🧬 biology

Investigating the fundamental characteristics of retinal age models

This study reveals that the association between retinal age gap and systemic diseases is inconsistent across age subgroups due to regression-to-the-mean effects and that model generalisability should be evaluated based on application-specific utility rather than age estimation accuracy alone.

Original authors: Yukun Zhou, Yiqun Lin, Ariel Ong, Matthew Wong, Yilan Wu, Wenyi Hu, Shubhank Verma, Broder Poschkamp, Lie Ju, Akshay Narayan, Sophie Martin, Maitrei Kohli, Fares Antaki, Alexander Heatley, Meng Wang

Published 2026-07-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yukun Zhou, Yiqun Lin, Ariel Ong, Matthew Wong, Yilan Wu, Wenyi Hu, Shubhank Verma, Broder Poschkamp, Lie Ju, Akshay Narayan, Sophie Martin, Maitrei Kohli, Fares Antaki, Alexander Heatley, Meng Wang, Ke Zou, Zheyuan Wang, Eden Ruffell, Dominic Williamson, Justin Engelmann, Hyunmin Kim, Takahiro Ninomiya, Rahul Jonas, Yansong Liu, Zijie Cheng, Hanyuan Zhang, Hejie Cui, Hualiang Wang, Zuozhu Liu, Bin Pu, Chubo Liu, Kenli Li, Jiayang Xu, Xinpeng Ding, Guokai Zhang, Huazhu Fu, Zongyuan Ge, Xujia Liu, Mohammad Eslami, Milen Raytchev, Tobias Elze, Michael Morley, Neil Oxtoby, Daniel Alexander, Anthony Khawaja, Ching-Yu Cheng, Lisa Zhu, Yih-Chung Tham, James Cole, Carol Cheung, Pearse Santos, Siegfried Wagner

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your eyes are like a high-tech security camera that doesn't just take pictures of what's in front of it, but also secretly records a "biological timestamp" of how fast your body is aging. Scientists call this Retinal Age. If your eyes look older than your actual birthday, it's called a Retinal Age Gap (RAG). For a long time, researchers thought a big gap meant "uh-oh, your body is aging too fast and you might get sick."

But this new study, led by a massive team of scientists from places like University College London and Harvard, decided to put that idea under a microscope. They looked at over 327,764 retinal images from 60,865 people to see if the story was really that simple.

Here is the twist they found: The "older eyes = sicker body" rule doesn't work the same way for everyone.

The "Gravity" Problem

The researchers discovered that the computer models used to guess retinal age have a weird habit called Regression-to-the-Mean (RTM). Think of it like a bouncy ball that always wants to land in the middle of the room.

If you are a young person (say, 20 years old) and your eyes look a bit worn out, the model might guess you are 25. That's a positive gap, and it correctly signals "hey, something is wrong." But if you are an elderly person (say, 80 years old) with very sick eyes, the model gets confused. Because it's trained to guess the "average" age, it pulls its guess down toward the middle. It might guess an 80-year-old with terrible eyes is only 75.

This creates a negative gap (the model thinks you are younger than you are), even though you are actually very sick. So, in older groups, a "negative gap" actually meant "sick," which is the exact opposite of what the model usually means! The study found that this "gravity" effect was much stronger in people with diseases like heart issues or eye problems than in healthy people.

The Age Trap

Because of this "gravity," the link between the age gap and sickness changes depending on how old you are.

In young and middle-aged people: A higher gap (older eyes) strongly suggests poor health.
In the elderly: That link gets weaker, and sometimes flips completely. The study showed that for people over 90, the usual warning sign of a high age gap basically disappears or even reverses.

The team tried to fix this by "calibrating" the model (like adjusting a scale) or by feeding it more balanced data, but the weird "gravity" effect kept happening. It turns out this isn't a bug you can just patch; it's a fundamental quirk of how these models work.

The Solution: Don't Guess, Ask

So, how do we fix this? The researchers suggest that instead of just looking at the age gap alone, we need to ask the model: "How old is this person, and what is their age gap?"

When they added this "interaction" (combining age and the gap) into their disease prediction math, it worked better.

For young and middle-aged people, adding the age gap improved the ability to spot sickness.
For older people, adding the age gap without the interaction actually made things worse (it started giving false alarms). But when they included the interaction, the model stopped making those mistakes and stayed accurate.

Does it work on other people?

The team tested their model on three other groups of people from the UK and Brazil, using different cameras.

The Accuracy: The model was pretty good at guessing ages in the UK group (error of about 3.80 years) but struggled more with the Brazilian groups (errors of 7.34 and 8.40 years).
The Surprise: Even though the model was bad at guessing the exact age in the Brazilian groups, the "age gap" still managed to spot sick people pretty well!

This suggests that you don't need a perfect age-guessing robot to find disease. Sometimes, even a "clumsy" robot can spot the signs of illness, as long as you know how to read its results for that specific group of people.

The Bottom Line

The paper doesn't say retinal age is useless. In fact, it confirms that your eyes do hold secrets about your health. However, it argues that we can't just use a single "one-size-fits-all" rule.

If you are young, a high age gap is a red flag. If you are very old, that same red flag might be a false alarm caused by the model's "gravity." To use this tool in the real world, doctors and scientists need to be careful: they must check if the model works for their specific group of patients and their specific goal (like guessing age vs. spotting disease).

The study suggests that before we start using these tools in hospitals, we need to stop assuming the rules are the same for a 20-year-old and an 80-year-old. It's a reminder that biology is messy, and our computers need to learn to be a little more flexible.

