By 2040, 288 million people worldwide are projected to have age-related macular degeneration (AMD).1 The increasing number of AMD cases calls for more frequent eye examinations. As a result, ophthalmologists will need more time to analyze patient data due to the heavier workload. However, there is promise in using artificial intelligence (AI)‒based analysis to predict the progression of AMD and evaluate its response to treatment. 

Artificial intelligence models can significantly assist in patient and treatment selection, drug development, and establishing critical endpoints for AMD trials.2 Machine learning algorithms diagnose AMD by analyzing big data from ophthalmic imaging modalities such as color fundus photography (CFP), fundus autofluorescence (FAF), near-infrared reflectance (NIR), and optical coherence tomography (OCT).3

What Is Machine Learning?

Diagnostic AI uses machine learning, which is a subset of AI that uses algorithms to analyze and learn from data, allowing it to make predictions and decisions without being explicitly programmed.3 It has numerous applications, including studying the retina.

Another subset of AI-based algorithms practical in AMD diagnosis is deep learning or deep neural networking. Deep neural networks have demonstrated accuracy comparable to clinical experts across various applications.4 Also, deep learning is adept at learning from images, which is ideal for AMD diagnosis, because AI tools are algorithms trained on large datasets of AMD-related images.3 By analyzing patterns and features in these images, the AI can identify the presence of AMD with high accuracy and can even determine whether it is wet or dry AMD.2

The use of advanced AI algorithms in health care helps professionals to identify the distinct characteristics of various types of AMD more quickly and accurately. This faster and more efficient diagnosis is crucial for timely treatment initiation, because early intervention can help prevent further progression of the disease and preserve vision. 

Several studies have incorporated these deep learning models to diagnose AMD. One such study, a recent Beijing Tongren Eye Center analysis, focused on AI algorithms’ ability to detect AMD in fundus images. The study revealed that these algorithms were nearly as practical as retinal specialists in detecting the disease.5 

Developing Effective Diagnostic Models With Artificial Intelligence

Factors to consider when developing diagnostic AI models are how these algorithms work, how researchers train the AI models, and what makes these models so effective. In the realm of medical diagnostics, it is crucial to involve researchers and human experts in the development of AI models. This collaborative approach, known as human-in-the-loop (HITL) machine learning, ensures that AI algorithms are not created in isolation.6 Instead, it fosters close collaboration between machine learning models and medical professionals, resulting in precise and reliable diagnostic tools.6 Human input is vital in identifying key image features to detect specific retinal disorders and interpret AI-generated results accurately.

Aside from HITL model setups, creating accurate AI algorithms for medical diagnostics also relies on large databases and high-quality data availability. In the case of retinal disorders, deep convolutional neural networks are trained using data sets, most commonly the National Eye Institute (NEI) Age-Related Eye Disease Study (AREDS) database, which contains more than 130,000 fundus photographs.7

It is important to acknowledge that the AREDS database was created in the 1990s and may need a more detailed understanding of hard drusen and age-related changes in the present clinical classification of AMD. Therefore, some outdated images may not be suitable for large-scale AI models. Additionally, it is important to note that some photos were digitized from film.8 Still, AREDS provides one of the most comprehensive datasets that features retinal images, including size and type of drusen and pigment abnormalities.9 The sheer robustness of AREDS data helps train retinal diagnostic models effectively.

The significance of data in AI medical diagnosis is further highlighted by the fact that deep learning models alone, without specific retinal imaging parameters, result in poorer predictive performance than the elaborate multiparametric approach used in creating these algorithms. According to a recent study, combining deep learning and machine learning alongside AMD-specific image parameter-generating algorithms enhances the sensitivity and specificity of automated color fundus photograph-based AMD prediction models.10

Do AI Diagnostics for Retinal Disorders Work?

The models discussed here explain how the development of AI software can facilitate rapid data processing and provide decision support for diagnosing, classifying, monitoring, and treating retinal diseases.11 It is crucial to consider the practicality of AI models based on their use cases. 

