Abstract

Age-related macular degeneration (AMD) is a major cause of vision loss in older adults. However, early detection is key to managing the condition. It is estimated that 288 million people worldwide will be affected by AMD by 2040. Traditional screening methods rely on human interpretation of retinal images, which can be time-consuming, expensive and prone to error. Artificial intelligence (AI), particularly deep learning, offers a promising solution by automating the analysis of retinal scans. AI can quickly detect early signs of AMD with high accuracy, enabling faster diagnoses and better outcomes for patients at potentially a lower cost. This technology may enhance screening efficiency and improve patient care. In this paper, we aim to synthesise the published literature regarding the potential benefits and implications of the implementation of AI in AMD screening, deconstructing case studies, cost effectiveness, performance and legal challenges related to AI in AMD. The purpose of using AI as a screening tool for AMD is to enhance the early detection and diagnosis of the disease, which is crucial for preventing vision loss. By leveraging AI, healthcare professionals can increase accuracy; AI algorithms can analyse retinal images more consistently and precisely than human clinicians, reducing the chances of missed or inaccurate diagnoses. AI can also accelerate the screening process by rapidly processing large data volumes for streamlined identification of at-risk patients to promptly commence medical intervention. AI can reduce healthcare costs by automating the screening process to save time and resources, especially for under-equipped settings in less developed countries. Finally, due to the early detection, AI can improve patient outcomes. Early detection allows for earlier treatment, which can slow the progression of AMD and prevent severe vision loss. 

Introduction

Age-related macular degeneration (AMD) is a leading cause of vision loss among older adults, posing a significant burden on individuals and healthcare systems worldwide. It is estimated that 288 million people worldwide will be affected by AMD by 2040 (1). Early detection and timely intervention are crucial to slowing the progression of this degenerative eye condition. In recent years, artificial intelligence (AI) has emerged as a transformative tool in healthcare, offering new possibilities for improving screening and diagnostic accuracy. By analysing retinal images with high speed and precision, AI-based systems have the potential to identify early signs of AMD more efficiently than traditional methods. This integration of AI into ophthalmology not only enhances the accessibility and affordability of screening services but also supports clinicians in making more informed decisions. As such, the use of AI as a screening tool for AMD represents a major advancement in preventive eye care and highlights the broader impact of digital innovation on healthcare delivery.

AMD is categorised into early, intermediate and late stages. Early and intermediate AMD is the most common and least severe form, defined by pigmentary abnormalities in the macula and accumulation of extracellular aggregates, which may remain asymptomatic but increase the risk of progression. Late AMD occurs in two distinct forms. Dry (atrophic) AMD, often referred to as geographic atrophy (GA), is characterised by the progressive death of retinal pigment epithelium (RPE) cells, photoreceptors and choroidal capillaries, leading to slowly enlarging central scotomas. Wet (neovascular) AMD, in contrast to the slow process of GA, can damage central vision acutely through choroidal neovascularisation, with abnormal vessel growth causing exudation, haemorrhage and fibrotic changes. The visual consequences of AMD are substantial. Patients may experience blurred or distorted central vision (metamorphopsia), reduced contrast sensitivity, impaired dark adaptation and difficulty with daily tasks such as reading or recognising faces. The peripheral vision generally remains intact, but central vision loss severely compromises one’s independence and quality of life (1).

IMPACT ON HEALTHCARE

AMD mainly affects people over 50 years of age. This disease has a significant impact on quality of life and places a growing burden on healthcare systems around the world. It damages the macula, the central part of the retina, leading to a gradual loss of sharp, central vision. As the global population ages, detecting and treating AMD early becomes increasingly important to slow the disease’s progression and preserve vision.

Traditionally, specialists diagnose AMD by manually interpreting retinal images. This process can take a long time, be inconsistent and may not be accessible in underserved areas. Recently, AI has become a powerful tool in eye care, offering quicker, more accurate and scalable ways to spot early signs of AMD. AI algorithms, especially those that use deep learning, can examine thousands of retinal images with great accuracy, often detecting subtle changes that might be missed by human eyes.

Using AI in screening not only improves diagnostic accuracy but also increases access to care, especially through telemedicine and primary care settings. This approach ensures earlier intervention and better use of specialist resources. Moreover, AI acts as a helpful support tool for clinicians, improving decision-making consistency and enabling more personalised care.

I. CASE STUDIES

DIAGNOSTICS

Colour fundus photos are at the forefront of modern retinal imaging, which take a high resolution of the fundus of the eye, the interior of the rear of the eye, and analyse the retina, optic nerve and choroid. The macula is a sensitive part of the retina responsible for converting the focused light into signals to be processed by the brain (5,6). However, an emerging competitor to fundus imagery is optical coherence tomography (OCT), which uses infrared light to create a 3D map of the layers of the eye. While both imaging techniques are able to highlight the presence of AMD, fundus imagery is more accessible globally, while OCT produces more detailed imagery for data processing (6,7). This discrepancy creates an inconsistency for AI integration; different systems have to be developed, tested and implemented according to the type and quality of photos used, creating a significant gap between developed and under-resourced medical environments.