Technical Summary: Investigating the Fundamental Characteristics of Retinal Age Models

Problem Statement
Retinal age, estimated from fundus images, serves as a biological ageing marker where the deviation from chronological age, known as the Retinal Age Gap (RAG), is widely used to predict adverse health outcomes. However, fundamental characteristics of retinal age models critical for interpreting RAG remain underexplored. Specifically, the consistency of RAG-disease associations across different age subgroups and the generalisability of these models across diverse cohorts and imaging devices are not fully understood. A key concern is the "regression-to-the-mean" (RTM) effect, a statistical phenomenon in regression models where predictions are systematically pulled toward the mean of the training population. This can lead to overestimation of age in young individuals and underestimation in older individuals. The authors hypothesize that differential RTM effects between healthy and unhealthy populations may cause inconsistent associations between RAG and health status across age groups, potentially leading to misinterpretation in clinical applications.

Methodology
The study utilized a large-scale dataset comprising 327,764 retinal images from the AlzEye cohort (internal) and three external cohorts (UK Biobank, BRSET, and mBRSET) with distinct populations and imaging devices.

Model Architecture: The retinal age model was developed using a Vision Transformer (ViT) backbone, specifically DINOv3-Large, pre-trained on the LVD142M dataset. The architecture processes 224×224 images, averages patch tokens to form a 1,024-dimensional feature vector, and passes them through a custom regression head (linear layer, ReLU, dropout, final linear output) to estimate chronological age.
Training and Evaluation: The model was trained on the AlzEye dataset. Performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Pearson's $r^2$ .
Association Analysis: RAG was defined as estimated retinal age minus chronological age. Logistic regression was employed to analyze the association between RAG and disease status (unhealthy vs. healthy, and specific disease subgroups like cardiovascular, chronic systemic, and ocular diseases), adjusted for chronological age.
Investigating RTM: The authors explicitly tested for differential RTM by modeling the interaction between age and health status. They explored whether the observed inconsistencies could be mitigated by:
1. Linear and quadratic age calibration of RAG based on the healthy subset.
2. Retraining the model on a balanced age distribution (uniform sampling across age bins).
Downstream Classification: To assess the utility of RAG in disease detection, three logistic regression models were compared: (1) Age only, (2) Age + RAG, and (3) Age + RAG + Age × RAG interaction. Performance was measured via Area Under the Curve (AUC) and threshold-based metrics (sensitivity, specificity, F1 score).
External Validation: The model was applied to external datasets without retraining ("off-the-shelf") and with external calibration (refitting the logistic regression coefficients and thresholds on external training splits) to evaluate generalisability across different metrics: age estimation accuracy, disease association strength, and classification performance.

Key Results

Internal Performance: The model achieved an MAE of 5.31 years and a Pearson's $r^2$ of 0.745 on the internal test set. A positive association was confirmed between higher RAG and various diseases (cardiovascular, chronic systemic, ocular) in the overall cohort.
Age-Dependent Inconsistency: A critical finding was the inconsistency of RAG-disease associations across age subgroups. While positive associations (higher RAG = higher disease risk) were observed in young and middle-aged groups, these associations attenuated and even reversed (negative association) in the oldest subgroup (>80 years).
Differential RTM: This inconsistency was attributed to differential RTM. The slope of RAG decline with age was significantly steeper in the unhealthy group (-0.314) compared to the healthy group (-0.184). This differential effect persisted even after applying linear or quadratic age calibration and even when the model was retrained on a balanced age distribution, suggesting it is a fundamental characteristic of the model rather than a simple artifact of data distribution or calibration.
Value of Interaction Terms: Incorporating an Age × RAG interaction term into disease classification models improved performance, particularly in older age groups where the simple addition of RAG to an Age-only model actually degraded performance (AUC dropped from 0.585 to 0.494). The interaction model maintained performance gains in younger groups and mitigated the degradation in older groups.
External Generalisability: External evaluation revealed that model performance varied significantly by cohort and application.
- Age Estimation: The model performed well in the UK Biobank (MAE 3.80) but showed larger errors in BRSET (MAE 7.34) and mBRSET (MAE 8.40).
- Disease Association: Despite large estimation errors in BRSET and mBRSET, significant positive associations between RAG and unhealthy conditions were preserved across much of the age range.
- Classification Utility: The added value of RAG for disease classification was not uniform. In BRSET, RAG provided significant added value despite estimation errors. In mBRSET, significant disease associations did not translate to significant improvements in classification performance over chronological age alone.

Significance and Claims
The paper claims that the interpretation of Retinal Age Gap (RAG) is more complex than previously assumed due to differential RTM effects between healthy and unhealthy populations. The authors assert that:

Age-Dependent Associations: RAG-disease associations are not constant across the lifespan; they vary significantly by age subgroup. Reporting a single overall association may be misleading, particularly for older populations where the relationship may reverse.
Limitations of Calibration: Standard age calibration techniques and balancing training data distributions do not eliminate differential RTM. This suggests that RTM is an intrinsic property of how these models learn from data containing pathological variations.
Application-Specific Evaluation: The generalisability of retinal age models cannot be judged by a single metric (e.g., age estimation accuracy). A model with poor age estimation accuracy may still retain strong disease associations and utility for specific clinical tasks, while a model with high accuracy may not necessarily improve downstream classification. Therefore, external evaluation must be multi-dimensional and tailored to the intended application (e.g., age estimation vs. risk stratification).
Modelling Recommendations: For reliable clinical application, models should explicitly account for the interaction between chronological age and RAG to avoid misleading risk interpretations in older subgroups.

The study concludes that while RAG contains valuable health-related information, its clinical utility requires careful, age-stratified interpretation and application-specific validation frameworks. These findings are suggested to extend to broader biological age estimation models beyond retinal imaging.