AI In the Eye: Driving Solutions for Public Health

Artificial Intelligence (AI) is defined as machines simulating human-like thought processing behaviours. Automation may be considered a fundamental component of AI, which makes simple task delegation possible by setting the input for a desired output. As an analogy, the production of a cake can be automated by providing the recipe ‘algorithm’ and its ingredients. AI further encompasses machine learning which, by processing sets of ‘training data’ (the input with its corresponding output), the algorithm itself can be learned. This notion of ‘self-learning’ underpins its integrity to be used for a myriad of applications in eye care. AI can work out the recipe by analysing the cake and its ingredients. The recipe can then be refined through evaluation or ‘data validation’ of the cake, which improves its sensitivity and specificity. Deep learning is a subset of machine learning, where there are multiple algorithms forming artificial neural networks like the human brain. It involves a multi-step learning process which enables it to evaluate the different outcomes and improve itself to achieve optimisation. So what are the applications of AI in optometry today and its projected implications for the future?

a recent study has proposed the idea of distributing wearable OphthoAI headsets to patients that capture and analyse images using machine learning technology

DIABETIC RETINOPATHY

The prevalence of diabetes is projected to rise, especially in urban areas, from 171 million in 2000 to 366 million by 2030.¹ This has been attributed to changes in diet and selective mortality of those with diabetes-related complications in rural regions because of poor access to healthcare.² With more than 10% of those with diabetes subsequently having visionthreatening diabetic retinopathy (DR) (classified as having either proliferative DR or diabetic macular oedema),³ it has now become an epidemic and the leading cause of blindness among working-age adults.⁴

Ocular imaging technologies including OCT (pictured), are quickly becoming a mainstay in clinical practice. In conjunction with artificial intelligence, they are directing a new future for eye care.

The Problem 

Although routine dilated eye examinations are necessary to properly evaluate diabetic associated eye complications and reduce the risk of blindness, it is estimated that 57% of individuals with diabetes do not undergo regular eye examinations.⁵ Reasons for this include poor access to diabetic eye care (perceived by both patients and physicians)⁶ and poor public awareness of the importance of eye examinations. Screening programs have been introduced to improve access and education to diabetic eye care but are limited in capacity due to the large number of individuals with diabetes and the reliance on manpower. Could AI be the answer?

The Current Solution 

The most successful AI system for DR screening currently is an AI algorithm called IDx-DR. This AI-based diagnostic system classifies fundus images as either positive or negative of having more than mild DR and hence determines whether referral to an ophthalmologist is indicated. The clinical trial using IDx-DR on 900 participants found it exceeded “all prespecified superiority endpoints at sensitivity of 87.2% and specificity of 90.7%”⁷ and allowed it to become the first FDAapproved device in any field of medicine that can make a decision without clinician input. With such promising results, IBM Watson Health is endeavouring to introduce it to Australia.⁸ 

Looking to the Future 

Screening for DR is just the tip of the iceberg for AI. Foreseeably, AI-assisted self-screening could automate the initial diagnosis of DR and improve eye care accessibility through remote monitoring. This may enable reallocation of resources by reducing the number of unnecessary referrals (through screening), and allow physicians and policy makers to focus on optimising the treatment pathways for vision-threatening DR.

GLAUCOMA

The number of people with glaucoma worldwide is estimated to be 79.6 million in 2020.⁹ Its prevalence is projected to increase to 111.8 million by 2040.¹⁰ The clinical assessment for glaucoma involves measurement of intraocular pressure, anterior chamber angle evaluation, optic nerve head (ONH) analysis, visual field testing and optical coherence tomography (OCT) imaging of the retinal nerve fibre layer (RNFL).

The Problem
Population-based studies suggest that over half of all glaucoma cases remain undiagnosed worldwide.^11–14 Reasons for this are numerous. Due to the chronic and progressive nature of glaucoma, patients may initially be asymptomatic and therefore not present for screening. Diagnosis of glaucoma also involves observation of structural deficits in ONH and RNFL as well as functional deficits (as measured by standard automated perimetry) which can often be discordant.^15,16 Finally, glaucoma diagnosis is resource-intensive, requiring multiple repeated tests to establish a reliable baseline for change analysis.

