Artificial intelligence (AI) has the potential to benefit both patients and providers across a multitude of health care specialties. For glaucoma specifically, however, roadblocks have slowed progress in application of AI, because the nature of its diagnosis can be subjective, and no clear standards exist for imaging. Despite the challenges, researchers have been investigating ways to implement AI in ophthalmology and glaucoma. Here, Glaucoma Physician speaks with 3 clinician-scientists who have been leading the charge in AI research. They share insights and recommendations for ophthalmologists who want to learn more.

Pearse Keane, MD

I’m a consultant ophthalmologist at Moorfields Eye Hospital in London, specializing in the treatment of retinal disease. I’m also a professor of artificial medical intelligence at University College London Institute of Ophthalmology. Although I’m not a computer scientist or engineer, I’ve come to find myself leading a multidisciplinary clinical research group that aims to develop and apply AI in health care. We use ophthalmology as an exemplar in this regard, but we are not limited to work in this specialty.

I first became interested in AI when I was reading about all the latest advances in the field over the last decade, in particular the way in which a form of AI called “deep learning” had begun to revolutionize the tech industry. We started to see this technology become widely adopted in areas such as computer vision, speech recognition, foreign language translation, and even in self-driving cars. These advances were driven by advances in computing power, as well as the availability of large data sets. Given the challenges, it seemed that, if implemented well, AI would have the potential to transform health care. Even better, it seemed that ophthalmology could be at the forefront of this revolution.

It seems to me that the best application of AI in glaucoma would be to help highlight which patients are being undertreated or overtreated. This could help make clinics more efficient, which is a necessity given how overwhelmed with patients most glaucoma clinics are. There may also be a role for AI in early detection of glaucomatous change, hopefully at a point where some of it is reversible.

I think the real promise of AI is how we can bring world-class expertise out of specialized centers into the community and even into the home. Even when we can successfully prevent blindness, chronic eye diseases such as glaucoma place a huge burden on patients, often requiring very frequent hospital visits. With the advent of AI, plus supporting telemedicine infrastructure, I think we’ll increasingly be able to monitor stable patients in the community.

With my friend and collaborator, Joe Ledsam, MD, I coined the term “artificial medical intelligence” to describe the next phase of AI in health care. From my experience in the last 5 years, I’ve come to learn that training a deep-learning model is only one component of a much larger translational pathway for AI-enabled health care. This pathway requires identification of novel clinical applications (“use cases”), aggregation and labelling of clinical data, training of models, robust clinical validation, regulatory approval, and health services delivery and implementation science. I think clinical researchers will play an increasingly leading role in these developments. As such, I think we’ll start to see more short courses and postgraduate certificates related to this vision. These courses will be far more than just teaching how to code a neural network but will encompass this much wider spectrum of activity. Separately, as AI systems become much more widely available in the coming decade, we will start to see their inclusion in medical school and residency curricula.

For practicing ophthalmologists who wish to learn more about AI, I would advise them to read 2 books: Deep Medicine by Eric Topol¹ and Artificial Intelligence by Melanie Mitchell.² These are amazing primers for the field. I would also encourage them to explore code-free deep learning (CFDL). This is an area we have published several papers about, led by a great member of our team, Edward Korot, MD. I encourage reading his paper in Nature Machine Intelligence describing how to use different code-free platforms for medical imaging.³ The exciting thing about CFDL is that it allows clinicians with minimal if any coding expertise to begin to train their own deep-learning models. Of course, such models can’t be used for direct care of patients, but I think they will allow clinical academics to explore proof-of-concept for the thousands of novel applications for AI that they will imagine in the coming years.

In 10 years, ophthalmologists will still be as busy as ever. However, our practice will have been transformed by AI. At each patient visit, we will likely have huge amounts of data to digest, from multiple different advanced imaging systems; from genomic, proteomic, and multiple other “omic” profiles; from multiple psychophysical tests, including virtual reality and augmented reality; from wearable technologies such as smart watches; from hyperlocal environment information; and more. We will then use a range of narrow AI tools to obtain actionable intelligence from each of these systems. As clinicians, our job will be to integrate these outputs to decide the best course of action for our patients.

Nicolas Jaccard, PhD

I am the principal architect at Orbis International, overseeing our portfolio of digital products, including our AI-enabled platform for eye care, Cybersight AI. My team is focusing on the development and validation of AI and machine learning algorithms for the screening of sight-threatening disease, such as diabetic retinopathy and glaucoma. We are also exploring the use of AI and machine learning for training and teaching, an area that we call “machine mentoring.”

The aspect of AI and machine learning that really got me interested in the field was the ability to derive novel and useful insights from data that has been around for decades. When looking at applications to eye care in particular, it quickly became apparent that information was “hiding in plain sight” in ophthalmic imagery. As an example, a research paper demonstrated that age and gender could be inferred from retinal photographs,⁴ something that was unknown to ophthalmologists at the time. This is where the technology becomes exciting for me: not to replicate and automate human processes, but to do things that are either currently unknown or thought impossible.

Glaucoma is complex to tackle from an AI standpoint due to various factors, the main one being the complexity of a diagnosis. Glaucoma diagnosis relies on a variety of modalities and fragmented clinical standards. This is in part because glaucoma is a group of conditions, and thus symptoms and biomarkers will vary greatly from one case to another. As a result, it is challenging to build data sets because there is significant variability between experts (lack of agreement regarding diagnosis and prognosis). This makes building robust AI solutions difficult, unlike in diabetic retinopathy where well-documented standards have been established for detection and progression assessment is based solely on retinal images.

