Governing AI-Assisted Diabetic Retinopathy Screening in Viet Nam: Image Quality, Referral Completion and Patient Safety
IMPLEMENTATION INSIGHTS
Last reviewed: 25 October 2026
A patient with type 2 diabetes attends a routine appointment at a hospital outpatient clinic. During the visit, retinal photographs are taken using a digital fundus camera. An artificial intelligence system analyzes the images and identifies findings suggesting that further examination by an ophthalmologist is needed.
The result is available before the patient leaves. A referral is recommended, and the screening encounter is recorded as completed. Several weeks later, however, the patient has still not attended an eye examination.
In another consultation, the AI system cannot produce a reliable result because the retinal photographs are insufficiently clear. The patient is told that the imaging was unsuccessful but leaves without a definite plan for another examination.
Both situations illustrate an important challenge in AI-assisted diabetic retinopathy screening. A system may generate results quickly and identify clinically relevant abnormalities, but its contribution to patient care depends on what happens before and after the analysis.
The right patients must be offered screening. Images must be acquired correctly, results must be handled according to the system's intended use, and patients requiring further assessment need a practical way to receive it.
For hospitals and diabetes care services in Viet Nam, introducing AI-assisted retinal screening therefore involves more than installing a fundus camera and connecting it to an algorithm. It requires a reliable clinical process linking screening eligibility, image acquisition, result interpretation, referral and subsequent eye care.
Successful AI screening should not be judged only by how many abnormalities are detected. It should be judged by whether patients with clinically important findings receive appropriate follow-up care.
AI may improve access to screening without necessarily improving referral completion. Similarly, identifying more potentially serious abnormalities does not automatically establish that more patients have received treatment or avoided vision loss.
The important question is whether the technology helps healthcare organizations deliver a more complete, timely and reliable diabetic eye care pathway.
Diabetic retinopathy screening is an entry point to care
Diabetic retinopathy is a complication of diabetes involving damage to the retinal microvasculature. It can progress to visual impairment or blindness, particularly when clinically important disease is not recognized and managed appropriately.
The condition may develop without noticeable visual symptoms. Patients can have retinal abnormalities even when they report no difficulty seeing, making regular eye assessment an important component of diabetes care.
Screening aims to identify people who may have disease requiring further examination, monitoring or treatment. However, screening and comprehensive ophthalmological assessment are not the same activity.
A screening program identifies patients who may benefit from further care. Subsequent clinical assessment determines the nature and severity of the findings and whether additional monitoring or treatment is indicated.
The American Diabetes Association's Standards of Care in Diabetes—2026 recognizes retinal photography with remote expert interpretation or US Food and Drug Administration (FDA)-approved AI algorithms as appropriate strategies for improving access to diabetic retinopathy screening [1].
Importantly, the ADA also emphasizes that these programs need pathways for timely referral for comprehensive eye examination when indicated.
Retinal photography does not replace every aspect of comprehensive eye care. In-person examination remains necessary when images are inadequate, abnormalities require further assessment or the patient's circumstances warrant specialist review.
The ADA recommends an initial dilated comprehensive eye examination for people with type 2 diabetes at diagnosis and for adults with type 1 diabetes beginning five years after disease onset. Subsequent examination intervals depend on retinal findings, glycemic control and other relevant clinical factors. In appropriately selected patients with reassuring previous examinations and good glycemic control, screening every one to two years may be considered.
These recommendations provide an international professional reference. They should not be interpreted as establishing identical mandatory intervals or service arrangements for every healthcare organization in Viet Nam.
For an AI-assisted program, the practical implication is that the service must define its intended population and distinguish people suitable for routine screening from those requiring a different clinical pathway.
A patient with previously diagnosed advanced diabetic retinopathy, for example, may need ongoing specialist follow-up rather than another routine automated screening examination.
Similarly, new visual symptoms or other concerning findings should not be managed solely according to an automated screening result.
AI can support access to retinal assessment, but it should operate within an appropriate diabetes and eye care program rather than function as an isolated technology.
Different retinal AI systems have different clinical roles
AI applications for retinal imaging vary in their intended purpose, required inputs and level of autonomy.
Some systems are designed to produce a screening result without requiring an ophthalmologist to interpret every image individually, provided the examination falls within the system's authorized intended use.
Other systems assist healthcare professionals by identifying suspicious regions, assigning preliminary grades or prioritizing images for review.
Some applications also assess abnormalities beyond diabetic retinopathy, including findings associated with glaucoma or macular disease.
These functions should not be assumed interchangeable.
An autonomous system evaluated for detecting more-than-mild diabetic retinopathy in a defined population is not automatically validated for diagnosing every stage of retinopathy, identifying all macular diseases or assessing glaucoma.
Likewise, a system designed to assist professional image interpretation should not automatically be considered suitable for independent screening without that professional review.
Clinical terminology also matters.
More-than-mild diabetic retinopathy, referable diabetic retinopathy and vision-threatening diabetic retinopathy are related but not necessarily equivalent categories. Their definitions depend on the grading system, reference standard and intended purpose of the application.
Some systems may also identify findings associated with suspected diabetic macular edema. However, color fundus photography does not provide the same information as optical coherence tomography (OCT). An AI result should not automatically be treated as definitive confirmation or exclusion of macular edema.
