Artificial intelligence is growing more sophisticated by the day, and researchers in dry eye disease and glaucoma are just starting to harness its power.
In the half-century plus since the advent of the first powerful computers, few technologies have garnered such breathless excitement as artificial intelligence (AI). Even before the advent of the transistor and modern silicon wafer semiconductors, human beings have had a remarkably prescient sense that one day our machines would become intelligent.
After years of promise, AI’s nascent potential is beginning to bloom. And healthcare is proving to be fertile ground for the seeds of AI to take root. Though the most cynical among us may note that the most high-powered AIs are reserved for the military and the markets, medicine is not far behind.
In eye care, great strides have already been made with AI algorithms. Optometrists, as primary eye care providers whose duties encompass screening, monitoring, and management of disorders of the eye, stand to gain greatly. Even today, algorithms for image processing for diseases like diabetic retinopathy, macular edema, and age-related macular degeneration are changing the field forever.
As two of the most critical ailments affecting the eye, glaucoma and dry eye disease (DED) would seem to follow. Depending on diagnostic criteria, the prevalence of dry eye in the population has been listed at between 5% and an eye-watering 50%. Glaucoma is also consistently one of the leading causes of moderate to severe vision impairment worldwide, especially in adults aged 50 and over, where it is the second leading cause of blindness behind cataract.1
Unleashing the AI beast on glaucoma and DED would thus alter countless lives around the world. But there’s a problem.
To err is human, to diagnose is divine
Glaucoma and DED are two of the most stubborn holdouts resisting the inevitability of the AI revolution. The roots of the difficulty adapting the problems of these two diseases for powerful machines, though, ultimately lie with humans.
The issue with dry eye and glaucoma is that diagnoses can both vary from clinician to clinician. Practicing optometrists know that DED symptoms can be present without any clear signals in medical imaging.
“You can have a very clean clinical picture and still have a very symptomatic patient [with dry eye]”, said Dr. Mark Eltis, a prominent Canadian optometrist, lecturer, and prolific author in both glaucoma and DED for optometry.
Early stage, and especially normal-tension glaucoma (NTG), can evade even the most scrupulous and experienced observers. And this is when the diagnosis is most critical. “Glaucoma is a disease that doesn’t have any signs or symptoms for the most part,” explained Dr. Eltis. “If glaucoma is easy to diagnose, then the doctor may have been missing it for a long time or the patient has not come in regularly enough. It shouldn’t be easy.”
The difficulties for optometrists in glaucoma and DED only get murkier from there. Diagnosis of both diseases can vary greatly from practitioner to practitioner. And in the case of DED, management is subjective and highly tailored to the individual patient.
“We have so many treatments for dry eye, [as well as] diagnostics,” Dr. Eltis said. “The bad news is that buying a piece of equipment does not make you an expert. Like with any other condition, there’s no replacement for experience and training.” Though shiny tools and fancy imaging machines may be mainstays of the modern optometrist, detection and management are still down to individual practitioners and the weight of their experiences. And they can still be wrong.
This poses massive problems for many of the most common iterations of AI being used in medicine. Whether man or machine, the strength of any decision-maker lies in the data set used to train their judgment. Human cognition is in many ways the sum of all of the experiences we have — our data set. And if there are issues with the data set, errors occur. An American who has only visited Europe (and not Malta), for example, would think that everyone in the world drives on the right until told otherwise.
Machine learning (ML), including many kinds of so-called ‘neural networks’, computer systems built to mimic the processing style of the human brain, is similarly restricted by the data set.
Mr. Daniel Eckert, a member of the computer vision and machine learning research group at Huawei, Switzerland, and developer of a smartphone-based corneal topographer, provided some insight into the problems that AI excels in. “One thing [machine learning methods] are really good at is if you optimize them to a specific problem or question. Then you have a data set with the correct answers and it learns to replicate those answers,” he explained.