We need better data sets, potentially including all modalities used in diagnosis, and better ways to accommodate the noisy nature of glaucoma grading. I do hope that we will get to a point where glaucoma can be reliably diagnosed and graded based on an optic disc photograph, which would enable mass screening of the condition, especially in low-income to middle-income countries where other modalities are not readily available.

Artificial intelligence is well suited for clinical decision support, ensuring that as much information is available to make the best possible decision for patients. Using AI to upskill a large contingent of health-care professionals and enable them to actively participate in clinical decision-making will likely have a significant impact on clinical practice and patients, especially in low-resource environments. For example, imagine an ophthalmic photographer being able to confidently refer patients for a wide variety of sight-threatening conditions.

The biggest challenges in AI are safety and ethical concerns associated with its use in health care. It is important to provide a robust basis for clinicians and health-care professionals to understand how the technology works and its limitations. Platforms like Cybersight will be key to making sure that appropriate learning material is available to as many trainees as possible, entirely free of charge. There is work needed to ensure that reference material is created to base courses on. There are a few efforts ongoing already.

To learn more about AI, I recommend hands-on experience if possible. There are machine-learning courses on platforms such as Coursera that take students from no programming knowledge to a level sufficient to develop useful AI algorithms.

Ten years ago, when deep learning was starting to explode, we all thought that AI would be everywhere by now. While this is true in consumer goods, it hasn’t had the expected impact in health care. It is a difficult problem to solve, and until everything aligns (availability of data, regulations, advances in technology, trust by patients and clinicians) we are unlikely to see major breakthroughs. Regulations will likely catch up, providing a clear pathway for AI-enabled systems to be integrated in clinical practice. Meanwhile, the field as whole will hopefully become better at dealing with critical issues such as bias.

I would expect AI-enabled systems to be commonplace in 10 years, but the rush that we currently see for autonomous AI will have subsided. In everyday life, AI will be an assistive tool. However, in eye care, AI will only live up to its promises if the entire community works together. The current approaches are very fragmented. Academia, the tech industry, health-care providers, and nonprofits need to work together to deliver a model that works for patients everywhere.

Lama Al Aswad, MD, MPH

I am a professor of ophthalmology; a professor of population health; the vice chair for innovations; the director of teleophthalmology, artificial intelligence, and innovations; and the director of the ophthalmology innovation fellowship at New York University (NYU) Medical School, Langone Health. I am currently working on several projects related to AI, including work on glaucoma diagnosis and predication of glaucoma progression, the use of federated AI for glaucoma and diabetic retinopathy diagnosis, and building the ophthalmology data hub at NYU for research and AI training.

What attracted me to AI research was the idea of improving population health by screening high-risk populations while improving the accuracy of diagnosis in a cost-efficient fashion. More than 50% of individuals who have glaucoma don’t know they have it, and it is more prevalent in high-risk and underserved communities. I started screening for glaucoma in New York City in 2004 and then launched the teleophthalmology mobile unit to screen for the 4 leading causes of blindness. The costliest aspect of this was manpower. If we can deploy AI to help with screening and then send only the suspects and individuals with disease to eye care specialists, then we can improve access to care efficiently, decrease the burden of blindness effectively, and help control health-care cost.

One major challenge to implementing AI in glaucoma is the lack of a unified gold standard and a consensus for glaucoma diagnosis. The biggest unmet needs in AI for glaucoma are large, diverse data sets to train AI and standardization in AI reporting and algorithmic fairness. In addition, many clinical validation studies demonstrate moderate to poor levels of transparency.

In glaucoma, AI can provide value in all aspects of clinical practice, screening, management, and treatment of glaucoma and remote monitoring. I believe AI will play an important role in better understanding glaucoma and its progression. Furthermore, it will improve its management and access to care.

The future is the merging of innovations, not only AI, in medical and ophthalmic education and care. It is important to start thinking outside the box by integrating innovations in ophthalmology. For example, I have launched and endowed the ophthalmology innovations fellowship at NYU. The curriculum has 3 arms: teleophthalmology and technology development, big data and AI, and technology transfer (taking an idea from concept to market). Our fellow this year is Dinah Chen, MD. She graduated from the ophthalmology residency at NYU last year and joined the ophthalmology innovations lab. We also need to create a curriculum for AI for ophthalmology residents. In 2020, I launched an AI curriculum for our residents at NYU. I would also encourage practicing ophthalmologists who are interested in learning more to access some of the many resources online that can be utilized to learn about AI in medicine and ophthalmology. In addition, I encourage students and residents to consider a master’s program in bioinformatics and data science.

Artificial intelligence is already in ophthalmic clinical practice, for example diabetic retinopathy screening. It is only a matter of time for AI to be integrated in all aspects of our daily patient workflow. Artificial intelligence technologies and innovations are driving the future of health care. Ophthalmologists should learn, embrace, and adopt to stay in the lead. GP

Pearse Keane, MD, is a consultant ophthalmologist at Moorfields Eye Hospital in London.

Nicolas Jaccard, PhD, is the principal architect at Orbis International.

Lama Al Aswad, MD, MPH, is a professor of ophthalmology at New York University Medical School, Langone Health.

Editor’s note: We would like to thank Malik Kahook, MD, for his guidance and coordination of this feature.

References

1. Topol E. Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books. 2019.
2. Mitchell M. Artificial Intelligence: A Guide for Thinking Humans. Macmillan. 2019.
3. Korot E, Guan Z, Ferraz D, et al. Code-free deep learning for multi-modality medical image classification. Nat Machine Intel. 2021;3:288-298.
4. Poplin R, Varadarajan AV, Blumer K, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng. 2018;2(3):158-164. doi:10.1038/s41551-018-0195-0