In terms of a referrable diagnosis, AI can help to identify patients who need immediate attention and referral to a specialist.11 Using advanced analytics algorithms, AI can analyze information from multiple sources to identify high-risk patients. This includes factors such as age, family history, smoking history, and the presence of other health conditions such as hypertension. By doing this, AI can assist health care professionals in making informed decisions about when to refer patients to specialists for further testing and treatment. 

A study by the National Yang-Ming University in Taiwan showcases the potential of using AI-powered diagnostic tools.12 Specifically, they introduce an intelligent cloud service for medical imaging diagnostics and telemedicine. The study aims to diagnose AMD via effortless uploading of OCT images on the website to determine if treatment is required.

The AI platform achieved exceptional image analysis accuracy that was on par with skilled retina specialists.12 Its detection accuracy consistently surpassed 90%.12 Additionally, the platform delivered treatment recommendations aligned with retina specialists.12

This study is not the only example of AI models effectively being used to diagnose retinal disorders. Another example is the effectiveness of a program known as iPredict (iHealthScreen).13 The program functions as an advanced diagnostic system using machine learning technology. It aims to accurately predict which individuals with early-stage disease are at higher risk of vision loss. 

When using the software, physicians capture retinal images via fundus cameras. These images are then securely sent to a centralized server. Once uploaded, a report is generated based on the images and patient data analysis. The report classifies the patient as either referable or nonreferable for AMD. The system assigns a prediction score for referable patients, quantifying their risk of developing late-stage disease within the next 2 years.13

In clinical trials, iPredict predicted a 2-year risk for progression to late AMD with 86% and 84% accuracy.13,14 The screening model has been validated and submitted to the FDA for approval to market the system to primary care practices by the end of 2023. This follows a 2018 FDA decision allowing the marketing of a medical device called IDx-DR that uses AI to detect diabetic retinopathy in severe adult cases.15 The IDx-DR works similarly to iPredict, though they diagnose different retinal disorders.

Still, these are not the only 2 AI diagnostic models for AMD that are being studied. The National Center for Biotechnology Information and researchers from the NEI have also successfully developed an advanced AI-based system.13 This system has undergone comprehensive training to identify reticular pseudodrusen. This pattern is strongly associated with an increased risk of progressing to the advanced stages of the disease.

The model uses data from 3,298 participants, comprising 80,000 images, from both AREDS and AREDS2. It showed superior predictive accuracy compared to 2 clinical standards used by retina specialists when tested against an independent data set.16 The study also found that combining deep learning grading and a survival approach enables accurate and automated predictions.16 

Limitations of Artificial Intelligence for AMD Diagnosis 

Despite the successful results of AI models, there are limitations and unanswered questions surrounding the predictive capabilities of AI models for AMD progression. Artificial intelligence algorithms are typically developed to see only 1 disease or sign, they are trained on limited data sets, and their effectiveness relies on the quality of the image for accurate diagnosis. As a result, they still need comprehensive and accurate data sets to diagnose or predict AMD progression effectively. Also, AI models cannot account for the complexity of the relationship of AMD to lifestyle factors such as diet, smoking, or environmental exposures. 

Implementing AI software poses additional challenges, such as feasibility and performance compared to clinical physicians, patient trust in machines, and the potential issues of a “black box” system, where physicians might miss false-negative cases, and patients must understand that referrals are necessary if AMD symptoms develop. 

Furthermore, machine learning and deep learning models have been criticized for their lack of generalization and robustness, often overfitting to specific datasets. To address this “black box” problem, Explainable Artificial Intelligence (XAI) techniques have been developed.17 These techniques aim to enhance the interpretability and trustworthiness of AI models, ultimately optimizing their performance in real-world applications.17


Despite their limitations, AI models hold great potential as a screening tool for AMD. As the AI algorithms continue to be refined, they will likely be able to detect subtle changes in the eye that could go unnoticed by clinicians. This could aid in early diagnosis and treatment decisions. Also, AI diagnostics allow the automation of AMD-specific tasks, such as grading the severity of AMD based on retinal images or monitoring disease progression over time. Automating these tasks can significantly reduce the workload associated with routine retinal evaluations for physicians, which can be a significant benefit as the global population ages and the number of AMD cases skyrockets.