The Current Solution 

Numerous machine learning models have been used to differentiate between normal and glaucomatous eyes, based on features such as visual fields, RNFL thickness and ONH analysis results. These approaches have shown high sensitivity and specificity with one classifier achieving an AUC result of 0.946.¹⁷ The various approaches may rely on either single or multiple modalities. Not surprisingly, by combining structural (OCT) and functional (standard automated perimetry) measurements as input, one study revealed that it provides improved diagnostic accuracy for detecting glaucomatous eyes compared to stand alone measurements.¹⁸ Other researchers have developed a method to detect visual field progression in glaucoma by quantifying pattern changes using an unsupervised archetypal analysis. This information can then be incorporated into the decision-making process of clinicians to yield more accurate diagnosis of glaucoma progression.¹⁹ More research and development in the areas of trend analysis and its interplay with visual function in glaucoma are required.

In response to growing public health concerns, emerging AI technology presents an opportunity to improve the affordability, quality and accessibility of eye care

Looking to the Future 

Recently, IBM Research in collaboration with New York University has conducted a study to diagnose glaucoma patients by analysing raw OCT images using a deep learning framework. The new approach allows it to by-pass the segmentation process.²⁰ The AI model achieves an estimation of the corresponding visual field index values from the OCT imaging data, with an overall error within 2% and further lays the groundwork for future technologies that can potentially use this type of analysis to gain a better understanding of the patient’s visual function.²⁰ This highlights the potential value of feature-agnostic AI algorithms in large population-based disease screening to help alleviate the public health problem of undiagnosed glaucoma.

AGE-RELATED MACULAR DEGENERATION

Age-related macular degeneration (AMD) is the leading cause of blindness in developed countries, accounting for 8.7% of all blindness worldwide and its prevalence is estimated to rise to 288 million by 2040.²¹ With the exponential growth of the ageing population, it is inherently an ongoing and an ever-more pressing public health concern. Currently, the mainstay of optometric management for early and intermediate AMD involves counselling on modifiable risk factors, providing an Amsler grid for self-monitoring and informing patients to return to clinic if a change in their vision is noted.²²

The Problem 

Choroidal neovascularisation (CNV) is the main cause of severe vision loss in late AMD and requires early detection for effective initiation of treatment. However, patients with CNV are often asymptomatic which leads to poor detection (by both patient and/or clinician).²³ As such, many clinicians rely on the Beckman classification scheme to estimate five-year risk of AMD patients to progress to late AMD and organise appropriate monitoring.²² Research shows that practicing clinicians nonetheless tend to underestimate the severity of AMD with other risk factors such as smoking, hypertension and body mass index not considered.²⁴

The Current Solution 

AI applied to the analysis of imaging biomarkers has allowed the personalised prediction of AMD progression.²⁵ One such machine learning tool ranks morphological features found in AMD based on their associated risk of conversion to CNV or geographic atrophy (GA). The study found that “differential pathways became visible for the neovascular and the atrophic pathways in AMD” with focal decompensation within drusen seen to be associated with CNV as opposed to general retinal ageing in GA. By further integrating genetic and demographic parameters, AI technology can be seen as a novel method in determining the risk profile of a patient in progressing to late AMD, enabling clinicians to implement a more effective management plan for their patients.²⁷ 

Looking to the Future 

By providing a more accurate, personalised estimate of the risk of progression to late-AMD, AI provides the potential for more timely treatment. Furthermore, a recent study has proposed the idea of distributing wearable OphthoAI headsets to patients that capture and analyse images using machine learning technology which is then uploaded for a clinician to review.²⁶ Concurrent advancements in cloud computing could also support continuous data collection and automatic analysis for on-going AMD management. This highlights the potential interplay of multiple technologies in conjunction with AI on a larger scale to alleviate the public health concern of vision loss due to AMD.

In response to growing public health concerns, emerging AI technology presents an opportunity to improve the affordability, quality and accessibility of eye care. AI is evolving and is already being used to help clinicians with grading the severity of DR. Further research in similar applications is underway to assist in diagnosing glaucoma and predicting the clinical outcome of patients with AMD. Although there are currently ethical debates on patient safety and privacy, a new AI-integrated eye care paradigm holds tremendous promise for improving the world’s major public health problems in eye care.

Judy Nam is a Research Assistant at the Centre for Eye Health in Sydney NSW. 

Angelica Ly is the Integrated Care Co-ordinator and Lead Clinician (Macula), at the Centre for Eye Health. She is also an Associate Lecturer at the School of Optometry and Vision Science, UNSW. 

Lisa Nivison-Smith is a Senior Research Associate at the Centre for Eye Health and the School of Optometry and Vision Science, UNSW. 