Deep learning (DL) models are the primary AI structure utilised in ophthalmology due to the image-based nature of diagnoses. Of them, convolutional neural networks (CNNs) are the most prominent form, a model that applies a series of filters on the image to detect certain patterns and structures within the image (8,9). A range of CNN and similar DL networks have been developed globally to accommodate a range of nuances: automated or semi-automated depending on the availability of medical expertise, single-condition or investigation of a series of common eye conditions and integrating medical records to account for sociodemographic factors (10). Some AI models have seen remarkable success in comparison to human graders – a promising endeavour especially for poorly equipped healthcare systems. However, a key limitation is a general sense of hesitancy when translating a tested AI model into a practical implementation; much work needs to be done to enable the transition of AI into everyday AMD screening.

When analysing AI data, the primary analytic is the area under the curve (AUC), which plots sensitivity, the accuracy of diagnosing true positives, against specificity, the accuracy of diagnosing true negatives (11,12). A key note is that in a medical scenario, false negatives must be minimised, highlighting a priority of sensitivity over specificity. Most studies employ a comparison of AI-analysed data to human graders, although there are inconsistencies in grading standards including an adjudicated ‘gold-standard’ (13).

Table 1: Diagnostic accuracy metrics for AI-enabled AMD screening.

As analysed from a range of studies, AI is a reliable diagnoser of AMD, and while there is still room for improvement in developing networks, an implementation of one of many systems into a healthcare system would be promising without concerns about significant misdiagnosis volume. The major concern is the feasibility of single-condition AI analysers; health providers may not save time if only one condition can be screened by the AI as patients may have another, potentially more adverse condition.

TREATMENT PLANNING AND FUTURE CARE

The treatment method and planning for age-related macular degeneration are highly dependent on the stage of the disease. In the early and intermediate phases, interventions focus on prevention and lifestyle modification. For instance, smoking cessation is considered one of the most effective measures, as smoking not only increases the likelihood of developing AMD but also reduces treatment success in later stages (13).

Other modifiable risks include high body mass index, hypertension and cardiovascular disease, all of which accelerate retinal degeneration (14). Nutritional strategies have also been explored, with diets rich in antioxidant vitamins, zinc and carotenoids – commonly sourced from fish and leafy or yellow vegetables – being associated with a reduced risk of progression to advanced AMD (15). The Age-Related Eye Disease Study 2 (AREDS2) was a large, multicentre randomised clinical trial that tested modifications to the original AREDS supplement formulation. After evaluating the addition of lutein, zeaxanthin and omega-3 fatty acids, the trial demonstrated that supplementation with lutein and zeaxanthin, in place of beta-carotene, significantly improved safety and provided protective effects against progression to late AMD. This established the AREDS2 formulation as the current standard care for nutritional supplementation, particularly for patients with intermediate AMD in one or both eyes (16). Nonetheless, these measures primarily serve as preventive strategies and evidence remains mixed, as several studies have concluded that neither smoking, dietary modification nor supplementation shows a significant link with the onset of early AMD.

In late-stage AMD, approaches diverge between atrophic (dry) and neovascular (wet) forms. Currently, no effective treatment exists for dry AMD, leaving patients reliant on supportive measures and low-vision rehabilitation. By contrast, wet AMD can be managed with injections of anti-vascular endothelial growth factor (anti-VEGF) drugs, such as ranibizumab (approved in 2007), aflibercept (approved in 2012) and brolucizumab (approved in 2020). These agents reduce blood vessel permeability, which inhibits the leakage of intraretinal fluid from choroidal vessels, the primary cause of worsening exudative AMD. It is demonstrated that vision clarity remains stable in more than 70% of treated eyes, with fewer than 20% improving in visual acuity after initial treatments with anti-VEGF (17). Although highly effective, they are expensive, averaging 1,000 euros per injection, and carry a small risk of endophthalmitis (0.029%). To balance outcomes with treatment burden, strategies such as “pro re nata” (PRN), where injections are administered only when disease activity is detected by OCT or vision loss, and “treat-and-extend”, which progressively lengthens the interval between doses, have been developed to personalise therapy and reduce unnecessary costly injections (18). During treatments, monitoring disease activity, which refers to exudative changes arising from macular neovascularisation, using optical coherence tomography (OCT) is critical. In cases where drug resistance emerges, switching between anti-VEGF agents has shown success in restoring response (19). 

Recent research demonstrates how artificial intelligence can further enhance treatment planning and patient care. AI-based cloud platforms are developed and tested, enabling OCT images to be uploaded and automatically processed, generating diagnostic and treatment recommendations. Such systems increase efficiency, support personalised treatment plans and are ideally suited for telemedicine applications. Validation results showed that their convolutional neural network (CNN) models achieved high accuracy and sensitivity in AMD detection, with VGG16 reaching 91.40%, InceptionV3 92.67% and ResNet50 90.73%, all surpassing 90% (20). These findings suggest that AI integration into clinical care not only facilitates earlier detection but also provides clinicians with data-driven guidance to optimise therapy.

TIME EFFICIENCY

According to Chen et al. (21), AI-assisted image-based diagnosis of AMD demonstrated improvements not only in diagnostic accuracy but also in time efficiency compared with conventional manual diagnosis performed by clinicians alone. In this diagnostic study, 24 clinicians from 12 institutions evaluated 2,880 AMD risk features across 240 patient samples over multiple rounds, with and without AI assistance. 