Editor’s note: This article is discussed on the Retina Podcast at


  1. Wang Z, Lim G, Ng WY, et al. Synthetic artificial intelligence using generative adversarial network for retinal imaging in detection of age-related macular degeneration. Front Med (Lausanne). 2023;10:1184892. doi:10.3389/fmed.2023.1184892
  2. Muntean GA, Marginean A, Groza A, et al. The predictive capabilities of artificial intelligence-based OCT analysis for age-related macular degeneration progression-a systematic review. Diagnostics (Basel). 2023;13(14):2464. doi:10.3390/diagnostics13142464
  3. Mehta N, Modi Y, Du W, et al. Predicting outcomes and treatment frequency following monthly intravitreal aflibercept for macular edema secondary to central retinal vein occlusion: a machine learning model approach. Invest Ophthalmol Vis Sci. 2022;63:2673.
  4. Zhou Y, Chia MA, Wagner SK, et al. A foundation model for generalizable disease detection from retinal images. Nature. 2023;622:156-163. doi:10.1038/s41586-023-06555-x
  5. Dong L, Yang Q, Zhang RH, Wei WB. Artificial intelligence for the detection of age-related macular degeneration in color fundus photographs: A systematic review and meta-analysis. EClinicalMedicine. 2021;35:100875. doi:10.1016/j.eclinm.2021.100875
  6. Suhir E. Human-in-the-Loop (HITL): Probabilistic predictive modeling (PPM) of an aerospace mission/situation outcome. Aerospace. 2014;1(3):101-136.
  7. Age-Related Eye Disease Study Research Group. The Age-Related Eye Disease Study (AREDS): design implications. AREDS report no. 1. Control Clin Trials. 1999;20(6):573-600. doi:10.1016/s0197-2456(99)00031-8
  8. AREDS2 Research Group, Chew EY, Clemons T, et al. The Age-Related Eye Disease Study 2 (AREDS2): study design and baseline characteristics (AREDS2 report number 1). Ophthalmology. 2012;119(11):2282-2289. doi:10.1016/j.ophtha.2012.05.027
  9. National Eye Institute. Age-Related Eye Disease Studies (AREDS/AREDS2). Accessed November 14, 2023.
  10. Bhuiyan A, Wong TY, Ting DSW, Govindaiah A, Souied EH, Smith RT. Artificial intelligence to stratify severity of age-related macular degeneration (AMD) and predict risk of progression to late AMD. Transl Vis Sci Technol. 2020;9(2):25. doi:10.1167/tvst.9.2.25
  11. Daich Varela M, Sen S, De Guimaraes TAC, et al. Artificial intelligence in retinal disease: clinical application, challenges, and future directions. Graefes Arch Clin Exp Ophthalmol. 2023;261(11):3283-3297. doi:10.1007/s00417-023-06052-x
  12. Hwang DK, Hsu CC, Chang KJ, et al. Artificial intelligence-based decision-making for age-related macular degeneration. Theranostics. 2019;9(1):232-245. doi:10.7150/thno.28447
  13. AI-based systems can help identify rapidly advancing age-related macular degeneration. March 14, 2023. Accessed November 14, 2023.
  14. Telemedicine in age-related macular degeneration. identifier: NCT04863391. Last updated April 28, 2021. Accessed November 14, 2023.
  15. US Food and Drug Administration. FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems. April 11, 2018. Accessed November 14, 2023.
  16. Peng Y, Keenan TD, Chen Q, et al. Predicting risk of late age-related macular degeneration using deep learning. NPJ Digit Med. 2020;3:111. doi:10.1038/s41746-020-00317-z
  17. Wang MH, Chong KK-l, Lin Z, Yu X, Pan Y. An explainable artificial intelligence-based robustness optimization approach for age-related macular degeneration detection based on medical IOT systems. Electronics. 2023;12(12):2697.