These distinctions influence how results should be communicated and which patients require further assessment.
Before adopting a product, a hospital should establish exactly what the system is designed to detect, which patients it is intended for, what imaging equipment it supports and what professional involvement is required.
It is equally important to distinguish regulatory authorization in one country from the requirements applicable in another.
An autonomous AI application authorized by the FDA is not automatically permitted to replace professional image interpretation or other clinical responsibilities under Vietnamese requirements.
The proposed deployment model needs its own clinical, professional and regulatory assessment.
Image quality is part of patient safety
The reliability of AI-assisted screening begins before the software analyzes an image.
Fundus photographs must provide sufficient visualization of the relevant retinal structures for the intended assessment. Poor focus, inadequate illumination, small pupils, media opacity, patient movement and incorrect positioning can reduce image quality.
Cataracts and other conditions that interfere with the optical pathway may make it difficult to obtain photographs that can be interpreted reliably.
These problems deserve particular attention when AI is introduced into diabetes clinics or services where retinal imaging is not part of every staff member's routine practice.
Appropriate training in image acquisition is therefore a clinical safety consideration, not merely a technical requirement.
Personnel need to understand the imaging protocol, recognize common causes of poor-quality photographs and know when another image or a different form of clinical assessment is necessary.
The use of pupil dilation also requires consideration. Some cameras and AI applications support imaging without routine dilation, while others may require different arrangements depending on the patient's condition and the equipment being used.
Organizations should follow applicable professional guidance, product instructions and clinical assessment rather than assume that one imaging method is suitable for every patient.
Viet Nam already has specific professional guidance relevant to this process.
Decision No. 2557/QĐ-BYT, issued by the Ministry of Health on September 20, 2022, promulgated the technical procedure for non-fluorescent fundus photography used in the early detection of diabetic retinopathy [2].
The document addresses patient preparation, personnel, equipment, image acquisition and interpretation. It specifies verification of the correct patient, correct eye and correct clinical indication.
The described technique includes taking at least two photographs of each eye, with images covering the optic disc and posterior retinal region.
These details matter when hospitals assess whether their proposed imaging workflow is consistent with applicable Vietnamese professional guidance.
They should not, however, be treated as universal technical specifications for every international AI product. The organization must also understand the manufacturer's intended imaging protocol and whether the relevant camera–software combination has been evaluated appropriately.
Technical compatibility alone does not establish that the resulting photographs are suitable for reliable clinical screening.
An inconclusive result must lead to another appropriate step
An important measure of screening performance is the proportion of examinations that produce images of sufficient quality to generate a valid result.
This may be described through measures such as imageability or gradability, depending on the study or system being evaluated.
However, the denominator matters.
A system may appear highly accurate when evaluated only on good-quality photographs, while a substantial proportion of patients attending the clinic cannot obtain an interpretable result.
A prospective real-world study published by Hunfeld and colleagues in Scientific Reports in January 2026 provides a useful illustration [3].
The researchers evaluated the IDx-DR system among 875 patients with diabetes in a German clinical setting. They found that 10.5% of patients did not yield a fundus image without pupil dilation, while 26.1% of examinations were not analyzable by the AI system.
The study identified several factors affecting acquisition and analyzability, including pupil size, patient age, visual acuity and the person performing the examination.
When adequate images were obtained, the system demonstrated encouraging performance for detecting more severe diabetic retinopathy. However, the proportion of examinations that could not be analyzed showed why image acquisition and patient characteristics are important to real-world effectiveness.
These results describe a particular system, population and examination protocol. They should not be interpreted as the expected failure rate of every retinal AI application.
For healthcare organizations, the practical lesson is that screening performance must be evaluated across all patients who present for examination, not only those who produce usable images.
Hospitals should distinguish the number of people who attend screening, the number who undergo image acquisition and the number for whom the system produces a valid result.
A patient whose photographs cannot be assessed has not necessarily completed an effective screening examination.
The organization needs to determine what happens next. Depending on the circumstances, this may involve repeating the photographs, using another imaging method or arranging appropriate clinical eye assessment.
Repeated imaging attempts should also be reasonable for the patient. Someone with limited mobility, for example, may face substantial inconvenience if unsuccessful imaging results in repeated visits without a clear plan.
A screening program should not allow inconclusive results to remain unresolved indefinitely.
The same principle applies when one eye can be assessed but the other cannot. The record should reflect the limitation rather than treating an incomplete bilateral examination as a normal result for both eyes.
An inconclusive AI result requires an appropriate next step. It must not be treated as evidence that clinically important disease is absent.
Correct patient identification and result attribution remain essential
AI-assisted retinal screening depends on linking the correct photographs and results to the correct patient.
Even a technically accurate algorithm can contribute to unsafe care if images are associated with another person's record or if results from different examinations are incorrectly combined.
This risk becomes more important when photographs move between cameras, local computers, external platforms and hospital information systems.
Patient identification, examination dates, image laterality and result attribution should therefore be managed consistently.
The right and left eyes need to be identified correctly, and the screening record should accurately reflect any limitation affecting the examination.
Decision No. 2557/QĐ-BYT specifically addresses verification of the correct patient, eye and indication during the imaging procedure [2].