In contrast to previous studies, this investigation prioritised practical applicability by evaluating not only AI-only and manual diagnoses but also manual plus AI as a third diagnostic approach, enabling a more comprehensive comparison with conventional manual diagnosis. In Round 1, the addition of AI assistance led to a mean diagnostic time reduction of 10.3 seconds, decreasing from 39.8 seconds to approximately 29.5 seconds per patient (95% CI, −15.1 to −5.5 seconds). Even in Rounds 2 through 4, where clinician performance improved due to learning effects, AI-assisted diagnosis remained significantly faster – by 6.9 seconds (95% CI, 0.2–13.7 seconds) to 8.6 seconds (95% CI, 1.8–15.3 seconds).

These findings suggest that AI-assisted workflows can help reduce clinician workload without compromising – and potentially improving – diagnostic accuracy, particularly in high-volume screening contexts. Moreover, the integration of AI has the potential to increase the number of patients that can be screened within a given timeframe, thereby contributing to the earlier detection of age-related macular degeneration and improved public health outcomes.

II. COST EFFECTIVENESS

PROVIDER PERSPECTIVE

The cost-effectiveness of AI in ophthalmology is a growing topic of interest, particularly as healthcare systems seek to improve outcomes while managing limited resources. From a provider’s perspective, AI presents both significant upfront costs and promising long-term savings. Initial investments often include purchasing software licenses, such as IDx-DR for diabetic retinopathy screening, as well as upgrading necessary hardware like retinal cameras, servers and GPUs. Additionally, integrating AI tools into a provider’s electronic health record (EHR) system can be complex and costly.

However, these initial expenses may be offset by long-term cost benefits. AI reduces the need for ophthalmologists to spend time on routine screenings, allowing for better task-shifting to non-specialist staff. This streamlining decreases the referral burden by reducing unnecessary patient transfers to higher levels of care. Perhaps most importantly, AI can enable earlier disease detection, which helps prevent costly interventions down the line; for instance, minimising the need for expensive anti-VEGF injections in late-stage AMD.

In terms of operational efficiency, AI dramatically improves the speed of diagnoses and triaging, particularly in high-volume screening programmes. By automating routine assessments, AI can handle larger patient loads without sacrificing quality, leading to fewer missed diagnoses. These improvements not only contribute to better clinical outcomes but also result in reduced long-term treatment costs, making care more sustainable over time.

From a revenue standpoint, AI can help providers increase throughput, enabling more billable services within the same timeframe. Some AI tools can be reused, enhancing return on investment. Furthermore, AI-driven tele-ophthalmology solutions can expand access to eye care in underserved or rural areas, reaching patients without the need for a full-time ophthalmologist on site.

AI also supports scalability and smarter resource allocation. It facilitates large-scale screening initiatives without requiring a proportional increase in specialist manpower. In areas facing a shortage of ophthalmologists, AI acts as a force multiplier, ensuring broader population coverage and more equitable access to eye care services.

Despite these advantages, providers must also consider the limitations and ongoing costs. The initial investment remains high and continuous maintenance, software updates and clinical validations are required to ensure reliability. Moreover, AI tools are currently most effective for specific diseases – such as diabetic retinopathy, AMD and glaucoma suspects – and cannot replace comprehensive clinical eye exams.

In conclusion, while AI involves significant upfront costs, it offers compelling long-term value by increasing efficiency, reducing workload and improving patient outcomes. For many providers, especially those involved in large-scale screening or working in resource-limited settings, AI represents a cost-effective and scalable solution, provided its limitations are carefully managed.

PATIENT PERSPECTIVE

For the patient, the cost of AI in age-related macular degeneration is a result of dollars, minutes, tension and vision outcomes. In the US, yearly healthcare costs per patient often add up to the mid-five figures, primarily in the form of constant anti-VEGF injections, imaging and follow-ups. One study estimated roughly $24,000 per year, even before travel or lost wages were considered. Anti-VEGF agents by themselves contribute to over one-third of direct costs, much of which is captured in the form of premiums, copays or drug maximums. “This can be covered by biosimilars and cheaper agents like bevacizumab, but follow-up and monitoring remain expensive, with evidence showing that patients initiated on ranibizumab or aflibercept had 268% and 272% higher costs, respectively, compared with those on bevacizumab.” (28)

AI offers the patient a mechanism of cost deferral. Large-scale risk triage may detect disease before it advances, reducing the number of patients to continue undetected resulting in a transition from dry to wet AMD. Deep-learning OCT algorithms may detect high-risk cases at specialist levels, enabling earlier treatment. AI tele-ophthalmology also removes travel, recurring work absence and caregiver burden by bringing screening into primary care or retail eye care. Retina specialists can then focus on those with the highest probable benefit. For AMD, multicondition programs bundled together spread platform costs, and screening is very cost-effective. Prior research shows that the economic cost of nAMD is large, averaging over $24,000 annually per patient in the US, with anti-VEGF treatment type and injection frequency driving costs most significantly (28). Also, studies show that AI in ophthalmology is generally cost-effective, showing that tele-ophthalmology can provide real economic value alongside clinical benefits (29). When used in tele-ophthalmology that minimises unnecessary visits and streamlines referrals, AI can save patients money and sight. The benefit grows as lower-cost anti-VEGF treatments become more common. Ultimately, AI-driven tele-ophthalmology offers a sustainable path forward, aligning earlier disease detection with smarter resource allocation, preserving vision for patients while reducing the long-term costs of AMD care.

