A Disease-specific Retinal Aging Biomarker for Predicting Progression to Late Age-related Macular Degeneration

Hexin Dong, Benjamin Hou, Puzhen Wu, Mingquan Lin, Emily Y. Chew, Zhiyong Lu, Yifan PengarXiv 2026 · arXiv

TL;DR

Researchers have developed Fundus-based AMD Age (FAA), an AI-derived biomarker that estimates a patient's "retinal age" specifically in the context of Age-related Macular Degeneration (AMD). By integrating this disease-specific aging clock into standard clinical models, they significantly improved the accuracy of predicting which patients will progress to late-stage vision loss compared to using chronological age alone.

What Problem Does This Paper Address?

Age-related macular degeneration (AMD) is a primary cause of blindness in older adults. While chronological age is the strongest risk factor, it is a blunt instrument: two 70-year-olds can have vastly different retinal health profiles.

Current AI models for AMD often act as "black boxes," providing a risk score without clinical context, or they are "biological clocks" trained only on healthy eyes. These healthy-eye models often fail to capture the accelerated, disease-specific aging processes occurring in AMD. The authors sought to create a biomarker that is both interpretable—expressed as a "retinal age"—and specifically tuned to the structural changes associated with AMD progression.

Methods

The core of the study is the development of the FAA estimation model.

  1. Architecture: The team used VMamba, a state-space model backbone, which is adept at capturing both local details (like tiny drusen) and long-range spatial dependencies in fundus images.
  2. Novel Loss Function: Instead of just training the model to predict chronological age, they introduced a ranking-based loss function. This forced the model to ensure that eyes with higher clinical AMD severity scores were assigned a higher FAA than eyes with lower severity, even if their chronological ages were the same.
  3. Survival Analysis: The predicted FAA was then plugged into a Cox proportional-hazards model. This classical statistical framework allowed the researchers to combine the AI biomarker with known clinical factors like smoking status and specific retinal phenotypes (e.g., drusen size, pigmentary abnormalities).
Figure 1: Overview of the study workflow for predicting late AMD progression. a-b, Flowchart de- tailing participant selection, data preprocessing, and exclusion criteria for building the AMD risk prediction models based on the AREDS and AREDS2 datasets. c, The workflow for predicting late AMD progression. FAA, as predicted by the VMamba model [30] from CFP, is combined with smoking status and clinically relevant AMD phenotypes (e.g., drusen size and pigmentary abnormalities), and the resulting model is integrated into a Cox proportional hazards model to predict the risk of AMD progression.

Datasets

The study utilized two landmark longitudinal datasets:

  • AREDS (Age-Related Eye Disease Study): Used for training and internal testing (36,782 images from 2,192 participants).
  • AREDS2: Used as an independent external validation cohort to ensure the model generalizes across different populations and grading standards.

Key Insights and Ideas

  • Disease-Specific vs. General Aging: The "aha moment" here is that a biological clock trained on healthy eyes isn't enough for prognosis. By using a ranking loss guided by AMD severity, the FAA captures "accelerated aging" that is unique to the pathology of the disease.
  • Beyond Manual Grading: While FAA is guided by severity scores during training, it actually outperforms models that use only those manual scores. This suggests the AI is picking up on subtle structural "aging signals" in the images that human graders might miss or that aren't captured by current grading scales.
  • Interoperability: By outputting an "age," the model provides a metric that is immediately intuitive to clinicians and patients, bridging the gap between complex deep learning and clinical communication.

Key Results

The FAA-based models showed consistent, significant improvements over models using chronological age across all metrics:

  • Predictive Accuracy: In the AREDS cohort, FAA improved the C-index (a measure of how well the model ranks risk) from 0.836 to 0.888.
  • Error Reduction: The 5-year Brier score (measuring prediction accuracy) dropped from 0.050 to 0.046.
  • External Validation: The gains held up in the AREDS2 dataset, with the C-index rising from 0.715 to 0.756.
  • Risk Stratification: Kaplan-Meier survival curves demonstrated that FAA could much more effectively separate patients into "high-risk" and "low-risk" groups than chronological age.
Figure 2: Comparison of prognostic performance between models using age (chronological age or FAA) and phenotypes graded by reading centers in AREDS and AREDS2. a, AREDS. b, AREDS2. P-values were calculated using two-sided bootstrap resampling (5,000 iterations). Statistical significance was indicated as *** for 𝑝< 0.001.

Limitations and Future Work

Despite the strong results, the authors note several areas for improvement:

  • Population Diversity: The AREDS cohorts are predominantly White. Future work must validate FAA in more racially and ethnically diverse populations to ensure the biomarker is "transportable."
  • Subtype Analysis: The current model predicts progression to "late AMD" generally, without distinguishing between the two main types: geographic atrophy (dry) and neovascular AMD (wet).
  • Biological Grounding: While FAA is a powerful imaging biomarker, it hasn't yet been correlated with molecular aging markers (like epigenetic clocks or blood-based biomarkers).
  • Interpretability: While the authors used FullGrad visualizations to show the model focuses on the macula and drusen, deep learning models remain somewhat "black boxes" regarding the exact features they prioritize.

Summary updated

Abstract

Age-related macular degeneration (AMD) is a leading cause of irreversible vision loss in older adults. While chronological age is a primary risk factor, it inadequately reflects individual variations in retinal aging, limiting effective AMD risk stratification and timely intervention. Here, we propose and validate fundus-based AMD age (FAA), which quantifies disease-specific retinal aging in AMD and captures accelerated AMD-related retinal aging, thereby improving prognostic precision over chronological age. Trained on 36,782 images from 2,192 participants in the Age-Related Eye Disease Study (AREDS) dataset and validated across AREDS and AREDS2 cohorts, our model demonstrates significant improvements over traditional chronological age-based approaches. Specifically, integrating the proposed FAA into Cox proportional-hazards models significantly enhances prediction performance. FAA improved the C-index from 0.836 to 0.888 and reduced the 5-year Brier score from 0.050 to 0.046 in AREDS, with consistent gains observed in AREDS2 (C-index 0.715 to 0.756; 2-year Brier 0.078 to 0.076). FAA showed improved prognostic performance across the two datasets and under different phenotype definitions (reading center-defined vs. specialist-defined grading), supporting its potential utility for AMD risk stratification. These findings highlight the advantage of incorporating FAA into AMD risk prediction frameworks, supporting further evaluation of FAA for AMD risk stratification.
BibTeX
@misc{dong2026disease,
  author = {Hexin Dong and Benjamin Hou and Puzhen Wu and Mingquan Lin and Emily Y. Chew and Zhiyong Lu and Yifan Peng},
  title = {{A Disease-specific Retinal Aging Biomarker for Predicting Progression to Late Age-related Macular Degeneration}},
  year = {2026},
  eprint = {2611.16955},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.16955},
}
Figure