Utilizing Artificial Intelligence for Enhancing Glaucoma Care

Glaucoma is the leading cause of irreversible blindness globally, with 3.6 million people blind and nearly 4.14 million visually impaired due to the disease in 2020. It accounted for 8.4% of all blindness globally and 4.9% in South Asia.¹ Application of artificial intelligence (AI) technologies is bringing about a tremendous revolution in all walks of life, including ophthalmology.² Three distinct but overlapping and related terminologies—AI, machine learning (ML), and deep learning (DL)—are often utilized.

Artificial intelligence is a field of data science concerned with building computers and machines that can reason, learn, and act in such a way that would normally require human intelligence or that involves data whose scale exceeds what humans can analyze.³ ML is a subset of AI where the computer is provided with data and the output. The computer identifies patterns from features in the data to predict the output. The ML model thus developed is then used to predict the output based on new data.⁴ Various learning methods can be utilized (supervised learning, unsupervised learning, reinforcement learning), and some of the classical ML algorithms include regressions, decision trees, LASSO, k-nearest neighbors (KNN), support vector machines (SVM), and random forests. DL is a specific ML method which is inspired from biological neuronal networks and is particularly suited to tasks of image analysis. “deep” in DL refers to the multiple layers of neurons which process the data to generate the output.⁵ Neural network-based methods were used for visual field analysis in the mid-1990s, but they primarily remained in research laboratory benches.^6,7 Advancements in computing, specifically the development of parallel processing techniques leveraging graphics processing units and hardware that accelerate DL tasks, have led to a surge in development of ML/DL-based applications and techniques for healthcare that have wide bedside utility, and AI/ML/DL is upending the way glaucoma care will be delivered in the near future.^8–10 AI research can be used for glaucoma detection, differentiation, quantification, prognostication, and disease prediction.¹¹

Key challenges have been identified for glaucoma in the context of global eye health, including: (1) affordable and effective screening approaches, (2) effective treatments that prevent glaucoma progression, and (3) monitoring patients for progression of glaucoma and adapting management.¹² These challenges need to be managed in a broader health system that promotes engaging communities, enhances equity, and addresses shortages of trained eye health personnel for glaucoma management. We believe AI can play a transformative role here.

Glaucoma Screening and Detection in Primary Care Settings

The level of undetected primary open-angle glaucoma (POAG) is a high 57% in North America and a staggering 90%+ in Africa and Asia, underscoring the importance of finding solutions to reduce this diagnostic gap.¹³ The solutions can be technological as well as health-systemic. The focus has largely been on using AI-enabled retinal fundus imaging. Fundus photography is recognized as a low-cost, noninvasive, and quick technique, and portable fundus cameras have allowed screening to be done at the doorstep and in primary eye care centers/vision centers. Multiple research papers have been published on the use of AI for automated classification of monoscopic fundus images based on convolutional neural networks (CNNs) architectures such as ResNet, InceptionNet, and VGGNet, and use of vision transformers. The ML models for screening typically aim for very high sensitivity to avoid missing true positives while maintaining high specificity to minimize false referrals. In general, the models have achieved sensitivity of 80%, and many of these systems have outperformed humans.^9,14,15 The Indian Medios model that works offline has shown sensitivity and specificity of 93.7 and 85.6%, respectively, in the detection of referable glaucoma and had high repeatability.¹⁶ Handheld nonmydriatic fundus cameras and smartphone-based fundus cameras have become commonplace in India, and costs have reduced, opening up an exciting new possibility where trained technicians can offer glaucoma screening at the vision center level. The cost-effectiveness of these models will need to be ascertained, but integrated screening models where the same fundus camera is used to offer diabetic retinopathy and glaucoma screening will surely help in reining in the costs.

While fundus photographs look for structural evidence of glaucoma, functional evidence in the form of visual fields changes can also be used for screening. This is where online perimetry delivered via laptop, desktop, or tablet computers will play an important role. Studies have demonstrated accuracy of Melbourne rapid fields (MRF) testing and online circular contrast perimetry (OCCP) test.¹⁷ Studies indicate similar sensitivity, specificity, area under curve (AUC) for diagnosing glaucoma, agreement with standard automated perimetry (SAP), and user preference to OCCP.^18–20 These online perimetry tests also utilize AI. In a tertiary hospital setting, AI-enabled analysis of optical coherence tomography (OCT) has gained traction as well. Machine-to-machine (M2M) approaches have shown promise in quantifying retinal nerve fiber layer (RNFL) loss from color fundus photographs.

Detection of primary angle closure (PAC) disease has remained an elusive frontier primarily due to lack of availability of equipment that allows visualization of the angle in a rapid, noncontact, noninvasive manner which is implementable in vision centers. Studies have attempted classifying PAC/primary angle closure glaucoma (PACG) based on volumetric data from swept-source OCT (SSOCT), Anterior Segment (AS)-OCT, gonioscopy photographs, and ultrasound biomicroscopy (UBM) images, qualitatively as well as quantitatively. However, these technologies are not generally available in primary eye care settings, more so in developing countries. Approaches that leverage simpler modalities such as anterior segment photographs taken using slit illumination with customized smartphone and slit-lamp device, diffuse lighting slit lamp photography, as well as fundus photographs, have been used to predict anterior chamber depth. Still, there is an unmet need of AI technologies that can help screen for PAC/PACG in low-resource settings.¹¹ Work also needs to be done to assess the application of these techniques outside of controlled settings in routine ophthalmic care situations with diverse patient demographics, differing imaging equipment, and nonstandardized operating procedures.

