Diabetic retinopathy (DR) is a leading cause of preventable blindness, yet current “one-size-fits-all” screening intervals ignore large differences in individual risk. This project will use UK Biobank’s large-scale, longitudinal and multimodal resources, including retinal images, biochemical measurements, primary care and hospital records, medications, and lifestyle factors, to develop trustworthy models that predict both the onset and progression time of DR at the individual level.
Our key research questions are: (1) how accurately can multimodal longitudinal models predict time to incident DR and transitions between DR severity stages compared with existing risk scores; (2) which combinations of imaging and non-imaging features are most informative, and how these patterns vary across demographic and clinical subgroups; (3) how to quantify uncertainty and provide explanation- and case-based outputs that make predictions clinically interpretable and suitable for shared decision-making; and (4) what is the potential impact of personalized screening intervals on missed sight-threatening DR and on overall examination burden.
Methodologically, we will build compact latent representations of irregular longitudinal trajectories, fuse retinal images with tabular and text data, and apply deep time-to-event and multi-state survival models with explicit uncertainty quantification and causal sensitivity analyses. We will rigorously assess discrimination, calibration, fairness and robustness, and simulate alternative, risk-stratified screening strategies informed by our models. The ultimate objective is to provide an evidence base for adaptive DR screening policies that better target limited ophthalmic resources, enable earlier identification of rapidly progressing patients, and reduce preventable vision loss in people with diabetes.