HIGH INCOME VS. LOW INCOME COUNTRIES

In developed countries, cost-effectiveness in applying AI in ophthalmology to manage AMD is generally compared against existing arrangements of diagnosis, specialist assessment and availability of medication. In such cases, the cost of AMD is not so much the inability to diagnose disease as an issue of high recurrent expense of treatment and follow-up. AI is justified on labour efficiency and productivity of patients: triage technology using deep learning can automate clinic visits, optimise scheduling and enable harried ophthalmologists to focus on high-risk or actively converting AMD cases (29). AI application in such settings generally consists of integrating DL algorithms into OCT devices, electronic health records or tele-ophthalmology pipelines. The upfront expense of the initial software license, IT and clinician training is steep, but the payoff can be huge. AI allows for efficient triage, prevents unnecessary follow-up exams and optimises the time of the ophthalmologist on high-risk cases, reducing redundant man-hours and improving productivity in high-throughput clinics. Economic evidence in such environments shows that AI-assisted screening or monitoring has a tendency to provide cost-effective cost-effectiveness ratios, not so much because they supplant the cost of treatment, which remains inordinately high, but because they maximise the utilisation of expensive specialist manpower and imaging time (30). In short, in affluent systems, AI is cost-effective primarily as a productivity tool for labour and workflows, conserving care and preventing delays in treatment initiation at considerable expenses. Research shows that physicians’ AI monitoring or screening tend to be cost-effective in the developed world, but that is typically due to its optimisation of resources and manpower, and not because it can substitute for the exorbitant cost of the treatment (29).

In low- and middle-income countries, the mathematics are different because the weight of AI is different. There, the highest cost of AMD treatment is not in specialist time, but its unavailability and the lack of infrastructure: OCT machines are few, retina specialists are concentrated in major cities and patients have to travel long distances to receive treatment. For its installation in these areas, there are related initial expenses such as acquisition of imaging equipment, provision of stable electricity and internet connection, and subscription costs of AI software in foreign exchange (28). Tele-ophthalmology with AI can prolong scarce resources through the ability of technicians to perform screenings and mark high-risk patients, avoiding massive indirect expenses in travel, lost productivity and disability from blindness. Thus, where in wealthier countries the economic value of AI is from optimising current high-cost systems, in poorer countries its cost-savings rely on access to low-cost diagnostic tools, with its economic potential only being achievable when also combined with available care chains.

Overall for the patients, AI in AMD holds out potential for cost savings through containment of indirect costs, improved detection at earlier disease stages and prevention of expensive late-stage treatment – although patients will also have risk of false positives and higher out-of-pocket expenses in these cases. On a global scale, the cost-effectiveness of AI is divided by the level of the country’s economic development: in high-income nations like the US, it takes the burden on labour and optimises expensive processes (28), while in poor nations it may be revolutionary, providing detection and treatment that otherwise would not be available (29). Ultimately, the value of AI in AMD cannot be systematically evaluated into a cost-effective industry as the roles vary, along with the effects of AI on each country.

III. LIMITATIONS AND POTENTIAL RISKS

POLICIES AND LEGAL RISKS

Although there is no denying the positive impact of AI on ophthalmology, specifically in a clinical setting, it is also important to recognise the possibility of misdiagnoses and the legal implications that can impact the patient and clinician or provider. 

A common point of discussion is the unresolved malpractice and liability allocation; as a group of researchers in Cell Report stated, “it is still unclear whether healthcare providers, developers, sellers or regulators should be held accountable if an AI system makes mistakes” (33). The question of who will be at fault for misdiagnosis, and therefore the medical endangerment of the patient, is still being debated. In the EU, the legal risk class of ophthalmic AI (e.g. AMD triage/monitoring) is considered very dangerous, triggering documentation, human oversight and post-market duties, all coming with challenges and costs of their own. As researchers Aboy, Mansen and Vayena state in NPJ Digital Medicine, “under the EU AI Act, UK systems in regulated digital medical products are classified as ‘high risk’” (33). To put it simply, AI, even when monitored, is still considered dangerous and not yet developed enough to provide a reliable source for human health care.

The question to ask is whether or not AI assistance is worth the extra time, possible legal implications and danger in misdiagnosis. Will it really positively impact ophthalmology enough to outweigh the risks associated with misdiagnosis, including but not limited to negative impacts on patient care and legal consequences in the event of negative AI interventions? In the United States, the FDA clarifies that “the traditional paradigm of medical device regulation was not designed for adaptive AI/ML technologies” (34). The FDA recognises and acknowledges the core regulatory challenge that impacts liability, change control and post-market oversight for AMD tools. These tools were not made to work hand-in-hand with AI technologies which can impact performance and reliability. Similarly in the UK, it is the duty and responsibility of the European legal regimes such as the Medical Device Regulation (MDR) and the General Data Protection Regulation (GDPR) to prevent misuse of AI and dictate the reliability and security of AI’s implementation in medical devices. It is important that “the development of medical devices incorporating AI falls under the EU MDR and may also be subject to the GDPR” (35) to ensure a comprehensive review of the legal framework for AI’s emerging role in medical technology, research and life-changing devices. While prime examples of the potential of AI technology in medicine can be related to the US and EU (both of which contain highly developed legal systems), AI simply does not make sense for smaller, less developed countries (in the medical sense) as the required legal framework, and possible legal challenges, outweighs the potential for medical innovation. 

