subsequent evaluation should determine whether identified patients receive the necessary investigations and whether confirmed CKD leads to appropriate risk-based management and follow-up.

If missing UACR measurements are the principal problem, strengthening laboratory ordering and result review may be the most effective response. If the main difficulty is referral completion, improving communication and care coordination may offer greater benefit than further refinement of an AI prediction model.

Expansion to additional departments or healthcare facilities should depend on demonstrated performance, manageable workload, reliable clinical processes and acceptable patient-safety risk.

The organization should also maintain the ability to continue essential assessment and follow-up when the AI system is unavailable.

The objective is to improve the reliability of CKD care, not to make clinical practice dependent on a particular technology.

From Earlier Recognition to Better Kidney Care

AI-assisted CKD case finding offers a potentially useful way to identify patients whose kidney abnormalities might otherwise remain overlooked, particularly when relevant clinical information is distributed across multiple encounters and healthcare systems.

Yet identifying a possible case of CKD is not the same as confirming the disease, understanding its prognosis or ensuring that the patient receives appropriate care.

A patient with a newly reduced eGFR needs clinical assessment that considers potentially acute causes. Someone with persistent albuminuria needs the finding to be confirmed and interpreted in context. A patient with established CKD needs appropriate risk-based management, while a patient requiring nephrology assessment needs a referral process that supports meaningful follow-up.

None of these steps is guaranteed by an AI-generated finding.

The opportunity lies in connecting earlier recognition with reliable clinical decisions and continuity of care. That requires clear professional responsibilities, access to appropriate investigations, communication across services and an effective way to determine whether the necessary next steps have occurred.

For healthcare leaders, success should not be measured simply by the number of patients identified by AI.

A more meaningful assessment asks whether previously overlooked patients receive timely evaluation, whether unnecessary diagnoses are avoided, whether their risks are appropriately understood and whether the organization can demonstrate improvement in clinical follow-up.

AI-assisted CKD case finding becomes valuable when it helps healthcare professionals turn relevant information into safer, more coordinated kidney care.

Selected References and Regulatory Sources

Vietnamese Clinical Guidance

Ministry of Health of Viet Nam. Decision No. 2388/QD-BYT, dated 12 August 2024, issuing Diagnosis and Treatment of Chronic Kidney Disease and Certain Kidney Diseases (Hướng dẫn chẩn đoán và điều trị bệnh thận mạn và một số bệnh lý thận).

International Clinical Guidelines

Kidney Disease: Improving Global Outcomes (KDIGO) CKD Work Group. KDIGO 2024 Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease. Kidney International. 2024;105(4S)–S314.

Levin A, Ahmed SB, Carrero JJ, et al. Executive Summary of the KDIGO 2024 Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease: Known Knowns and Unknown Knowns. Kidney International. 2024;105:684–701. DOI: 10.1016/j.kint.2023.10.016.

American Diabetes Association Professional Practice Committee for Diabetes. 11. Chronic Kidney Disease and Risk Management: Standards of Care in Diabetes—2026. Diabetes Care. 2026;49(Suppl 1)–S260. DOI: 10.2337/dc26-S011.

European Society of Cardiology, in collaboration with the European Renal Association. 2026 ESC Guidelines for the Management of Cardiovascular Disease and Chronic Kidney Disease. European Heart Journal. 2026. DOI: 10.1093/eurheartj/ehag098.

Clinical AI and Digital Implementation Evidence

Wu J, Zheng Y, He Z, et al. Leveraging Time-Series Electronic Health Records With Large Language Models for Chronic Kidney Disease Diagnosis in Primary Care. BMJ Health & Care Informatics. 2026;33. DOI: 10.1136/bmjhci-2025-101830.

Zheng X, Hui M, Yang H, et al. Clinical Decision Support for Chronic Kidney Disease in Primary Care: A Cluster Randomized Clinical Trial. JAMA Network Open. 2026;9(5). DOI: 10.1001/jamanetworkopen.2026.11112.

Goldhaber SZ, Rashedi S, Pfeferman M, et al. Electronic Alert-Based Computerized Decision Support for Detection of Chronic Kidney Disease in Type 2 Diabetes Mellitus: CKD-DETECT. The American Journal of Medicine. 2026;139(9):1228–1236. DOI: 10.1016/j.amjmed.2026.04.022.

Relevant Vietnamese Legal and Regulatory Framework

National Assembly of Viet Nam. Law on Medical Examination and Treatment No. 15/2023/QH15, as amended. Consolidated Document No. 26/VBHN-VPQH, dated 26 February 2026.

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.

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 on Determination of Risk Levels and Management Measures for Medical Devices and Related Regulatory Amendments, effective 1 July 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.

Editorial and Clinical Notice

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

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

The approaches discussed for AI-assisted CKD case finding, diagnostic confirmation, risk stratification, referral follow-up and performance evaluation are organizational implementation considerations. They do not constitute a mandatory Vietnamese Ministry of Health AI implementation protocol, an independently validated clinical decision instrument or a determination of the regulatory status of any particular software product.

This article does not provide individualized medical or legal advice, establish a diagnosis, replace professional clinical judgment or constitute regulatory approval. Clinical assessment, diagnosis, treatment and referral decisions remain the responsibility of appropriately qualified healthcare professionals, consistent with applicable Vietnamese requirements, current clinical guidance and locally approved procedures.

The research studies discussed do not establish that AI-assisted CKD case finding necessarily improves kidney or cardiovascular outcomes in Vietnamese healthcare settings. Appropriate local validation, clinical oversight and ongoing monitoring remain necessary before claims of clinical effectiveness can be made.

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