Artificial intelligence in glaucoma diagnosis and management: Has its time come?

Abstract

Artificial intelligence (AI) has been used in glaucoma diagnosis and management. The uses range from summarizing patient notes from electronic medical records to diagnosis based on fundus or optical coherence tomography (OCT) images, detecting progression on imaging or visual fields, and, less commonly, assessing target intraocular pressure (IOP) and predicting surgical failure.[1–6] Using AI-based classification of disc photographs for population-based screening of undetected glaucoma in India based is an attractive proposition. The implementation of fundus photography-based diabetic retinopathy screening programs across some states is an opportunity. Adding an AI-based grading for the disc is a logical value addition to the program. However, there are challenges from both logistic and AI perspectives. The logistic concerns are applicable to any screening program for glaucoma, the increased pressure on the health care system with these referrals, the high proportion of ungradable images, the false positives and false negative rates, and so on. From the AI standpoint, the replicability of many algorithms in the population has been far below that in a controlled study environment. In addition, algorithms do not always perform adequately in populations of other ethnicities than the test cohort.[2] Any algorithm would need to be adequately validated in appropriate populations before it is rolled out on a larger scale. Fortunately, there are Indian AI disc photograph-based diagnostic algorithms, though adequate large-scale validation is a challenge.[7] The use of AI to detect progression of OHT to glaucoma or predict progression of existing disease earlier than current techniques could help identify those eyes at greatest risk for closer follow-up. This could result in reduced patient visits while helping prevent glaucoma-related morbidity. However, there are very few datasets globally with adequately documented longitudinal clinical data; these include large randomized control trials conducted in the West, which do not necessarily apply to all at-risk individuals.[5] The lack of good longitudinal India datasets is a challenge in this regard since the applicability of progression datasets across ethnicities is still an unknown quantity. The number of practices in the country that would do serial visual field and imaging in glaucoma patients using appropriate progression software with Electronic Medical Records (EMR) that capture other parameters to make such assessments is very limited. Having to input large amounts of data and images from fragmented sources is too time-intensive to be of practical use. While some of the diagnostic tools may have relevance in large eye hospitals to triage referrals to the glaucoma clinic, it would be more practical to do a comprehensive evaluation of a glaucoma suspect by a glaucoma specialist instead. In health care settings where primary care is by a nonophthalmologist, these tools fit better with the referral model of the “at risk” patient. For the vision center model, these could be valuable additions by reducing the pressure on the tele-reading centers and reducing unnecessary referrals.[7] One fundamental challenge with any glaucoma diagnostic algorithm is the definition of glaucoma.[8] Unlike diabetic retinopathy where the presence, absence, and distribution of clinically obvious retinal findings help in clear definition of disease, the definition of glaucoma is challenging. Using a visual field-based definition can potentially result in missing early pre-perimetric disease; using only imaging-based definitions can result in overdiagnosis of glaucoma. Different groups have used varied definitions for glaucoma, and this makes comparing algorithms and clinical utility difficult. Any technology comes with costs, and this is an unanswered question with the use of AI in glaucoma. In a screening setting, it would be beneficial by reducing reading center workloads. Would a non-glaucoma specialist be willing to pay for additional software that would help them diagnose disease or predict progression? That is a difficult question to address, but I suspect that the answer would be that they would not. Ophthalmology is already a very equipment-intensive field, and additional costs without very clear benefits are unlikely to be adopted. A per use model may be more acceptable, but how the cost would be passed on to the consumer has to be addressed. The most acceptable system would probably be if the model was integrated into the hardware itself and this has happened especially with fundus cameras. Given that most standard perimeters and OCT systems are not made in India, any AI model that incorporated into these devices in the future would need to have been tested on adequate Indian datasets, especially for diagnosis. Any deployed algorithms should also be upgradable as the model improves its sensitivity for diagnosis. One should also keep in mind that the less experienced clinicians or trainees who are most likely to rely heavily on AI can also be most easily misled by an incorrect AI response. From “red disease,” we may move to era of “AI disease.”[9] Currently, most ophthalmologists are very suspicious of the use of AI in clinical ophthalmology and that would hinder wider acceptability.[10] This could be ameliorated by the adoption of explainable AI models where the clinician understands the parameters that went into making the decision instead of a “black box” model that does not specify this. AI-based diagnostic assistance is already employed in some screening programs today. Validating these models in different ethnic groups and standardizing the definition of glaucoma for diagnosis are other crucial areas that would improve acceptance. AI-assisted clinical care is likely to become more widely adopted for glaucoma care in the future as and when some of these challenges are resolved. Adoption of EMR with networked diagnostic equipment will facilitate the process. One concern, in the long term, with greater adoption of these technologies is the risk of further atrophy of glaucoma diagnosis skills in ophthalmic trainees.[9] This would drive a more technology and instrument-based practice which could drive up costs with potentially poorer clinical outcomes. Adoption of AI-based tools is inevitable, but it should supplement a good clinical evaluation and not replace it.About the authorProf. Ronnie George Professor Ronnie George is currently a senior consultant for glaucoma services at Sankara Nethralaya, Chennai. He also serves as the Research Director for the organization in addition to being the President of Vision Research Foundation. His areas of research focus have included glaucoma epidemiology, glaucoma genetics, and glaucoma diagnostics. He was part of the Vellore Eye Survey follow-up studies that reported seminal data on the progression of angle closure suspects. He was also a clinical investigator on the Chennai Glaucoma Study and the Chennai Eye Diseases Incidence Study which contributed significantly to the understanding of glaucoma epidemiology in India and the natural history of the disease. He was part of global consortia that reported the first genetic associations for Primary Angle Closure Glaucoma. He has over 200 Pubmed indexed publications with an h-index of 57. He has served as the Secretary Glaucoma Society of India, President, ARVO India chapter and serves on multiple editorial boards. He is a member of the prestigious Glaucoma Research Society and Affiliate Professor, Deakin University, Australia.