These requirements remain relevant when AI is introduced.
Software cannot compensate for incorrect patient identification or unreliable attribution of clinical information.
Negative and positive AI results require appropriate interpretation
False-negative findings are an important safety concern in any screening program.
An AI system may fail to identify clinically relevant abnormalities because of limitations in its detection performance, image quality, retinal appearance or differences between the patient population and the data used during development.
A negative result becomes particularly problematic if it is interpreted as confirmation that no further eye care is necessary.
An AI result indicating that referable diabetic retinopathy was not detected does not necessarily exclude every form of diabetic eye disease. It also does not rule out unrelated ophthalmic conditions.
Patients with new visual symptoms, concerning clinical findings or previously diagnosed retinal disease may require appropriate assessment regardless of the automated screening output.
Screening eligibility is especially important for populations that were not adequately represented in the evidence supporting a particular product.
These may include certain pediatric patients, pregnant patients with preexisting diabetes or people with known advanced retinal disease.
Such patients should be managed according to relevant professional guidance and the specific system's intended-use limitations rather than assuming that a routine automated pathway is appropriate.
A valid negative result also needs to be documented and incorporated into the patient's ongoing diabetes care plan.
The patient should remain within the appropriate periodic eye assessment arrangements. A reassuring screening result is not a reason to stop future eye care.
False-positive findings create a different concern.
A screening system may identify abnormalities that are not subsequently confirmed by an ophthalmologist. While these findings do not necessarily cause harm when appropriate assessment follows, they may increase patient anxiety, generate unnecessary visits and place additional pressure on specialist services.
The consequences become more significant when referral capacity is limited.
If AI produces a large number of positive results, the organization needs to understand whether referrals are clinically appropriate and whether patients with more urgent conditions continue to receive timely attention.
However, referral appropriateness cannot be judged solely by whether treatment is eventually provided.
A referral may be justified because further examination is necessary to exclude serious disease, even when the specialist ultimately determines that treatment is not required.
The evaluation should therefore consider the original indication, the information available at screening and the findings of subsequent clinical assessment.
The goal is to recognize disease appropriately without creating avoidable burdens for patients and the health system.
What international evidence tells us
Prospective studies have demonstrated that appropriately developed AI systems can identify diabetic retinopathy with clinically relevant diagnostic performance in defined populations.
An important early example is the pivotal trial published by Abràmoff and colleagues in npj Digital Medicine in 2018 [4].
The study enrolled 900 people with diabetes who had no previously diagnosed diabetic retinopathy at primary care clinics. It evaluated an autonomous AI system against a reference assessment involving retinal photography and OCT.
The system achieved a reported sensitivity of 87.2% and specificity of 90.7% for detecting more-than-mild diabetic retinopathy, with an imageability rate of 96.1%.
These findings supported the first FDA authorization of an autonomous AI diagnostic system for diabetic retinopathy detection.
However, the results describe a specific system, patient population, imaging protocol and reference standard. They should not be generalized to every retinal AI application.
More importantly, diagnostic accuracy does not establish that patients with positive screening results subsequently receive appropriate care.
This distinction has become increasingly relevant as research moves from algorithm development toward real-world implementation.
On April 17, 2026, Leigh and colleagues published a systematic review and meta-analysis in npj Digital Medicine examining referral uptake following diabetic retinopathy screening with AI-assisted care pathways [5].
The review included six eligible studies comparing AI-assisted screening pathways with existing approaches.
The pooled analysis found higher referral uptake in the AI-assisted groups, with a relative risk of 1.89 and a 95% confidence interval of 1.18 to 3.03.
These findings suggest that AI-supported screening pathways may help more patients attend recommended specialist assessment.
However, the studies were highly heterogeneous, with an I² value of 91.9%.
The comparison pathways differed. Some studies compared AI-supported targeted referrals with routine referral arrangements, while others compared AI with manual image grading or asynchronous teleophthalmology.
Four of the six studies also incorporated patient-facing behavioral interventions enabled by the immediate availability of AI screening results.
These interventions may have influenced whether patients attended subsequent appointments.
The authors considered the overall certainty of evidence low because of important differences between studies and potential sources of bias.
The findings therefore do not establish that the algorithm alone caused the improvement in referral uptake.
Instead, they suggest that the value of AI may depend on combining rapid screening results with appropriate referral arrangements, patient communication and practical follow-up support.
The review also does not establish that every AI-assisted screening program reduces visual impairment or prevents blindness.
Its most important contribution for healthcare organizations is showing why evaluation must extend beyond detection accuracy to consider what happens after an abnormal result is identified.
Performance may change across clinical settings
Another important question is whether a retinal AI system performs consistently when used outside the environment in which it was originally developed or evaluated.
A study by Quellec and colleagues, published in Scientific Reports on September 10, 2026, examined the performance of an AI system for diabetic retinopathy screening across different evaluation settings [6].
The researchers assessed the OphtAI system using several independent evaluation frameworks, including clinical datasets and screening environments with different populations and imaging arrangements.
The study demonstrated that performance can vary meaningfully according to the evaluation setting, even when an algorithm has performed well on established benchmark datasets.
This is particularly relevant to healthcare organizations introducing systems developed using data from another country or healthcare environment.
