This poster presents LongRETSurv, a Transformer-based longitudinal survival model that predicts future Alzheimer’s disease diagnosis from retinal fundus images. Using the large-scale SCONe dataset and RETFound-Green image embeddings, the model analyses retinal changes over time to estimate individual risk trajectories. LongRETSurv achieved strong long-term predictive performance (AUC up to 0.89) and outperformed single-visit approaches. This work highlights the potential of non-invasive retinal imaging and artificial intelligence for early dementia risk assessment and personalised screening.

Keywords: Alzheimer's disease, dementia, retinal imaging, fundus photography, survival analysis, longitudinal data, Transformer, RETFound, artificial intelligence, risk prediction, biomarkers, SCONe.