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

Technical Summary: REVEAL++

Problem Statement
Alzheimer's disease (AD) involves a long preclinical phase where pathological changes accumulate before clinical symptoms appear. While the retina offers a noninvasive window into neurodegenerative risk, existing Vision–Language (V-L) modeling approaches, such as the REVEAL framework, rely on Group-Aware Contrastive Learning (GACL) to improve prediction. A critical limitation of current GACL methods is their reliance on discrete phenotypic grouping. These approaches assign individuals to fixed clusters based on hard similarity thresholds, treating subjects within a group as multi-positive pairs. This design imposes rigid supervision and decouples group formation from representation learning. Biologically, however, neurodegenerative risk evolves along continuous, overlapping trajectories shaped by heterogeneous factors. Hard group assignments create artificial boundaries that fail to reflect the graded, spectrum-like nature of disease vulnerability and prevent the grouping mechanism from adapting during training.

Methodology: REVEAL++
The authors propose REVEAL++, a framework that replaces discrete grouping with a differentiable phenotypic weighting mechanism within a multi-modal contrastive learning setup. The architecture consists of the following components:

1. Multi-Modal Encoders:

   • Image Encoder: Utilizes RETFound to process retinal fundus images.
   • Text Encoder: Utilizes GatorTron to process structured clinical risk narratives.
   • Clinical Report Generation: Structured questionnaire data (demographics, lifestyle, cognitive variables) from the UK Biobank are converted into synthetic clinical narratives using LLaMA-3.1. This ensures compatibility with pretrained text encoders while maintaining numerical fidelity by mapping tabular entries directly to template fields without imputation.
   • Projection: Both modalities are projected into a shared latent space ($d=1024$) and $\ell_2$-normalized.

2. Differentiable Phenotypic Weighting:
   Instead of hard clustering, the model computes intra-modality similarity matrices ($S_{ii}$ for images, $S_{tt}$ for text) using cosine similarity between normalized embeddings.

   • These similarities are transformed into soft membership signals using a sigmoid gating function with learnable thresholds ($\tau$) and sharpness parameters ($g$).
   • The image and text signals are combined via a differentiable probabilistic union operator to produce a continuous phenotypic weighting score $W_{pq} \in [0, 1]$ for every pair of subjects $(p, q)$.
   • $W_{pq}$ represents the degree of phenotypic alignment, serving as a soft target rather than a binary label.

3. Soft-Target Contrastive Objective:
   The framework optimizes a soft-target multi-positive contrastive loss. The loss function incorporates $W_{pq}$ to modulate the supervision strength:

   • When $W_{pq} \to 1$, the pair is treated as a strong positive match.
   • When $W_{pq} \to 0$, the pair is treated as a negative.
   • Intermediate values allow for graded supervision, enabling the model to learn cross-modal alignment and phenotypic structure jointly in an end-to-end manner.

Key Contributions

1. Differentiable Phenotypic Weighting: The paper introduces a continuous formulation of phenotypic structure, replacing hard threshold-based grouping with similarity weights derived directly from retinal and clinical embeddings. This enables smooth, data-driven cohort modeling.
2. Soft Multi-Positive Contrastive Learning: A new objective function is proposed that incorporates phenotypic similarity into cross-modal alignment, facilitating graded multi-positive supervision instead of binary pair assignments.
3. State-of-the-Art Performance: The framework achieves new state-of-the-art results for incident Alzheimer's disease prediction using UK Biobank retinal imaging data, outperforming existing V-L baselines and discrete group-based contrastive methods.

Experimental Results
The framework was evaluated on the UK Biobank dataset (approx. 39k subjects) for predicting incident AD.

• Baselines: The study compared REVEAL++ against standard SVMs, image-only foundation models (RETFound), general biomedical V-L models (BiomedCLIP, PMC-CLIP), and retinal-specific V-L models (RETCLIP, MM-Retinal). It also included a baseline REVEAL model without Group-Aware Contrastive Learning (GACL) and the original REVEAL with discrete GACL.
• Performance: REVEAL++ achieved the highest performance across all metrics:
  • AUROC: 0.678 (vs. 0.658 for REVEAL with discrete GACL and 0.642 for RETFound+GatorTron).
  • Balanced Accuracy: 0.613.
  • F1-Score: 0.236.
  • MCC: 0.168.
• Analysis: The results indicate that soft, differentiable phenotypic alignment leads to more coherent multi-modal representations. The model benefits from allowing subjects with similar risk profiles to contribute proportionally during training, rather than relying on rigid group assignments.

Significance and Claims
The paper claims that treating phenotypic similarity as a learnable, continuous signal rather than a fixed grouping rule provides a principled foundation for population-scale neurodegenerative risk modeling. By modeling relationships as a differentiable, attention-like process, REVEAL++ captures the continuous and heterogeneous biological variability underlying AD risk more faithfully than discrete methods. The authors argue that this approach better reflects the pathophysiology of preclinical AD, where risk accumulates gradually and manifests heterogeneously, allowing the model to represent intermediate phenotypic states that correspond to early pathological changes. This offers a robust strategy for modeling structured variability in multi-modal biomedical data, with potential applications in chronic disease risk prediction and precision medicine.