Although the ultimate responsibility for mandatory AI limitations falls with the government and medical governing body, it is also important to emphasise clinician accountability and clarity when handling AI in a clinical setting. As the Royal College of Ophthalmologists says, “accountability and explainability must be prioritised when deploying AI technologies in ophthalmology” (33). Clinician accountability and transparency is expected for AI use in eye clinics (e.g. the patient must be fully aware that their data and imaging is being reviewed by AI and must provide consent to be treated using AI). Not only is it the job of the clinician to ensure the patient’s diagnosis is complete and correct, it is also important to comprehend that the clinician is ultimately responsible for the patient’s health, care, safety and diagnosis. The clinician must be transparent with their intended use of AI and must provide a certain level of governance and oversight when resulting in the use of AI in diagnosis assistance. The clinician is also not only responsible for the physical care of the patient, but also for the care of the patient’s data privacy and protection of information. It is the clinician’s responsibility to “ensure data privacy and patients’ confidentiality… [with] the research protocol… reviewed and approved by the Institutional Research Ethics Committee” (38). Although reviewed by the Institutional Research Ethics Committee, it still falls upon the provider to ensure the privacy of the patient, continuing to follow patient-doctor confidentiality protocols. Region-wide ophthalmology guidance shows ethics-review guidelines and the privacy requirements. AMD imaging/OCT AI can revolutionise the diagnosis process resulting in quick and accurate results; however, without the use of clinical guidance, false positives or negatives could impact the wellbeing of the patient and lead to serious legal consequences, which could be counterproductive as the time saved in the clinic could be spent handling legal ramifications or implications. 

If a patient is misdiagnosed, due to AI’s role in treating age-related macular degeneration (AMD), who would be legally responsible? According to the EU MDR, it would be the clinician as it is the duty of the clinician to provide guidance for AI, and provide patient care. When all is said and done, the clinician has the responsibility to ensure the proper use of AI in the medical setting and must ensure that although AI can be referenced and used as a tool, ultimately, it is the final judgement and decision of the physician in regards to treatment plans, diagnosis and patient care. Therefore, the physician is still responsible for the life of the patient and in turn, all legal implications. 

MISDIAGNOSIS AND IMPACTS ON PATIENT CARE

An estimated 288 million individuals across the globe will be living with AMD by 2040 (39). Early diagnosis of age-related macular degeneration is crucial in order to retain vision by medically intervening as soon as possible and preventing the disease from progressing. AI-based algorithms specialised in detecting AMD, typically from colour fundus images and optical coherence tomography scans, have shown encouraging potential in revolutionising ophthalmology. However, there is always a risk of misinterpretation of certain features triggering AI’s response, leading to misdiagnosis, complexities in treatment planning and follow-up care.

Misdiagnosis in relation to retinal conditions (AMD, diabetic retinopathy, glaucoma etc.) involves mistaking some conditions for others, bringing about improper conclusions regarding treatment and opportunities for effective assistance. 

First, the performance of various AI-based algorithms is heavily affected by the quality and variety of included images in the process of algorithm development. Although fundus images are still used for training AI in detecting AMD, Treder et al. (40) have shown promising results in developing a DCNN model to detect exudative AMD on the basis of spectral domain optical coherence tomography (SD-OCT). Venhuizen et al. (41) have come up with an ML system capable of autonomously classifying OCT scans according to the AMD severity scale, achieving an AUC of 0.98 with SEN and SPE of 0.982 and 0.912, respectively. Furthermore, Yoo et al. (42) reported that multimodal imaging – a synthesis of colour fundus and OCT images – can enhance AMD diagnostic accuracy over fundus images alone, minimising the chance of misdiagnosis. Thus, with the utilisation of multimodal image analysis, the chances of being misdiagnosed could be significantly reduced. 

Second, AI’s diagnostic accuracy is dependent on data size for training and validation. Since algorithms are trained on limited datasets, the exposure to larger validation datasets ultimately results in lower diagnostic accuracy. It is important to recognise that the ability of AI to predict disease progression in real life remains considerably less accurate than clinicians would require for optimal application in healthcare. Bhuiyan et al. (43) designed a CNN algorithm in the age-related eye disease study (AREDS) database. The model was trained to automatically recognise the stage of a disease (early/none vs intermediate/late), obtaining 99.2% accuracy. Although achieving almost perfect accuracy, the prediction model for two-year incident of late AMD pulled off 86.36%, notably lower performance when a specific type of late AMD (66.88% for late dry AMD and 67.15% for late wet AMD) was to be detected independently. We also note that data diversity is crucial: while a lot of advanced studies are performed in More Economically Developed Countries (MEDCs), other world regions are under-represented. This may contribute to insufficient representation, since the data AI-based algorithms have been trained on will take neither demographic data nor personal characteristics of the patient into account. 

