REVEAL++ Improves Alzheimer's Risk Prediction Using Differentiable Retinal Phenotypic Grouping

The retina offers a unique, non-invasive window into neurodegenerative diseases, revealing subtle structural patterns linked to future cognitive decline, such as Alzheimer's disease (AD). Previous vision-language frameworks, like REVEAL, have demonstrated that combining retinal fundus images with structured clinical risk narratives can enhance early AD prediction. A critical aspect of these methods involves phenotypic grouping, where individuals with similar risk profiles are treated as multi-positive pairs during contrastive learning. However, existing phenotypic grouping methods typically rely on discrete, hard assignments to fixed clusters, which can impose rigid supervision and separate the process of group formation from the actual representation learning. This research proposes REVEAL++, a novel approach that formulates phenotypic structure as a continuous concept within contrastive learning. Instead of assigning samples to static groups, REVEAL++ models inter-subject similarity using a differentiable weighting function. This function is derived from intra-modality embedding similarities in both retinal images and risk profiles. These continuous weights define "soft" multi-positive relationships through a continuous aggregation operator, allowing for graded supervision that better reflects the spectrum nature of disease risk. The framework also introduces a soft-target contrastive objective that simultaneously learns cross-modal alignment and phenotypic structure in an end-to-end manner. Evaluations on UK Biobank retinal imaging data for incident AD prediction show that REVEAL++ consistently outperforms discrete group-based contrastive learning and standard vision-language baselines, providing a more principled and robust foundation for population-scale neurodegenerative risk modeling.