In this interview, section editors Nima Ghadiri and Arun James Thirunavukarasu ask Ophthalmic Surgeon and NIHR Doctoral Fellow Shafi Balal about the development of his AI model for predicting keratoconus progression risk, with the longer-term objective of using AI on anterior segment data for oculomics discovery.
Shafi Balal.
Tell us about the models you developed and what they do.
We developed three models that predict, at a patient’s first or second clinic visit, whether their keratoconus will progress over the following two years: a multimodal deep learning ensemble (MMDINO) combining anterior segment optical coherence tomography (AS-OCT), Placido topography and tabular clinical data; a long short-term memory (LSTM) that refines this prediction using data from a second visit; and an XGBoost model on tabular data alone as a device-agnostic fallback. Outputs are used to triage patients into three care pathways: community optometry review, continued hospital monitoring, or expedited referral for corneal cross-linking. In addition, we have developed models that automate diagnosis of normal versus keratoconus (99% area under the curve) and automate monitoring. This is particularly important as there is currently up to a 44-week wait in parts of the UK to see a corneal specialist for diagnosis, in which time scarring and progression of disease may occur. It is one of the reasons why keratoconus remains the number one indication for a corneal transplant in the West.
Where did the data come from, and how did this strengthen validation?
Data came from Moorfields Eye Hospital, curated and accessed through the INSIGHT health data research hub, with external validation from Croydon University Hospital. Holding out an entire institution rather than a random split provides a more honest test of generalisability and helps protect against the kind of site-specific overfitting that has undermined many published keratoconus AI models.
Sensitivity is under 100% – is the risk of missing progressors acceptable when extending follow-up intervals?
Sensitivity at a 0.5 probability threshold cut-off is the wrong frame for this question; what matters clinically is the negative predictive value at the operating threshold, and we set this to ≥90% (i.e. fewer than one in ten patients allocated to community review will progress over two years) with the option to tighten to 95% where services prefer. This risk is acceptable provided patients allocated to community review have a clearly defined safety-net pathway back into hospital care should symptoms or refractive change emerge – and it should be compared against the current alternative, which is over-monitoring of low-risk patients at the cost of delayed access for higher-risk ones.
How do you envisage AI triage being integrated into ophthalmology?
Initially, as a decision-support layer attached to the imaging device itself, returning a risk classification at the point of scan acquisition that the clinician uses alongside their own assessment to set follow-up intervals. Over time, a fully automated pathway (from diagnosis to triage to monitoring). Early diagnosis could lead to early preventative cross-linking treatment that halts progression and reduces corneal transplant burden. And triage of stable, low-risk patients into community optometry pathways would free hospital capacity for higher-risk and treatment-requiring cases.
What needs to happen before deployment?
Three things in parallel: a period of silent prospective running at the deployment site to verify that predictive values match those reported here in the local population; UKCA/CE marking as a Class IIa medical device with the supporting clinical evaluation file; and a structured co-design exercise with the clinical and optometry teams who will operate the triage pathway, particularly around threshold setting and the safety-net referral route.
Now, considering oculomics and discovery science, how transferable are these latent features across devices?
The MS-39 inputs (Placido topography and AS-OCT cross-sections) capture the same underlying corneal geometry as Pentacam Scheimpflug or other AS-OCT platforms, so the biological signal should generalise – but the raw image features are device-specific, so direct deployment on a different platform without retraining is unlikely to work. The most likely transfer route is to fine-tune the KERAFound encoders on a smaller labelled dataset from each new device, exploiting the label efficiency demonstrated by foundation models to keep the retraining burden manageable. We have also trained a separate model which uses Pentacam tabular data to automate diagnosis, monitoring and triage.
Could latent features identify new phenotypes of progression?
Yes – unsupervised clustering of the KERAFound feature embeddings, rather than the supervised classification used here, could reveal whether progressors fall into distinct morphological sub-types (for example, central vs paracentral cones, epithelium-dominant vs stroma-dominant patterns) that current grading systems collapse into a single ‘keratoconus’ label. This is a natural next study and would benefit from the same probabilistic semi-supervised framework that Kandakji, et al. applied to subclinical disease [1].
Have you correlated model features with systemic or environmental risk factors?
Not yet. The present datasets capture tomography and basic demographics but not atopy, rubbing behaviour, sleep position, hormonal status or connective tissue traits in structured form. Linking the KERAFound embeddings to a phenotype-rich cohort (for example, through INSIGHT-linked questionnaire data or a prospective recruitment study) is a next step and one of the more promising avenues for generating biologically meaningful hypotheses about progression drivers. Could this be extended into an ‘anterior segment oculomics’ platform? Yes, and this is the most ambitious vision for the work. A shared AS foundation model produces a single learned representation of corneal shape, epithelial thickness, posterior elevation and higher-order aberrations that could be regressed against systemic traits (atopy, connective tissue genotypes, metabolic markers and systemic disease) and against polygenic risk scores in linked biobank-scale data – the AS analogue of the retinal oculomics programme exemplified by RETFound [2]. In fact, we have already demonstrated that there are hidden systemic bio signals within AS imaging [3]. These signals could be potentially exploited to predict systemic disease (oculomics similar to RETFound) but further work (including linkage to general practitioner health records) is required.
References
1. Kandakji L, Balal S, Stupnicki A, et al. Data-driven detection of subclinical keratoconus via semi-supervised clustering of multidimensional corneal biomarkers. Ophthalmol Sci 2025;6(2):100998.
2. Zhou Y, Chia MA, Wagner SK, et al. A foundation model for generalizable disease detection from retinal images. Nature 2023;622(7981):156–63.
3. Balal S, Cox L, Khan A, et al. Investigating the capability of deep learning models to predict age and biological sex from anterior segment ophthalmic imaging: a multi-centre retrospective study. BMJ Open 2025;15(10):e107196.
Declaration of competing interests: None declared.