An algorithm may be technically capable of analyzing retinal photographs, but differences in camera characteristics, image quality, disease prevalence and patient populations may affect its practical performance.
These findings do not mean that AI screening is inherently unreliable.
They reinforce the importance of assessing a specific system under the actual conditions in which it will be used.
For hospitals, external validation and local verification should therefore be treated as important steps toward responsible implementation rather than administrative formalities.
What experience in Viet Nam tells us
AI-assisted diabetic retinopathy screening is already being implemented in Vietnamese healthcare services.
In July 2024, Orbis International, working with NIDEK and local healthcare partners, launched a project combining digital fundus photography with the Cybersight AI platform.
The project was intended to improve access to retinal screening and strengthen local eye care capacity.
According to a progress update published by NIDEK on October 30, 2025, a total of 14,463 patients had undergone imaging and AI-assisted screening across six project sites by the end of August 2025 [7].
Of these patients, 3,582, representing approximately 24.8% of those screened, were identified by AI as having signs associated with referable diabetic retinopathy, glaucoma and/or macular abnormalities.
The project reported that these patients were referred to ophthalmologists for further assessment and appropriate management.
These results provide a useful example of how retinal imaging, AI-supported assessment and specialist referral can be organized within diabetes care services.
However, the figures require careful interpretation.
The 3,582 patients were identified with one or more potentially referable eye abnormalities. They should not all be counted as confirmed cases of diabetic retinopathy.
Likewise, the number reported as referred does not establish how many attended specialist appointments, received confirmed diagnoses or completed indicated treatment.
The published progress update does not provide sufficient patient-level follow-up information to calculate those outcomes reliably.
These are program-reported implementation data rather than results of an independent controlled evaluation of clinical effectiveness.
Orbis has separately described how the program combined equipment, AI-assisted screening and training for local eye care professionals [8].
At Nghe An Hospital of Endocrinology, for example, the initiative supported the introduction of non-mydriatic fundus photography and training intended to strengthen local diabetic eye care services.
Such experience is valuable because it shows that the introduction of AI can be connected to broader improvements in service organization and professional capacity.
A further development was reported on October 1, 2026, when Roche Pharma Viet Nam and Orbis announced the handover of a non-mydriatic fundus camera with AI-assisted image analysis software to Thu Duc Regional General Hospital in Ho Chi Minh City [9].
The activity formed part of the Vì đôi mắt Việt initiative, which aims to strengthen diabetic eye care through collaboration between general healthcare services and ophthalmology providers.
The participating organizations described a model connecting early detection, specialist assessment, referral and subsequent treatment. The initiative also includes professional capacity-building activities.
This development is relevant because it reflects growing interest in integrating AI-assisted screening into a broader clinical care pathway rather than using the technology solely for image analysis.
However, the October 2026 announcement describes a program implementation milestone and its intended direction. It should not be interpreted as independent evidence that the program has already improved treatment completion or long-term visual outcomes.
Taken together, these Vietnamese experiences show that AI-assisted retinal screening can be introduced within healthcare services through collaboration between clinical organizations, technology providers and professional training partners.
They do not establish that every retinal AI system will perform equally well in Viet Nam or that additional screening has already reduced long-term visual impairment.
For healthcare organizations, the practical lesson is that technology adoption needs to be accompanied by suitable imaging competence, specialist capacity and referral arrangements.
The challenge is not simply to screen more people. It is to ensure that those requiring further care have a realistic opportunity to receive it.
A referral recommendation is not a completed referral
When AI identifies a finding requiring further assessment, several activities may need to occur before the patient receives the intended benefit.
The screening result must be communicated. A referral must be issued to an appropriate eye care service. The patient needs to obtain and attend an appointment. The specialist must then determine the findings and whether further monitoring or treatment is indicated.
These activities are connected, but they are not equivalent.
A hospital may record a referral as completed when the document is issued or electronically transmitted. Yet the patient may never attend the recommended appointment.
From a patient-safety perspective, the care pathway remains incomplete when the necessary assessment has not taken place.
This matters because patients may face practical barriers to follow-up.
Travel distance, transportation costs, appointment availability, health literacy and concerns about medical expenses can all influence attendance.
Some patients may not understand why an eye appointment is necessary when they have no visual symptoms. Others may believe that AI screening has already provided a definitive diagnosis.
A screening program should therefore explain the purpose of referral clearly and help patients understand what needs to happen next.
Where feasible, healthcare organizations may consider appointment coordination, reminders or assistance with accessing specialist services.
The appropriate arrangements depend on available resources, patient needs and institutional capabilities.
The essential requirement is a reliable method for identifying patients who have not received recommended assessment.
A referral pathway should not be considered successful simply because information has been transmitted. The organization needs to understand whether the patient reached an appropriate service and whether further clinical action occurred when indicated.
This is particularly important when screening and ophthalmology services operate at different facilities.
Responsibilities for communication, appointment coordination and follow-up should be understood by the participating organizations.
Referral urgency should reflect clinical significance
Not every abnormal retinal screening result requires the same response.
The urgency of further assessment depends on the suspected condition, clinical findings, patient circumstances and relevant professional guidance.
