Artificial Intelligences Creates Inroads to Ophthalmology Clinical Practice

SAN DIEGO — As some experts sound the alarm about the potential dangers of artificial intelligence (AI), ophthalmologists have begun embracing AI as a means to improve the quality and efficiency of patient care.

During a first-of-its-kind session at the American Society of Cataract and Refractive Surgery meeting, speakers discussed ongoing efforts to develop and apply AI to clinical decision-making related to diagnosis, treatment, follow-up, and prognosis of multiple ophthalmologic conditions.

“Whether you’re a seasoned professional or a curious newcomer, we invite you to join us on our journey to revolutionize eye care,” said program co-chair Eric Rosenberg, DO, of SightMD in Long Island, New York, at the opening of the inaugural Digital Ophthalmic Society (DOS).

Co-chair John W. Kitchens, MD, of Retina Associates of Kentucky in Lexington, said the DOS presence in the ASCRS program affirms a “call to action … that crosses all specialties, professions, boardrooms, and even political borders.”

The utility of AI begins with data points, said Karl Stonecipher, MD, of Laser Defined Vision in Greensboro, North Carolina. The more data loaded into a computer or network, the more AI can help with patient care.

Dry Eye Diagnosis

As an example of how to use AI in clinical practice, Stonecipher described the application of AI to the diagnosis of dry eye, using a software platform known as CSI Dry Eye. The program has general information about the condition, and the ophthalmologist adds patient-specific data gleaned from a 50-item questionnaire that is automatically incorporated into the software.

Additional data come from patients’ subjective scoring on two dry eye risk factor surveys. Photos representing different levels of severity are added to the mix. The platform offers the flexibility to allow a practitioner or group to add other data elements, including personal preferences, that are meaningful to their clinical practices and to specific patients.

“I don’t have to do everything, and I don’t have to have every machine to do everything,” said Stonecipher. “I just want you to be able to do everything that you want to do. Put in as much [data] as you can, because that makes the machine more robust.”

The data form the basis for ongoing studies to develop computer models of dry-eye severity and dry-eye type. In the current phase of work, the models inform decision making related to dry eye diagnosis, but the long-term goal is to develop aids for making diagnostic and treatment decisions.

Almost 500 doctors are contributing data and using the computer platform. Data from 25,000 assessments and 22,000 questionnaires have been input. A recent analysis showed that the severity model had an area under the receiver operating characteristic curve (AUC-ROC) of 0.79 and AUC-precision recall (AUC-PR) of 0.61 for predicting dry eye severity. The type model had an AUC-ROC of 0.91 and AUC-PR of 0.94.

“So how is this helping me?” said Stonecipher. “If I am better able to diagnose a problem then I’m more likely to get to the final point of success. We’re now inputting everything that we can possibly know… . What I think this software will help you with is making the correct diagnosis but also, ultimately, the correct treatment.”

Outcomes with Intraocular Lenses

Another ongoing developmental program applies AI and machine learning to outcomes in cataract and refractive-lens surgery. The goal is to provide personalized medicine based on a large pool of data with objective inputs and conclusions, said Mark Packer, MD, co-founder of Oculotix, a start-up that has the goal of developing AI to achieve the personalized medicine goal.

“I’ve been doing multifocal and premium lenses for 25 years, and it’s still not perfect,” said Packer. “I cannot guarantee a happy patient, but I feel pretty good about my results after all that time. But I’m still human. What happened yesterday is going to impact how I feel today. If I have a patient who is very unhappy and wants a lens explanted tomorrow, I think I’m probably going to be more cautious about recommending that lens to the next person who comes down.”

“But that’s a fallacy that has no real validity because it’s just my experience from yesterday, not pulling in my experience over the 10,000 multifocal lenses I have implanted over the last 25 years. Wouldn’t it be nice if I had a tool that would give me a fresh start with each patient, based on all the data, not only that I have generated, but that all of you have generated… . That’s what we’d like, to get to his [patient’s] personalized medicine based on a large pool of data with a more objective set of inputs and conclusions rather than just based on what happened to me this week.”

The concept behind Oculotix is to have a “helper” take the optics of different intraocular lenses (IOLs) and use machine-learning, based on outcomes of previous patients, and suggest an optimal outcome for a specific patient, Packer continued. The platform also incorporates patient-reported postoperative outcomes that are reported through a cellphone app.

“We have a more objective way to integrate patient-reported outcomes into the feedback loop to help us achieve higher levels of patient satisfaction,” he said.

Natural Language Processing

Ophthalmologists recognize the abbreviation NLP as “no light perception,” but in the world of AI, NLP refers to natural language processing, said Gurpal Virdi, MD, an intern at the University of Missouri in Columbia, and founder and CEO of EyeLabs AI, a company that develops AI solutions for eye care.

“Natural language processing is a subfield of AI and linguistics that helps computers to understand, interpret, and generate human language that is both meaningful and useful,” said Virdi. “This can be done in the form of spoken word or written.”

Typically, NLP is created by means of transformer neural network architecture, a deep-learning model that can generate and understand human language. The advanced technology allows input, analysis, and interpretation of larger amounts of text, as compared with earlier, slower forms of architecture that processed text on a word-by-word basis. Transformer architecture can be trained on large-text data and fine-tuned for specific NLP tasks, said Virdi.

A variety of NLP tasks have evolved. One common task is named entity recognition, which refers to identification and classification of names, dates, medications, surgical procedures, and words or terms in a body of text, such as clinic notes, Virdi continued. Sentiment analysis can recognize the emotional tone within text, recognizing whether the tone is positive, negative, or neutral.

Speech recognition has been around for years. With respect to NLP, speech recognition included not only translating spoken words into text but voice-based interaction with digital systems. NLP also can perform text summarization of long documents and large amounts of text, such as surgical notes or research papers.

Most recently, NLP has evolved to include chatbots and conversational agents that can be trained to help with patients with preoperative and postoperative questions and to assist with scheduling.

Within the field of ophthalmology, NLP transformers can be trained to identify valuable information from free-text narratives, such as clinic and operative notes and abbreviations, Virdi continued. NLP can be trained to extract meaningful insights from unstructured data, for example, key terms and language patterns. Common tasks, such as prior authorization requests, can be automated with the aid of NLP. The technology also can be trained to identify social determinants of health, such as “can’t afford medication” in clinical notes.

Still other tasks within the realm of NLP include surgical planning, summarization of notes to simplify and contextualize terminology to improve understanding and prevent errors, note review for patients assessed for clinical trials, and development of ophthalmology-specific digital scribes to improve documentation, reduce physician burnout, and increase patient satisfaction.

Future directions in NLP include a great opportunity to develop ophthalmology-specific large language models (LLMs) to replace current open-source LLMs that are expensive and not HIPAA compliant, said Virdi. Transforming vast amounts of electronic health record data to LLMs has the potential to improve patient care and enhance research.

Other speakers during the session described the use of AI to perform power calculations for IOLs and choose the best IOL formula for any eye, develop a framework for surgical guidance, enhance ophthalmic imaging, perform retinopathy screening, and predict conversion to neovascular age-related macular degeneration.

Please enable JavaScript to view the comments powered by Disqus.