“I think it’s really powerful when it comes to these kinds of classification tasks where you need a yes or no response based on a well-defined problem,” Mr. Eckert continued.
The trouble, of course, is that DED and glaucoma are a classification task that is not so great with correct answers. As mentioned above, DED can have a negative clinical picture but be symptomatic. The nebulous nature of diagnosing early-stage glaucoma makes it difficult to discern what is glaucomatous and what is not.
DED to rights
At first glance, the prospects for the use of AI for DED and glaucoma in optometry may seem dismal. Depending on regional regulations, most optometrists need tools that will aid them in screening, managing, and monitoring DED and glaucoma.
But glaucoma and DED remain fundamentally classification problems, and new research into applications in the field are promising. In dry eye disease, machine learning algorithms, and especially DL algorithms and deep neural networks (DNN), are causing the biggest stir.
Tear break-up time (TBUT) is a common way of screening for DED. Thus far, four studies have been done with varying ML methods, and machines are able to get fairly close to human performance, including a 91% accuracy in detecting dry areas compared to optometrists.2
Artificial intelligence algorithms are also being fed images from interferometers and slit lamps to see if they can be of assistance. The tear film lipid layer is a font of information, and this direction of investigation is promising. Of particular interest to optometrists are studies being done to directly aid in the somewhat subjective diagnosis process, including a 100% accurate system for predicting ocular redness from slit lamp images.2
Artificial intelligence has also been used in the hunt for less conventional biomarkers to aid in DED diagnosis. Research has shown promising results using convolutional neural networks (CNN) on various features of corneal nerves,2 which were gold for detecting meibomian gland dysfunction (MGD), a key contributor to DED. Tear osmolarity has been looked at, but the more promising avenue for optometrists is the use of tear proteins. One model combining DNN and discriminant analysis had an astonishing 89.3% accuracy when distinguishing aqueous-deficient DED, MGD, and a healthy eye using the tear proteome.2
OCT is another interesting avenue for optometry in detecting DED, though it is still in its infancy. Corneal epithelium thickness has often been associated with DED, and a model using a CNN showed impressive accuracy in predicting severe DED,2 making this method a must for future research into image analysis.
Artificial intelligence doesn’t need to knock it out of the park with diagnosing dry eye to be useful. Time savings are also critical, especially in underserved areas, and AI showed great promise in time savings. For example, most, but not all, clinicians have little problem with meibography for looking at the meibomian glands. But a DL algorithm was found to outpace experienced clinicians in quantifying meibomian glands, and a higher degree of accuracy in determining meibomian gland atrophy.2
Just as with DED, uncertainty in diagnosis is the name of the game for glaucoma. And just like DED, modern machine learning algorithms are opening up doors in screening and monitoring unimaginable before AI and DL algorithms came to town.
The success of these algorithms, however, is more muted. This is because variability in screening for and diagnosing glaucoma, and especially challenging early-stage and normal tension cases, takes on a decidedly different flavor than dry eye. In DED, symptoms can be present without signals in imaging and biomarkers. In glaucoma, the disease can be asymptomatic and be nearly undetectable from imaging and biomarkers.
Dr. Eltis thinks part of this has to do with human error and a kind of imaging examination burnout. “There is a lot of manual looking and seeing [with glaucoma]. The truth is, regardless of how intelligent or perceptive or detail-oriented you are, you’re gonna miss stuff because that’s just the nature of being bombarded with too much information,” he said.
Sign AI up for the job? Not so fast. It’s true that artificial intelligence algorithms don’t succumb to human quirks like getting bored, distracted, or burned out. And with image processing and classification, AI does not really miss even minute variations in things like fundus imaging or OCT. But only when it is presented with a robust and concrete data set to train on.
This is exactly where the trouble begins for glaucoma. ODs and MDs do not always agree on when an imaging set indicates glaucoma. This also means that some are not always right. And even the most advanced artificial intelligence algorithms don’t jive with that.