The ADA's 2026 recommendations identify diabetic macular edema, moderate or worse nonproliferative diabetic retinopathy and proliferative diabetic retinopathy as conditions warranting prompt referral to an ophthalmologist experienced in diabetic retinopathy management [1].
However, an AI system may not reliably distinguish all these conditions or establish their full clinical significance without specialist examination.
Healthcare organizations should therefore understand how the system's outputs correspond to the categories used in their referral arrangements.
The pathway should distinguish routine follow-up from cases requiring prompt or urgent clinical attention.
A patient with new or worsening visual symptoms may require assessment independently of the automated screening category.
Similarly, an image that cannot be interpreted should not automatically lead to the same follow-up arrangement as a valid negative result.
The next step should reflect the patient's condition and the reason the examination was incomplete.
An AI result generated within minutes may provide limited benefit if the patient cannot obtain an indicated eye examination within an appropriate period.
The purpose of AI-assisted screening is therefore not merely rapid detection. It is timely access to clinically appropriate care.
Patient communication should be clear and accurate
Patients may reasonably wish to understand why retinal photography is recommended, whether AI is being used and what the result means.
The explanation should reflect the actual purpose of the technology.
A positive screening result indicates a reason for further assessment within the system's intended detection scope. It does not necessarily confirm a diagnosis or establish that treatment is required.
A negative result does not guarantee that every retinal or ophthalmic disease has been excluded.
An inconclusive result means the examination has not provided sufficient information for its intended purpose and that another appropriate step may be necessary.
These distinctions should be explained without unnecessary technical language.
An automatically generated report may contain terms that patients find difficult to understand, particularly when they have limited experience with ophthalmic care.
Where appropriate, a qualified or appropriately trained healthcare professional should explain the result and the recommended next step.
Communication also needs to consider people who have existing visual impairment, difficulty reading or other barriers to understanding written instructions.
Family members or caregivers may provide practical assistance when appropriate and consistent with the patient's preferences and applicable privacy requirements.
The objective is not simply to deliver information. It is to help patients understand what they need to do next and why that step matters.
Professional responsibilities must remain clear
An AI-assisted diabetic retinopathy screening program commonly involves several professional groups.
Diabetes care clinicians may identify eligible patients and incorporate eye screening into ongoing diabetes management.
Personnel performing retinal photography must meet relevant professional and technical requirements and recognize when acquisition is unsuccessful.
Healthcare professionals responsible for the screening service need to understand the system's intended use, how results are managed and which findings require further review.
Ophthalmologists provide specialist assessment and determine appropriate management within their professional scope.
Quality and patient-safety personnel can help examine screening performance, referral completion and clinically important problems arising during implementation.
The division of responsibilities may differ between institutions, but the relevant functions need to be assigned clearly.
In Viet Nam, particular attention is required to the existing technical procedure for retinal photography.
Decision No. 2557/QĐ-BYT specifies the professional categories involved in performing the procedure and describes how results are interpreted [2].
Where photography is performed by specified personnel other than ophthalmologists under this procedure, the document provides for initial interpretation by a first-level reader and verification by an ophthalmologist.
When the interpretations differ, the images may require further specialist review. Unclear images may also require referral for clinical examination.
Where an ophthalmologist performs the photography, the procedure provides for that ophthalmologist to interpret the result, with further consultation or referral when appropriate.
These arrangements matter when healthcare organizations consider autonomous AI applications.
The existence of technology capable of producing results without individual image interpretation does not automatically establish that it may replace the professional interpretation and verification arrangements described in Vietnamese guidance.
Hospitals should evaluate the proposed workflow against applicable professional requirements, product-specific conditions and the legal framework governing the service.
Introducing AI does not eliminate responsibility for appropriate patient selection, reliable image acquisition, result management and referral.
Local validation must reflect actual screening conditions
Evidence from international studies is valuable, but performance established in one healthcare system may not transfer directly to another.
Differences in retinal cameras, imaging protocols, patient populations, disease prevalence and clinical workflows can influence performance.
A model evaluated using a particular camera or image format should not automatically be assumed to perform identically with every alternative device.
Even when software is technically compatible, differences in image characteristics may affect detection.
The organization should therefore evaluate the specific combination of camera, software and workflow being proposed.
The assessment should include photographs obtained during routine operations rather than only selected images of high technical quality.
Patient characteristics also matter.
Older patients, people with cataracts and those who have difficulty remaining still may be more likely to produce images that cannot be interpreted reliably.
Excluding these patients from performance calculations without reporting them separately may produce an incomplete picture of the service.
The organization should also examine performance across clinically relevant groups and different levels of retinal disease.
Sensitivity and specificity are useful measures, but their interpretation depends on the detection threshold, reference standard and study population.
Evaluation should distinguish patient-level performance from eye-level performance and account appropriately for missing or inconclusive examinations.
There is also an important methodological concern when assessing screening systems after deployment.
If only patients with positive AI findings receive specialist examination, false-negative cases may remain undiscovered.
This can introduce verification bias and make estimates of diagnostic performance unreliable.
An appropriate validation or audit design should therefore consider how clinically important disease among patients with negative results can be assessed using suitable methods.
Hospitals do not necessarily need to repeat every international clinical trial. However, they should obtain sufficient evidence that the particular system is suitable for its intended use and local circumstances.