Third, most of the algorithms are developed to detect solely one ocular condition or symptom, resulting in the unsuccessful detection of other eye disorders. This engenders cases of both misdiagnosis and missed diagnosis.

When a patient is misdiagnosed, the impact on patient care is immediate. There are two categories of misclassification: false positives and false negatives. A false positive occurs when a condition is mistakenly reported to be present when it is not. Even though the outcomes of false positive cases are not as severe as those of false negatives, they are capable of causing potential harm. For instance, false positives in AMD testing may cause unnecessary treatment, resulting in additional clinic visits and medications. Such cases can result in additional expenses, increased anxiety and even possible complications from the interventions themselves. Moreover, patients might not be willing to participate in future screenings due to their negative experience, potentially impeding the detection of the actual condition they suffer from.

Conversely, a false negative occurs when an AI-based algorithm misses AMD activity following a diagnostic test. Early detection of AMD signs is particularly important and if an algorithm fails to do so, there may be serious consequences. For example, in wet AMD patients, missed AMD activity will delay anti-VEGF treatment, leading to irreversible vision loss. Relationships between doctors and patients may deteriorate as a result of such occurrences: if patients learn that AI was wrong, they may stop trusting their ophthalmologist and start questioning the trustworthiness of digital health.

CONCLUSION

In conclusion, advances in AI, particularly DL mechanisms, hold enormous promise in transforming AMD care by early detection, enhancing clinical efficiency and making cost-saving opportunities appear realistic. However, this promise is counterbalanced with the risks imposed by AI: unresolved liability, misdiagnosis, data privacy and so on. Particularly, misdiagnosis can delay treatment or expose patients to unnecessary interventions, while unresolved legal frameworks create uncertainty with accountability when errors occur. Also, in more economically developed countries, AI has great potential to be cost-effective due to advanced infrastructure and access to treatment; hence, the benefits outweigh potential risks. By contrast, in less economically developed countries, the implementation of AI as a screening tool will be challenged by the absence of effective and affordable treatment and lack of resources. What is crystal clear from this example is that AI is not universal everywhere; it is highly dependent on legal policies, clinician accountability and equitable access to treatment. Crucially, AI should function as a supportive tool rather than a replacement for ophthalmologists, with specialist oversight ensuring accuracy and clinical relevance. AI must be used as a tool for ophthalmologists to aid decision-making, without replacing the process of determining next steps in a patient’s treatment. Therefore, all the aspects regarding AI implementation in AMD screening must be taken into account in order to better navigate the future steps in ophthalmology.