According to a study assessing the issues facing AI in glaucoma,3 it’s all about the definition of what glaucoma actually is. The AI arrow needs a target to fly true, but this target is currently ill-defined. Once researchers can decide what glaucoma is, then it becomes manageable for AI.
“It’s [still] a classification problem,” explained Mr. Eckert. “Whether you define two categories or if you define 50 categories, it’s still the same class of problem.”
This is a simple-sounding, yet enormous task. There is wild variability in the three tests that show the most promise for AI analysis in screening — fundus imaging, OCT, and visual field testing. OCT is still wildly expensive and not widely available. There is no accepted global, and in some cases even regional, standard for any of the tests. This is particularly pressing in fundus imaging, which shows promise for low-resource environments, but is performed on anything from smartphones to clinical-standard equipment. Even some of the most promising research is being done on ethnically homogenous populations, which opens up another can of worms in data selection bias.
The list goes on. But standardization efforts are underway in the form of projects like the Crowd-Sourced Glaucoma Study at Dalhousie University in Canada, which aims to shore up OCT and visual field tests with specialist assessments of glaucoma likelihood, which would provide standardized inputs for AI to learn from and grow into valuable tools for optometrists screening for glaucoma.
A future beyond screening
Though AI glaucoma and DED have been slow out of the gate, they are well on their way to becoming prized ponies in the optometric toolbox. AI has demonstrated great ability in opening up new avenues for screening, but there is also untapped potential elsewhere.
Management is one key area that optometrists can hope to get a boost from with the helping hand of a machine, and this comes from predicting the course of both diseases. “For [prediction], machine learning is an amazing tool,” said Mr. Eckert. “The hope, of course, is that the algorithm can detect patterns that humans are not familiar with or that maybe a doctor doesn’t even know yet, and then maybe gain deeper insight into the mechanisms behind it.”
Another comprehensive study4 on DED illuminated how the predictive power of AI assessment of dry eye can optimize custom treatment protocols for the management of dry eye using not only AI analysis of imaging test results, but also from medical records and population-based studies.
Similar, wide-ranging work can be done on glaucoma for predicting the progression of the disease once the standardization puzzle is unlocked. The possibilities, from scanning patient records using AI-derived tools like natural language processing, to flagging patients at high risk for more frequent checkups, would further streamline and simplify the process of catching and stopping glaucoma in its tracks
Whatever the result is, AI can be a powerful tool. But whatever the prevailing sentiment on artificial intelligence in medicine, one thing is clear to Dr. Eltis: Nothing will ever replace a human doctor.
“Ultimately, the executive decision is still left to the captain of the ship, and that’s the doctor. That’s why, on a ship, you have experts in terms of navigation and everything else, which translate the data or transmit the data to the captain. But ultimately, it’s up to the captain to see the big picture and make an executive decision,” Dr. Eltis concluded.
- GBD 2019 Blindness and Vision Impairment Collaborators, & Vision Loss Expert Group of the Global Burden of Disease Study. Causes of blindness and vision impairment in 2020 and trends over 30 years, and prevalence of avoidable blindness in relation to VISION 2020: the Right to Sight: an analysis for the Global Burden of Disease Study. Lancet Glob Health. 2021;9(2):e144-e160.
- Storås AM, Strümke I, Riegler MA, Grauslund J, Hammer HL, Yazidi A, Halvorsen P, Gundersen KG, Utheim TP, Jackson CJ. Artificial intelligence in dry eye disease. Ocul Surf. 2022;23:74-86.
- Lee EB, Wang SY, Chang RT. Interpreting Deep Learning Studies in Glaucoma: Unresolved Challenges. Asia Pac J Ophthalmol (Phila). 2021;10(3):261-267.
- Yang HK, Che SA, Hyon JY, Han SB. Integration of Artificial Intelligence into the Approach for Diagnosis and Monitoring of Dry Eye Disease. Diagnostics (Basel), 2022;12(12):3167.
Editor’s Note: This article was first published in COOKIE magazine Issue 10.