What hospitals should measure after introducing AI
Healthcare organizations should define their expected improvement before introducing AI-assisted diabetic retinopathy screening.
One objective may be to increase screening coverage among eligible patients. Another may be to reduce delays in identifying referable abnormalities or improve attendance at specialist appointments.
These objectives require different measures.
Screening coverage is an important starting point. The organization should understand the proportion of eligible patients who receive the intended examination within the applicable clinical interval.
The denominator should reflect the eligible population, rather than only patients who attended the screening service.
Image quality also needs to be monitored.
Relevant measures include the proportion of examinations producing interpretable images, the number requiring repeat acquisition and the proportion remaining inconclusive.
These measures should be interpreted in relation to patient characteristics, equipment and staff experience where appropriate.
Diagnostic performance is another consideration.
Sensitivity and specificity for the intended detection threshold can provide useful information when a suitable reference standard and evaluation design are available.
However, performance should not be judged only among patients whose images are interpretable while overlooking those who cannot obtain a valid result.
The organization should also measure what happens after an abnormal result is generated.
The number of patients identified as requiring further assessment should be distinguished from the number for whom referrals are issued.
The number who attend specialist examination should be recorded separately.
Where feasible, hospitals should also examine how many patients receive a confirmed diagnosis, how many require treatment and whether indicated care begins within a clinically appropriate period.
A completed specialist examination does not necessarily mean that treatment is required. Equally, a patient who receives no treatment has not necessarily experienced an unsuccessful referral.
The appropriateness of the subsequent clinical decision matters.
For patients who do not attend, the organization should assess whether follow-up arrangements can identify and address avoidable barriers.
False-positive and false-negative results should also be considered through appropriate quality and patient-safety review.
Unnecessary referrals may increase patient burden and specialist workload, while missed clinically important disease may delay necessary care.
Another consideration is whether the program improves access equitably.
Hospitals may examine screening completion, image quality and referral attendance across relevant patient groups, locations or service settings to identify differences requiring attention.
Long-term visual outcomes remain important. However, meaningful evaluation generally requires reliable follow-up and consideration of factors beyond AI use, including diabetes management, disease severity, treatment availability and subsequent care.
A short implementation study cannot establish that additional screening has prevented future blindness.
The number of AI-detected abnormalities is not a sufficient measure of successful implementation.
A meaningful evaluation combines screening coverage, image quality, appropriate referral, completed specialist assessment and subsequent clinical care.
Monitoring should extend beyond routine performance reports
Once AI-assisted retinal screening becomes part of routine care, performance may be affected by changes in equipment, software, staffing or patient characteristics.
Replacing a camera, modifying the imaging protocol or updating software may change image quality or the behavior of the AI application.
Hospitals should understand when such changes require further review or verification before continued use.
The same applies when the screening service expands into new facilities or patient populations.
A system evaluated in a specialist-supervised environment may require additional assessment before use in a different clinical setting.
Monitoring should also include concerns raised by frontline personnel.
Repeated inconclusive results, unexpected changes in referral rates or suspected missed retinal disease may indicate that the screening process needs investigation.
The organization should have a practical way to examine these findings and determine whether changes to training, equipment, workflows or AI use are necessary.
However, every disagreement between an AI result and a subsequent specialist assessment should not automatically be classified as an AI-related adverse event.
Review should consider the original image quality, intended detection target, information available at the time and clinical significance of the difference.
The aim is to identify problems that matter for patient safety and improve the screening service, rather than assume that every disagreement represents a system failure.
Vietnamese professional and legal responsibilities
AI-assisted diabetic retinopathy screening operates within Viet Nam's existing framework for medical examination and treatment, ophthalmic practice, artificial intelligence, medical devices and personal data protection.
These frameworks address related but distinct responsibilities.
A healthcare organization should not assume that the technical availability of an AI application or its authorization in another country automatically permits every proposed clinical use in Viet Nam.
Medical examination, treatment and ophthalmic practice
The Law on Medical Examination and Treatment No. 15/2023/QH15, effective January 1, 2024, as subsequently amended, provides an important legal foundation for professional practice, healthcare services and patient rights in Viet Nam [10].
The Ministry of Health has also issued professional guidance directly relevant to diabetic retinopathy.
Decision No. 2558/QĐ-BYT, dated September 20, 2022, promulgated the guidance on diagnosis, treatment and management of diabetic retinopathy [11].
The document addresses early detection, clinical examination, disease classification, treatment and the organization of care across healthcare service levels.
It also recognizes that AI technologies have been developed and applied internationally to assist with reading and classifying retinal photographs.
However, this reference should not be interpreted as blanket authorization for every AI application or as establishing that autonomous AI may replace existing clinical responsibilities.
Decision No. 2557/QĐ-BYT, issued on the same date, provides the technical procedure for non-fluorescent fundus photography used in the early detection of diabetic retinopathy [2].
The procedure addresses professional personnel, patient preparation, image acquisition, interpretation and further review of uncertain findings.
Its arrangements for professional interpretation and verification are especially relevant when hospitals consider introducing AI.
The organization should determine whether its proposed workflow complies with applicable professional requirements, the facility's authorized services, staff qualifications and the regulatory status of the selected product.
