Why REVEAL++'s 'Continuous Learning' Could Be the Key to Predicting Alzheimer's from a Simple Eye Exam

Traditional vision-language models for Alzheimer's prediction from retinal images rely on discrete phenotypic groupings—hard clusters that assign patients to fixed risk categories. This rigid approach decouples group formation from representation learning and fails to capture the continuous nature of disease risk. REVEAL++ replaces these hard assignments with a differentiable weighting function that measures inter-subject similarity directly from intra-modality embeddings of both retinal images and clinical risk narratives. These weights define soft multi-positive relationships, allowing the model to learn graded supervision that reflects the full spectrum of cognitive decline.

Evaluated on the UK Biobank dataset for incident Alzheimer's prediction, REVEAL++ consistently outperforms both discrete group-based contrastive learning and standard vision-language baselines. By treating phenotypic similarity as a learnable continuous signal, the framework provides a principled and robust foundation for population-scale neurodegenerative risk modeling. Accepted at MICCAI 2026, this work points to a future where routine retinal scans could serve as a noninvasive, scalable screening tool for Alzheimer's disease, enabling earlier detection and intervention.