REVEAL++: Differentiable Phenotypic Grouping for Vision-Language Retinal Modeling of Alzheimer's Disease Risk

The article introduces REVEAL++, a novel framework for differentiable phenotypic grouping in vision-language models aimed at predicting Alzheimer's disease risk from retinal images and clinical narratives. This approach replaces discrete group assignments with a continuous formulation of phenotypic similarity, utilizing a differentiable weighting function to enhance contrastive learning through soft multi-positive relationships. Evaluated on UK Biobank data, REVEAL++ outperforms traditional discrete methods, offering a more flexible and effective strategy for modeling neurodegenerative disease risk at scale.