AI assistance does not independently expand a practitioner's legally permitted professional scope or remove the responsibilities of healthcare facilities and clinical personnel.
Artificial intelligence regulation
Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on March 1, 2026 [12].
Decree No. 142/2026/NĐ-CP, effective May 1, 2026, provides implementing provisions concerning AI activities, including risk classification and related responsibilities [13].
Decision No. 33/2026/QĐ-TTg, effective August 15, 2026, promulgates the list of high-risk AI systems [14].
The healthcare-specific entries expressly identified in that list concern defined AI-assisted surgical and robotic surgical applications. Ordinary AI-assisted diabetic retinopathy screening software is not separately identified among those entries.
However, this should not be interpreted as a definitive determination that every retinal AI application is legally low-risk.
The classification and applicable obligations depend on the specific system's functions, intended purpose, deployment arrangements and relevant legal criteria.
A system intended to support professional image interpretation may also have a different clinical and regulatory profile from one designed to produce autonomous screening results.
Statutory AI risk classification must be distinguished from an organization's clinical risk assessment.
A system may create significant patient-safety concerns if inappropriate outputs lead to missed disease or delayed specialist care, regardless of whether it falls within a particular statutory high-risk category.
Healthcare organizations should therefore assess both the applicable legal requirements and the clinical consequences of deploying the technology.
Medical device regulation
Medical device requirements may apply to fundus cameras and AI software according to their intended medical purpose, functions and regulatory classification.
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 framework [15].
Circular No. 24/2026/TT-BYT, effective July 1, 2026, further addresses risk determination and management measures for products classified as medical devices [16].
Hospitals should assess the regulatory status of the specific camera, software and relevant integrated functions proposed for clinical use.
The regulatory status of a fundus camera should not automatically be assumed to establish the status or permitted use of an AI application connected to it.
Likewise, foreign authorization, certification or published clinical validation does not independently demonstrate compliance with all Vietnamese requirements.
The assessment must be specific to the actual product and intended use.
Personal data protection
Retinal photographs and screening records may be associated with identifiable patient information, including medical history, diabetes status and examination results.
The Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP took effect on January 1, 2026 [17, 18].
Healthcare organizations should assess the lawful basis and purposes of processing, applicable consent requirements or exceptions, access controls, information security, retention and arrangements with technology providers.
Particular attention is needed when retinal images are transferred to external platforms or processed through cloud services.
The organization should understand where information is stored, who can access it and whether the technology provider intends to use it for purposes beyond the immediate screening service.
Secondary use for algorithm development, model training or commercial improvement requires separate assessment of the applicable legal and institutional requirements.
The use of AI for screening should not automatically be treated as authorization to reuse identifiable patient information for unrelated purposes.
Where a system is being evaluated as part of clinical research, applicable research ethics and participant protection requirements should also be considered separately.
These legal considerations should inform technology selection and implementation without replacing professional judgment or patient-safety evaluation.
A practical direction for healthcare organizations in Viet Nam
For Vietnamese hospitals and diabetes care services, responsible AI-assisted retinal screening should begin with an assessment of the existing care pathway.
The organization needs to understand how patients become eligible for screening, how they are invited or scheduled and where examinations are performed.
It should also establish how photographs are acquired, who interprets results and how patients requiring specialist assessment are identified and referred.
If the existing process cannot reliably identify patients who need screening, adding AI may not resolve the underlying problem.
Similarly, if patients with abnormal findings frequently fail to reach specialist services, faster detection alone may provide limited benefit.
Hospitals should therefore define the improvement they expect.
An organization may wish to increase screening coverage in diabetes services where ophthalmologists are not routinely available. Another may aim to improve the proportion of patients who attend specialist appointments after referral.
These objectives are related but require different evaluation methods.
Clinical leaders from diabetes care and ophthalmology should participate in selecting the intended use case and reviewing the evidence.
Personnel involved in imaging should receive training consistent with applicable professional requirements and the specific equipment being used.
The organization should also establish how results will be communicated, how referrals will be initiated and what happens when a patient does not attend the recommended examination.
A supervised introduction may be appropriate where permitted by applicable requirements and supported by sufficient clinical oversight.
During implementation, the hospital can assess whether the technology performs as expected and whether the referral pathway operates reliably.
A practical fallback is also necessary when AI is unavailable, imaging equipment fails or photographs cannot be interpreted.
Patients should continue to have access to appropriate clinical assessment through established arrangements.
The goal is not simply to demonstrate that the software works. It is to determine whether the technology improves an existing care process without introducing avoidable delays, unnecessary referrals or new risks for patients.
From successful screening to better diabetic eye care
AI-assisted diabetic retinopathy screening offers an opportunity to improve access to eye care for people living with diabetes.
International evidence demonstrates that appropriately developed systems can identify clinically relevant retinal abnormalities, while newer research suggests that AI-supported care pathways may improve attendance at recommended specialist appointments.
Experience in Viet Nam also shows that AI-assisted retinal imaging can be introduced into diabetes care services through collaboration between healthcare organizations, technology providers and professional training partners.
However, the success of these programs should not be assessed solely by the number of photographs taken or abnormalities identified.