1. Wong, W.L., Su, X., Li, X., Cheung, C.M., Klein, R., Cheng, C.Y. & Wong, T.Y. (2014). Global prevalence of age-related macular degeneration and disease burden projection for 2020 and 2040: A systematic review and meta-analysis. Lancet Glob. Health, 2, e106–e116. https://doi.org/10.1016/s2214-109x(13)70145-1
2. Lim, L. S., Mitchell, P., Seddon, J. M., Holz, F. G. & Wong, T. Y. (2012). Age-related macular degeneration. The Lancet, 379(9827), 1728–1738. https://doi.org/10.1016/S0140-6736(12)60282-7
3. Grassmann, F., Mengelkamp, J., Brandl, C., Harsch, S., Zimmermann & M.E., Linkohr, B. et al. (2018). A deep learning algorithm for prediction of age-related eye disease study severity scale for AMD from color fundus photography. American Academy of Ophthalmology, 125(9), 1410–1420. https://doi.org/10.1016/j.ophtha.2018.03.038
4. Ting, D.S.W., Pasquale, L.R., Peng, L., Campbell, J.P., Lee, A.Y. & Raman, R. et al. (2018). Artificial intelligence and deep learning in ophthalmology. British Journal of Ophthalmology, 103(2), 167–175. https://doi.org/10.1136/bjophthalmol-2018-313173
5. Macular Society (2025). What is the macula? Macular Disease Society. https://www.macularsociety.org/macular-disease/macula/
6. Cleveland Clinic (2023). Retinal Imaging. [Internet] Cleveland Clinic Health Library. https://my.clevelandclinic.org/health/diagnostics/25046-retinal-imaging/
7. Cleveland Clinic (2024). Fundus Photography. [Internet] Cleveland Clinic Health Library. https://my.clevelandclinic.org/health/diagnostics/fundus-photography
8. IBM (2025). What are convolutional neural networks? [Internet] IBM. https://www.ibm.com/think/topics/convolutional-neural-networks
9. Keen, M. (2021). What are Convolutional Neural Networks (CNNs)? [Internet Video] IBM Technology. https://www.youtube.com/watch?v=QzY57FaENXg
10. Bhuiyan, A., Wong, T.Y., Ting, D.S.W., Govindaiah, A., Souied, E.H. & Smith, R.T. (2020). Artificial Intelligence to Stratify Severity of Age-Related Macular Degeneration (AMD) and Predict Risk of Progression to Late AMD. Translational Vision Science & Technology, 9(25). https://doi.org/10.1167/tvst.9.2.25
11. Geeks for Geeks (2025). AUC ROC Curve in Machine Leaning. [Internet] GeeksforGeeks. https://www.geeksforgeeks.org/machine-learning/auc-roc-curve/
12. National Library of Medicine (2025). In brief: Sensitivity and specificity. National Center for Biotechnology Information. https://www.ncbi.nlm.nih.gov/books/NBK611900/
13. Trucco, E., Ruggeri, A., Karnowski, T., Giancardo, L., Chaum, E. & Hubschman, J.P. et al. (2013). Validating Retinal Fundus Image Analysis Algorithms: Issues and a Proposal. Invest Ophthalmol Vis Sci, 54(5):3546-3559. https://doi.org/10.1167/iovs.12-10347 
14. Chen, R., Chang, W., Song, F., Yu, H., Cao, D. & Zhen, Y. et al. (2024). Translating color fundus photography to indocyanine green angiography using deep-learning for age-related macular degeneration screening. NPJ Digit Med, 7:34. https://doi.org/10.1038/s41746-024-01018-7
15. Burlina, P.M., Joshi, N., Pekala, M., Pacheco, K.D., Freund, D.E. & Bresser, N.M. (2017). Automated Grading of Age-Related Macular Degeneration From Color Fundus Images Using Deep Convolutional Neural Networks. JAMA Ophthalmol, 135(11):1170-1176. https://doi.org/10.1001/jamaophthalmol.2017.3782
16. Wei, Q., Chi, L., Li, M., Qiu, Q., Liu, Q. (2025). Practical Applications of Artificial Intelligence Diagnostic Systems in Fundus Retinal Disease Screening. Int J Gen Med, 18:1173-1180. https://doi.org/10.2147/IJGM.S507100
17. Stahl, A. (2020). The diagnosis and treatment of age-related macular degeneration. Dtsch Arztebl Int, 117(29–30):513–520. https://doi.org/10.3238/arztebl.2020.0513
18. Lee, S., Song, S.J. & Yu, H.G. (2013). Current smoking is associated with a poor visual acuity improvement after intravitreal ranibizumab therapy in patients with exudative age-related macular degeneration. J Korean Med Sci, 28(5):769–774. https://doi.org/10.3346/jkms.2013.28.5.769
19. Van Leeuwen, E.M., Emri, E., Merle, B.M.J., Colijn, J.M., Kersten, E. & Cougnard-Grégoire, A. et al. (2018). A new perspective on lipid research in age-related macular degeneration. Prog Retin Eye Res, 67:56–86. https://doi.org/10.1016/j.preteyeres.2018.04.006
20. Van Leeuwen, R., Boekhoorn, S., Vingerling, J.R., Witteman, J.C.M., Klaver, C.C.W., Hofman, A. & De Jong, P.T.V.M. (2005). Dietary intake of antioxidants and risk of age-related macular degeneration. JAMA, 294(24):3101–3107. https://doi.org/10.1001/jama.294.24.3101 
21. Augsburger, M., Sarra, G.M. & Imesch, P. (2019). Treat and extend versus pro re nata regimens of ranibizumab and aflibercept in neovascular age-related macular degeneration: a comparative study. Graefes Arch Clin Exp Ophthalmol, 257(9):1889–1895. https://doi.org/10.1007/s00417-019-04404-0
22. Iida, T., Gomi, F., Yasukawa, T., Yamashiro, K., Honda, S., Maruko, I. & Kataoka, K. (2025). Japanese Retina and Vitreous Society Guidelines Committee for Neovascular Age-Related Macular Degeneration. Japanese clinical guidelines for neovascular age-related macular degeneration. Jpn J Ophthalmol, 69(4):639–660. https://doi.org/10.1007/s10384-025-01240-0
23. Hwang, D.K., Hsu, C.C., Chang, K.J., Chao, D., Sun, C.H., Jheng, Y.C., Yarmishyn, A.A., Wu, J.C., Tsai, C.Y., Wang, M.L., Peng, C.H., Chien, K.H., Kao, C.L., Lin, T.C., Woung, L.C., Chen, S.J. & Chiou, S.H. (2019). Artificial intelligence-based decision-making for age-related macular degeneration. Theranostics, 9(1):232–245. https://doi.org/10.7150/thno.28447