Monitoring Progression and Predicting Outcomes

Being a chronic disease, the ability to detect and predict glaucoma worsening is an important clinical and public health imperative. This has been targeted through DL models that use visual fields,^21,22 retinal flowmetry,²³ and EMR data containing clinical notes and various indices of glaucoma tests.^24–26 The outcomes have ranged from future glaucoma measures (such as MD, PSD, VFI, RNFL loss) to the need for surgery. Integration of multimodal data, including functional assessments, and extensive validation in real-world clinical settings has been recommended to enhance prediction of glaucoma progression.²⁷

Supporting Patients Through Large Language Models

Rapid developments in the works of large language models (LLMs) have taken place in the last 4–5 years. LLM technologies like ChatGPT and DeepSeek and personal digital AI assistants are perhaps the next revolution after the smartphone. This has also opened the possibility of medical LLMs that can answer glaucoma-related queries with evidence-based answers in layman terms. LLMs can be used to improve a patient's experiences in health centers, facilitate patient triage, history capturing, and enhance ophthalmic education.²⁸ The LLMs can not only respond to text questions but can also consume image data, voice inputs, PDF reports, etc. Through techniques of prompt engineering and fine-tuning, the LLMs are trained to give more accurate responses for glaucoma. Researchers have used chat models to diagnose glaucoma based on demographic, clinical, ocular, visual field, optic nerve head photo, and history of disease parameters.²⁹ LLM chatbot has been demonstrated to outperform glaucoma specialists in diagnostic and treatment accuracy.³⁰ However, researchers have also highlighted issues with commonly available LLMs such as nonspecific and outdated training, no access to current knowledge, plausible-sounding “fake” responses or hallucinations, lack of critical literature analysis, and ethical and copyright issues.³¹ The World Health Organization has launched a smart AI resource assistant for health (SARAH), which is also available in Hindi,³² and in testing by the authors, it gave reasonable responses to questions about glaucoma from a layman's perspective.

Limitations of Artificial Intelligence Models

The development of DL/AI models requires a large dataset and is limited by the quantity and quality of the supplied dataset. It is difficult to estimate how much data is actually required to reliably train the DL systems, as this depends on the quality of data supplied as well as the complexity of the task. Additionally, the generation of AI models requires creation of initial datasets that are labeled by human graders and are thus noisy and depend on the expertise of the grader. Although these models generate a simple output (e.g., fundus camera image labeled as normal or glaucoma), the image features used to identify glaucoma are not highlighted (known as the “black box” problem). The highest accuracy of AI models is possible when data from both structural and functional tests are combined (also incorporating genetic data in the future). Use of fundus photos along with OCT (RNFL + GCC) and visual field data would be the best option for an accurate diagnosis; however, this would be a very costly venture for healthcare systems. Using AI for a very early diagnosis (e.g., preperimetric glaucoma) may also be counterproductive, leading to early treatment and high costs that may overwhelm the healthcare system. Use of AI models for screening should focus on picking up moderate disease so that the entire process is cost-effective.

Challenges and Way Ahead

The WHO has listed six ethical principles for use of AI in health, namely: (1) protect autonomy, (2) promote human well-being, human safety, and public interest, (3) ensure transparency, explainability, and intelligibility, (4) foster responsibility and accountability, (5) ensure inclusiveness and equity, and (6) promote AI that is responsive and sustainable. These need to be assessed in the context of glaucoma as well. Explainability of models is a major focus of research currently.³³ There is a need to develop large, diverse, longitudinal databases that can be used for developing AI algorithms. The diversity needs to be in terms of the ethnicities of subjects, full spectrum of disease diversities and glaucoma disease mechanisms, and acquisition mediums, and collected from a range of settings (primary, secondary, and tertiary care). We need to leverage simpler and more widely available imaging modalities that will be implementable in primary care settings. Another promising direction is use of multimodal AI and large multimodels (LMMs) where more than one input data source and model are utilized. This may include, for example, combining VF and color fundus photographs in one common AI model.

In public health research, there is a clear distinction between assessing efficacy in controlled settings and assessing operational effectiveness in the real world. Finally, there is a need to assess these applications in real-world settings, varying from tertiary care to primary care scenarios in developing countries. While connectivity challenges are being addressed through enhancements in mobile network penetration and access to smartphones, there are still challenges in making complex glaucoma assessment modalities available at district hospitals and lower levels in underdeveloped economies. We certainly do not see HFA or OCT being available at primary health facilities any time soon. Therefore, there is a need to prioritize leveraging simpler assessments that can be made available, such as smartphone-based fundus photography, computer-based or tablet-based perimetry, or simpler anterior segment imaging, which will enable wider applicability in low-resource settings. In conclusion, AI/ML/DL technologies for glaucoma hold great promise, which is being realized in research and some tertiary care settings. It is only a matter of time before benefits percolate down to primary eye care settings and enhance glaucoma care worldwide.