EYESAVE: AI-enabled triage for Glaucoma

Glaucoma – a heterogeneous, chronic disease affecting the optic nerve - is the leading cause of irreversible blindness worldwide and the second most common cause of certifiable sight loss in the developed world.

Older age is a significant risk factor for glaucoma, with approximately 4% of the over-50s in the UK affected by the disease. Its proper management is critical to preserving vision.

Early detection is crucial, but the symptoms of the disease typically go unnoticed, initially involving only peripheral vision, until the underlying damage has reached an advanced stage. The rate of disease progression varies greatly across individuals. The damage it causes is irreversible, and there is no single diagnostic or prognostic test for glaucoma.

To gauge disease presence and assess risk of vision loss, ophthalmologists must interpret results of multiple disparate assessments, including both objective measures (e.g. intraocular pressure, corneal thickness) and subjective measures of visual field extents. Acquiring, collating and evaluating these measures is time-consuming, inefficient, and bias-prone.

Additional risk factors associated with glaucoma, particularly ethnicity, cardiovascular disease and other co-morbidities, must also be included in clinical decision-making.

There is huge need for improved efficiency and accuracy of risk-stratification for patients with suspected or confirmed glaucoma, to prevent sight loss.

Non-invasive imaging via Optical Coherence Tomography (OCT) allows more accurate measurements of glaucomatous damage to critical retinal structures, in particular the retinal nerve fibre and ganglion cell layers, and is already incorporated in some glaucoma assessment pipelines (e.g. the Hood Glaucoma Report).

However, the direct interpretation of OCT images or quantitative descriptors extracted from these via proprietary imaging software requires trained ophthalmologists, now in short supply in the UK. Artificial intelligence (AI) has significantly improved automated OCT image analysis.

Our research group has developed AI models which analyse retinal structure to detect early signs of neurodegenerative diseases, and differentiate glaucomatous from non-glaucomatous optic neuropathy.

Our goal is to create an AI tool that combines analysis of longitudinal retinal OCT images, fundus images, and non-imaging-based clinical data to stratify the risk of visual loss in glaucoma patients, alerting clinicians to those who most urgently need medical or surgical intervention. The project has three work streams: (1) Database creation.

We will create an annotated OCT image database linking all available successive retinal scans and clinical data on longitudinal disease status, measurements, and interventions, for thousands of patients in local NHS Trust glaucoma services. (2) Development of AI tool.

Current deep learning methods in glaucoma diagnosis are based on particular features in individual imaging modalities, and often focus on single scans, instead of.

We aim to create an AI tool that forecasts progression from successive clinical and imaging data points by: (a) fusing information from different imaging modalities; (b) including other clinical data, particularly longitudinal data where available; and (c) creating interpretable visualisations of the AI system’s decision-making to enhance its trustworthiness. (3) Demonstration of feasibility and acceptability.

The final risk stratification tool will quantify disease severity and risk of visual loss, supporting clinical decision-making concerning subsequent interventions.

We will test the effectiveness and generalisability of the AI tool on previously unseen data from diverse patient groups, comparing the tool’s classifications with known clinical outcomes. The unseen data will include supplementary data from local NHS Trusts as well as our external collaborators.

Through our public, patient and clinician engagement work, we will further evaluate the feasibility of deploying the software in the clinic by assessing clinicians’ views on user interface features as well as essential elements of the software-generated report.

Patient and public panel members will provide input on the acceptability of the software and AI-assisted decision-making in general, as well as on accessibility of patient-focused outputs from the software.

By creating a tool that accurately predicts disease severity and risk of visual loss, we will guide and improve efficiency of clinicians’ decision-making, save clinicians’ time and maximise their impact, and ultimately save the eyesight of more patients.