24. Chen, Q., Keenan, T.D.L. & Agron, E. et al. (2025). AI Workflow, External Validation, and Development in Eye Disease Diagnosis. JAMA Netw Open, 8(7):e2517204. https://doi.org/10.1001/jamanetworkopen.2025.17204
25. Li, H., Li, G., Li, N., Liu, C., Yuan, Z., Gao, Q., Hao, S., Fan, S. & Yang, J. (2023). Cost-effectiveness analysis of artificial intelligence-based diabetic retinopathy screening in rural China based on the Markov model. PLoS One. https://doi.org/10.1371/journal.pone.0291390 
26. Karabeg, M., Petrovski, G., Hertzberg, S.N., Erke, M.G., Fosmark, D.S., Russell, G., Moe, M.C., Volke, V., Raudonis, V., Verkauskiene, R., Sokolovska, J., Haugen, I.K. & Petrovski, B.E. (2024). A pilot cost-analysis study comparing AI-based EyeArt® and ophthalmologist assessment of diabetic retinopathy in minority women in Oslo, Norway. Int J Retina Vitreous. https://doi.org/10.1186/s40942-024-00547-3
27. Lin, S., Ma, Y., Xu, Y., Lu, L., He, J., Zhu, J., Peng, Y., Yu, T., Congdon, N. & Zou, H. (2023). Artificial Intelligence in Community-Based Diabetic Retinopathy Telemedicine Screening in Urban China: Cost-effectiveness and Cost-Utility Analyses With Real-world Data. JMIR Public Health Surveill. https://doi.org/10.2196/41624
28. Almony, A., Keyloun, K.R., Shah-Manek, B., Multani, J.K., McGuiness, C.B. & Chen, C.C. et al. (2021). Clinical and economic burden of neovascular age-related macular degeneration by disease status: a US claims-based analysis. Journal of Managed Care & Specialty Pharmacy, 27(9):1260–72. https://doi.org/10.18553/jmcp.2021.27.9.1260 
29. Wu, H., Jin, K., Chee, Y. C., Koh, V. & Ye, J. (2024). A systematic review of economic evaluation of artificial intelligence-based screening for eye diseases: From possibility to reality. Survey of Ophthalmology, 69(4):499–507. https://doi.org/10.1016/j.survophthal.2024.03.008 
30. Li, R., Yang, Z., Zhang, Y., Bai, W., Du, Y. & Sun, R. et al. (2022). Cost-effectiveness and cost-utility of traditional and telemedicine combined population-based age-related macular degeneration and diabetic retinopathy screening in rural and urban China. The Lancet Regional Health – Western Pacific, 23:100435. https://doi.org/10.1016/j.lanwpc.2022.100435 
31. Wu, X., Wu, Y., Tu, Z., Cao, Z., Xu, M. & Xiang, Y. et al. (2024). Cost-effectiveness and cost-utility of a digital technology-driven hierarchical healthcare screening pattern in China. Nature Communications, 15(1). https://www.nature.com/articles/s41467-024-47211-w
32. Davio, K. (2018). UK Health System Wins the Right to Treat AMD With Bevacizumab. [Internet] Center for Biosimilars. https://www.centerforbiosimilars.com/view/uk-health-system-wins-the-right-to-treat-amd-with-bevacizumab
33. Royal College of Ophthalmologists (2024). AI in Ophthalmology: Position Statement (London: RCOphth), pp. 1–3.
34. U.S. Food and Drug Administration (2025). Artificial Intelligence and Machine Learning (AI/ML) in Software as a Medical Device. U.S. Food and Drug Administration.
35. Aboy, M., Minssen, T. & Vayena, E. (2024) Navigating the EU AI Act: implications for regulated digital medical products. NPJ Digital Medicine, 7(1):237. https://doi.org/10.1038/s41746-024-01232-3 
36. Li, Z., Wang, L., Wu, X., Jiang, J., Qiang, W., Xie, H., Zhou, H., Wu, S., Shao, Y. & Chen, W. (2024). Artificial intelligence in ophthalmology: The path to the real-world clinic. Cell Reports Medicine, 15(4):107166. https://doi.org/10.5662/wjm.v15.i4.107166
37. Rodrigues, I.A., Hassan, S.U. & Khaw, P.T. (2023). Approval and certification of ophthalmic artificial intelligence devices in the European Union. Eye, 37:2177–2186. https://doi.org/10.1038/s41433-023-02575-8
38. Li, S., Wang, Y. & Xu, X. et al. (2023). APAO and APOIS joint position statement on the application of artificial intelligence in ophthalmology. Eye and Vision, 10, 57. https://doi.org/10.1186/s40662-023-00372-0 
39. Wong, S., Li, C. & Klein, C. et al. (2014). Global Prevalence of Age-related Macular Degeneration and Disease Burden Projection for 2020 and 2040: A Systematic review and Meta-analysis. [Internet] Lancet, 2:106–16. https://www.thelancet.com/action/showPdf?pii=S2214-109X%2813%2970145-1
40. Treder, M., Lauermann, J.L. & Eter, N. (2017). Automated detection of exudative age-related macular degeneration in spectral domain optical coherence tomography using deep learning. [Internet] Graefe S Archive for Clinical and Experimental Ophthalmology, 256(2):259–65. https://doi.org/10.1007/s00417-017-3850-3
41. Venhuizen, F.G., Van Ginneken, B., Van Asten, F., Van Grinsven, M.J.J.P., Fauser, S. & Hoyng, C.B. et al. (2017). Automated Staging of Age-Related Macular Degeneration using Optical Coherence Tomography. [Internet] Investigative Ophthalmology & Visual Science, 58(4):2318–28. https://doi.org/10.1167/iovs.16-20541
42. Yoo, T.K., Choi, J.Y., Seo, J.G., Ramasubramanian, B., Selvaperumal, S. & Kim, D.W. (2019). The possibility of the combination of OCT and fundus images for improving the diagnostic accuracy of deep learning for age-related macular degeneration: a preliminary experiment. Med Biol Eng Comput, 57(3):677-687. https://doi.org/10.1007/s11517-018-1915-z 
43. Bhuiyan, A., Wong, T.Y., Ting, D.S.W., Govindaiah, A., Souied, E.H. & Smith, R.T. (2020). Artificial Intelligence to Stratify Severity of Age-Related Macular Degeneration (AMD) and Predict Risk of Progression to Late AMD. Transl Vis Sci Technol, 9(2):25. https://doi.org/10.1167/tvst.9.2.25