References 

1. Wild S, Roglic G, Green A et al. Global prevalence of diabetes: estimates for the year 2000 and projections for 2030. Diabetes Care 2004; 27: 1047-1053. 
2. Lee R, Wong TY, Sabanayagam C. Epidemiology of diabetic retinopathy, diabetic macular edema and related vision loss. Eye Vis (Lond) 2015; 2: 17. 
3. Yau JW, Rogers SL, Kawasaki R et al. Global prevalence and major risk factors of diabetic retinopathy. Diabetes Care 2012; 35: 556-564. 
4. Zheng Y, He M, Congdon N. The worldwide epidemic of diabetic retinopathy. Indian J Ophthalmol 2012; 60: 428-431. 
5. Fisher MD, Rajput Y, Gu T et al. Evaluating Adherence to Dilated Eye Examination Recommendations Among Patients with Diabetes, Combined with Patient and Provider Perspectives. Am Health Drug Benefits 2016; 9: 385-393. 
6. Hartnett ME, Key IJ, Loyacano NM et al. Perceived barriers to diabetic eye care: qualitative study of patients and physicians. Arch Ophthalmol 2005; 123: 387-391. 
7. Abramoff MD, Lavin PT, Birch M et al. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. NPJ Digit Med 2018; 1: 39. 
8. IBM. IDx and IBM Watson Health Forge Alliance for Eye Health. In. CAMBRIDGE, Mass. and Iowa City, Iowa, 2017. 
9. Quigley HA, Broman AT. The number of people with glaucoma worldwide in 2010 and 2020. Br J Ophthalmol 2006; 90: 262-267. 
10. Tham YC, Li X, Wong TY et al. Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmology 2014; 121: 2081-2090. 
11. Quigley HA. Glaucoma. Lancet 2011; 377: 1367-1377. 
12. Keel S, Xie J, Foreman J et al. Prevalence of glaucoma in the Australian National Eye Health Survey. Br J Ophthalmol 2019; 103: 191-195. 
13. Chua J, Baskaran M, Ong PG et al. Prevalence, Risk Factors, and Visual Features of Undiagnosed Glaucoma: The Singapore Epidemiology of Eye Diseases Study. JAMA Ophthalmol 2015; 133: 938-946. 
14. Wong EY, Keeffe JE, Rait JL et al. Detection of undiagnosed glaucoma by eye health professionals. Ophthalmology 2004; 111: 1508-1514. 
15. Susanna R, Jr., De Moraes CG, Cioffi GA et al. Why Do People (Still) Go Blind from Glaucoma? Transl Vis Sci Technol 2015; 4: 1. 
16. van der Schoot J, Reus NJ, Garway-Heath DF et al. Accuracy of matching optic discs with visual fields: the European Structure and Function Assessment Trial (ESAFAT). Ophthalmology 2013; 120: 2470-2475. 
17. Silva FR, Vidotti VG, Cremasco F et al. Sensitivity and specificity of machine learning classifiers for glaucoma diagnosis using Spectral Domain OCT and standard automated perimetry. Arquivos Brasileiros de Oftalmologia 2013; 76: 170-174. 
18. Bowd C, Hao J, Tavares IM et al. Bayesian machine learning classifiers for combining structural and functional measurements to classify healthy and glaucomatous eyes. Invest Ophthalmol Vis Sci 2008; 49: 945-953. 
19. Wang M, Shen LQ, Pasquale LR et al. An Artificial Intelligence Approach to Detect Visual Field Progression in Glaucoma Based on Spatial Pattern Analysis. Investigative Ophthalmology & Visual Science 2019; 60: 365-375. 
20. Maetschke S, Antony B, Ishikawa H et al. A feature agnostic approach for glaucoma detection in OCT volumes. PLoS One 2019; 14: e0219126. 
21. Wong WL, Su X, Li X et al. Global prevalence of age-related macular degeneration and disease burden projection for 2020 and 2040: a systematic review and meta-analysis. Lancet Glob Health 2014; 2: e106-116. 
22. Optometry Australia. Diagnosis, treatment and management of Age-Related Macular Degeneration. In. 2019 Clinical Practice Guide, 2019. 
23. Schwartz R, Loewenstein A. Early detection of age related macular degeneration: current status. Int J Retina Vitreous 2015; 1: 20. 
24. Ly A, Nivison-Smith L, Zangerl B et al. Advanced imaging for the diagnosis of age-related macular degeneration: a case vignettes study. Clinical & experimental optometry 2018; 101: 243-254. 
25. Schmidt-Erfurth U, Waldstein SM, Klimscha S et al. Prediction of Individual Disease Conversion in Early AMD Using Artificial Intelligence. Invest Ophthalmol Vis Sci 2018; 59: 3199-3208. 
26. Das A, Rad P, Choo K-KR et al. Distributed machine learning cloud teleophthalmology IoT for predicting AMD disease progression. Future Generation Computer Systems 2019; 93: 486-498.