Artificial intelligence use is growing across medicine. From imaging to chatbots, ophthalmology is the center of one of these transformations.
The phrase "artificial intelligence" is everywhere in the public consciousness — both a buzzword promising a brighter tomorrow and a curse looming over our collective heads.
The public debate concerning the ethics and validity of artificial intelligence (AI) is likely to rage on for decades, but the evolving role of AI in ophthalmology suggests its vast potential to enhance clinical care and patient outcomes. AI imaging technologies may better track ophthalmic disease development, while AI chatbots could inform the next generation of medical leaders.
However, regardless of this excitement, ophthalmologists will face significant challenges. Given the rapidly expanding scope of artificial intelligence, the specialty will need to overcome ethical concerns and issues with interpretation to safeguard patient outcomes.
“We already know, just like any other technology, the first and second generations are not going to be widely used, and they’re going to have to work out their kinks,” said Jonathan Jonisch, MD, partner, Vitreoretinal Consultants of New York. “I think we’re a way away from using machine learning to guide our treatments, but I think the value of artificial intelligence imaging is here already and will just continue taking steps to move forward.”
The role of AI may take the form of an added tool in the growing armamentarium of clinicians, beginning with imaging technologies. A variety of deep and machine learning models have been deployed across the specialty, being worked and reworked to improve imaging and better detect ophthalmic disease and disease progression.
“When you have AI algorithms that have been trained to look at imaging, and perhaps using biomarkers that we may not see with the naked eye, there’s a potential for AI to allow for decision support that may be even better than what can be done by humans,” said Rajeev Muni, MD, MSc, a vitreoretinal surgeon in the department of ophthalmology at St. Michael’s Hospital and Unity Health Toronto.1
Data from these studies have indicated the benefit of AI in tracking disease developments. An analysis of real-world data in China found an AI-based fundus screening system had the ability to detect 5 prevalent ocular conditions, with a particularly favorable efficacy for diabetic retinopathy, retinal vein occlusion (RVO), and pathological myopia.
Investigators described the potential benefits of the clinical application of AI including its ease of use and limited need for resources, particularly for fundus screening, and ability to collect epidemiological data.2
Multiple studies presented at the 83rd Scientific Sessions of the American Diabetes Association (ADA) focused on the translation of AI systems to detect diabetic eye diseases. This included the real-world deployment of an autonomous AI system at Johns Hopkins School of Medicine being linked to improved testing adherence for diabetic eye diseases across primary care clinics.
In particular, the deployment improved access and equity for those traditionally disadvantaged in medical care. Investigators suggested the use of AI to overcome historic disparities will not only benefit ophthalmology, but medicine as a whole.3
Another analysis found machine learning models allowed for the accurate and feasible identification of the progression of diabetic retinopathy. The AI model predicted approximately 91% of the ultra-widefield images with the correct labels, often indicating greater disease progression than human graders. These algorithms, as a result, may further refine patient risk and introduce personalized screening intervals.4
According to Jonisch, machine learning may allow for the analysis of nuanced features on images and better prediction of the disease progression. With this knowledge, ophthalmologists could determine the most beneficial therapy for patients, as well as better determine the risk of failure or chance of success in a relevant clinical trial.
“Artificial intelligence and machine learning do a really good job of taking many more data points than we could analyze as a human at one time,” Jonisch said. “I would envision a time where machine learning can help us predict disease progression better.”
Racial bias in imaging, however, could be a residual concern for ophthalmologists. Race, although sociologically a social construct, is a phenotypic feature that can affect image-based classification performance.
An AI system has the capability to be deployed at a greater scale than an individual clinician. Thus, the potential harm from these biases may be increased, particularly when introduced in demographics different from those on which the system trained.
A diagnostic study conducted at Oregon Health & Science University found AI imaging could infer self-reported race from retinal fundus images and vessel maps previously believed to not contain information relevant to race – something human graders cannot do.5
As the use of AI grows in medicine, clinicians and researchers may need to place focus on strategies to mitigate AI biases, from the data collection stage to the evaluation and post-authorization deployment stage.
Recent analyses indicate the role of artificial intelligence chatbots could soon be extended from a fun novelty to a test preparation tool in ophthalmology.
New data suggest the increasing benefit of the popular AI chatbot, ChatGPT, for preparation for ophthalmology board certification. The investigative team from the University of Toronto found ChatGPT 4.0 answered 84% of multiple-choice practice questions taken from OphthoQuestions, a common practice resource for board examination, correctly in July 2023.6
Based on previous findings from the investigative team, the chatbot only answered 46% of multiple-choice questions correctly in January and improved slightly to 58% in February.7
“We can see almost in real time how this AI chatbot has evolved in terms of its ophthalmic knowledge and the gains in the performance of the chatbot, we’re seeing in virtually every subspecialty area of ophthalmology, from cornea to glaucoma and retina,” said Marko M. Popovic, MD, MPH, a resident physician in the department of ophthalmology and vision sciences at the University of Toronto, and one of the study investigators.1
While there remains work to be done, Popovic suggested the dramatic advances in the capability of ChatGPT in preparing for board certification in a short period of time lend credence to its future potential.
Another analysis suggested large language models provide appropriate ophthalmic advice to patient questions. Investigators from Stanford University noted, in particular, that the generated AI answers did not significantly differ from an ophthalmologist regarding incorrect information or the likelihood of harm.8
However, there are notable limitations to a chatbot’s benefit, stemming from its capability for "hallucinations," when a large language model responds with incorrect information or facts that aren't based on real data.9
In a study from the New England Eye Center, a large language model-based platform provided largely inaccurate information on questions regarding vitreoretinal diseases, including age-related macular degeneration (AMD), diabetic retinopathy, and retinal vein occlusion, with inconsistencies on repeat inquiries. Exactly half (50.0%) of answers were materially different, even after no functional changes were made to the platform between the first and second question submissions.10
Popovic suggested that inaccurate or incomplete responses could lead to suboptimal care and issues down the line, particularly if young physicians rely on answers from ChatGPT in the beginning stages of their careers.
“I think the bottom line here is at the end of the day, in the diagnosis and treatment of patients, the AI chatbot cannot be held accountable for what it provides,” he said.1 “And that’s particularly challenging in the situation where you ask ChatGPT what the symptoms of X condition are, and it provides 10 symptoms and only 9 of them are correct.”
As disease patterns and treatment responses may be recognized much faster with the use of AI tools, these tools could mark a forward leap for the field. Jay Duker, MD, the president, and chief executive officer of Eyepoint Pharmaceuticals, believes relying on AI for certain abilities may not be the worst thing for the specialty.
“I’ve been saying for a while to young residents that in 10 years, we’re not going to be diagnosticians,” Duker said.11 “The optical coherence tomography is going to tell you what the patient has, and everyone says ‘Oh, that won’t be fun anymore.’ It will, because now we’re going to concentrate on the patient, instead of concentrating on what they have, and we’re going to connect with them at a more personal level.”
Still, these specialists indicate more data is required for validation and the specialty should be cautious when implementing artificial intelligence into full-time clinical care. There is also a creeping, and understandable, fear of a future where AI replaces human intuition with ones and zeroes.
But it may be important to remember what these machines can and cannot do. In conjunction with a specialist’s expertise, an AI system could improve patient outcomes without sacrificing the Hippocratic oath – to do no harm.
“I think when these technologies are initially rolled out, we don’t need to fully trust it,” Jonisch said. “It can be used in addition to our current therapy, not instead of. That is how a lot of areas of medicine incorporate newer technologies, you don’t initially fully rely on them. You do it in conjunction with what you’re already doing.”