Appropriate patient selection, image quality, correct identification, reliable interpretation and management of inconclusive results remain essential.
A positive screening result needs an appropriate referral. A referral needs a practical pathway to specialist examination. When treatment or further monitoring is indicated, the patient needs access to suitable care.
These activities cannot be replaced by algorithmic accuracy.
For hospitals in Viet Nam, responsible AI implementation means connecting the technical performance of retinal screening with the responsibilities of clinical care delivery.
Healthcare organizations need to understand who requires screening, who may be missed, whether images are interpretable and what happens after results are generated.
They should also recognize the limitations of current evidence and avoid suggesting that increased detection alone proves a reduction in blindness or improved long-term visual outcomes.
The most meaningful measure of success lies beyond the screening encounter.
AI-assisted diabetic retinopathy screening creates clinical value when reliable detection is connected to appropriate referral, completed specialist assessment and timely care for patients who need it.
Professional Education Notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized medical, ophthalmic, diabetes management, legal, regulatory or patient-safety advice.
The opening clinical scenarios are illustrative and do not represent documented patient cases.
This article focuses on AI-assisted retinal photography for diabetic retinopathy screening, primarily among adults receiving routine diabetes care, and on the subsequent referral pathway. It does not establish the effectiveness of AI systems for diagnosing all retinal or ophthalmic diseases, making autonomous treatment decisions or replacing comprehensive ophthalmological examination.
Different AI products have different intended uses, target populations, supported imaging systems and regulatory authorizations. Evidence supporting one application should not automatically be generalized to another.
International recommendations from the American Diabetes Association and other professional organizations reflect particular evidence assessments and healthcare contexts. They should not automatically be interpreted as binding requirements for healthcare organizations in Viet Nam.
The organizational approaches described, including image quality monitoring, management of inconclusive results, referral coordination, professional review and evaluation of subsequent care, are proposed considerations for responsible implementation and healthcare quality improvement. They are not mandatory Vietnamese Ministry of Health AI screening protocols or independently validated assessment instruments.
Existing Vietnamese professional guidance concerning fundus photography and diabetic retinopathy management must be considered separately from the regulatory status and intended use of an individual AI application. References to AI in professional guidance do not automatically authorize autonomous replacement of required clinical assessment or interpretation.
Program-reported screening and referral figures should not be interpreted as independently confirmed disease diagnoses, completed treatment episodes or proof of reduced visual impairment.
Evidence showing improved screening accuracy, coverage or referral uptake does not automatically establish that an AI-assisted program reduces blindness or improves long-term visual outcomes.
The applicable legal and regulatory obligations associated with individual AI systems, medical device software, ophthalmic services, professional practice, personal data processing and clinical research require separate assessment under the relevant Vietnamese legal framework.
References
American Diabetes Association Professional Practice Committee for Diabetes. 12. Retinopathy, Neuropathy, and Foot Care: Standards of Care in Diabetes—2026. Diabetes Care. 2026;49(Suppl 1)–S276. DOI: 10.2337/dc26-S012.
Ministry of Health of Viet Nam. Decision No. 2557/QĐ-BYT Promulgating the Technical Procedure for Non-Fluorescent Fundus Photography for Early Detection of Diabetic Retinopathy. Issued September 20, 2022.
Hunfeld E, Tayar A, Paul S, et al. Real-World Performance of the AI Diagnostic System IDx-DR in the Diagnosis of Diabetic Retinopathy and Its Main Confounders. Scientific Reports. 2026;16:4349. DOI: 10.1038/s41598-026-36970-9.
Abràmoff MD, Lavin PT, Birch M, Shah N, Folk JC. Pivotal Trial of an Autonomous AI-Based Diagnostic System for Detection of Diabetic Retinopathy in Primary Care Offices. npj Digital Medicine. 2018;1:39. DOI: 10.1038/s41746-018-0040-6.
Leigh JA, Sherrington A, Barber ARJ, Turner AW, Kidd M, Powell J, Pope C. Referral Uptake After Diabetic Retinopathy Screening With Artificial Intelligence-Assisted Care Pathways: A Systematic Review and Meta-Analysis. npj Digital Medicine. 2026;9:468. Published April 17, 2026. DOI: 10.1038/s41746-026-02616-3.
Quellec G, Lamard M, Matta S, et al. Performance of AI-Based Diabetic Retinopathy Screening Is Highly Dependent on Evaluation Setting: A Five-Year, Multi-Framework Study. Scientific Reports. Published September 10, 2026. DOI: 10.1038/s41598-026-70676-2.
NIDEK Co., Ltd. Progress Update on the Diabetic Retinopathy Screening Project in Vietnam. Corporate news release. Published October 30, 2025.
Orbis International. 14,000 Screened for Diabetic Retinopathy Using AI in Vietnam. Program update. Published September 29, 2025.
Roche Pharma Viet Nam. Roche Pharma Viet Nam and Orbis International in Viet Nam Donate Fundus Imaging Equipment and Support Professional Capacity Building Under the “Vì đôi mắt Việt” Program. Corporate news release. Published October 1, 2026.
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. Decision No. 2558/QĐ-BYT Promulgating the Guidance on Diagnosis, Treatment and Management of Diabetic Retinopathy. Issued September 20, 2022.
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.