👁️ ophthalmology

RetFit: A Novel Deep Learning-Based Biomarker of Cardiorespiratory Fitness Derived From the Retina

This study introduces RetFit, a novel deep learning-based biomarker that estimates cardiorespiratory fitness from retinal fundus images, demonstrating strong prognostic value for cardiovascular events and mortality while offering a scalable, non-invasive alternative to traditional exercise testing.

Original authors: Quintas, I., Bontempi, D., Bors, S., Trofimova, O., Boettger, L., Iuliani, I., Ortin Vela, S., Liefers, B., Vargas-Quiros, J., Klaver, C. C. W., Meloni, I., Elwakil, A., Bergin, C., Tomasoni, M., Berg

Published 2026-01-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Quintas, I., Bontempi, D., Bors, S., Trofimova, O., Boettger, L., Iuliani, I., Ortin Vela, S., Liefers, B., Vargas-Quiros, J., Klaver, C. C. W., Meloni, I., Elwakil, A., Bergin, C., Tomasoni, M., Bergmann, S., Presby, D. M.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

The Big Idea: A Fitness Score from a Snapshot

Imagine you want to know how fit a person is. Usually, you'd ask them to run on a treadmill until they are out of breath while wearing a mask to measure their oxygen. This is the "gold standard," but it's expensive, takes a lot of time, and requires special equipment and a coach.

This paper introduces a new tool called RetFit. Think of RetFit as a "fitness detective" that looks at a simple, non-invasive photo of the back of your eye (a retinal image) and uses a super-smart computer brain (Artificial Intelligence) to guess your cardiorespiratory fitness level.

The authors call this a "biomarker," which is just a fancy word for a biological clue. They are saying: Your eyes hold a secret map of your heart and lung health, and we found a way to read it.

How They Built the "Eye-Scanner"

The researchers used a massive database called the UK Biobank, which contains health data from over 57,000 people.

The Training: They took photos of these people's eyes and paired them with their actual fitness scores (which were estimated using a sub-maximal exercise test, a slightly easier version of the treadmill test).
The AI: They taught a state-of-the-art AI model (a "Vision Transformer") to look at the eye photos and learn the patterns that match high or low fitness.
The Result: The AI created a new score called RetFit. It's not a perfect copy of the exercise test (they only matched about 20% of the time), but it captures a unique and powerful signal.

What RetFit Tells Us (The "Crystal Ball" Effect)

The most exciting part of the paper is what RetFit predicts about the future. The researchers tested if RetFit could predict who would get sick or pass away, and it worked surprisingly well.

The Heart Attack Predictor: If you have a low RetFit score, you are at a higher risk of having a heart attack or stroke in the future. The paper found that RetFit was actually better at predicting these events than the traditional exercise test score, even after accounting for other risks like high blood pressure or smoking.
The Longevity Predictor: A low RetFit score was also linked to a higher risk of death from any cause.
The "Extra" Information: The paper claims that RetFit and the exercise test score (SETCRF) are looking at the body through different lenses. They are only moderately related (like two friends who know some of the same things but have different secrets). RetFit seems to pick up on "subclinical" vascular health—meaning it sees tiny changes in your blood vessels that standard tests miss.

The "Genetic Mystery"

To understand why RetFit works, the researchers looked at the DNA of the participants.

Different Genes: They found that the genes linked to RetFit were completely different from the genes linked to the exercise test.
The Analogy: Imagine the exercise test is like checking the engine of a car (how the heart muscle works). RetFit, however, seems to be checking the road the car drives on (the blood vessels and the nervous system).
Brain Connection: Interestingly, RetFit was linked to genes involved in brain development and vision, suggesting it might be a window into how the nervous system and blood vessels work together, not just the heart.

The "Map" the AI Looked At

How did the AI know what to look for in the eye photo?

The Attention Map: The researchers asked the AI to show them where it was "looking" on the eye photo. The AI focused heavily on the blood vessels in the retina.
The Key Feature: The single most important feature the AI used was the number of times the arteries branch out (bifurcations). Think of it like a tree: a healthy, complex tree with many branches suggests a healthy, efficient system. A sparse tree suggests a system that is struggling.

Testing in Other Towns (Validation)

To make sure this wasn't just a fluke specific to the UK, the researchers tested their AI on two other groups of people in the Netherlands and Switzerland.

The Result: Even though these new groups didn't have exercise test data, the RetFit scores still successfully predicted heart disease and death. This proves the tool works in different populations, not just the one it was trained on.

The Bottom Line

The paper concludes that RetFit is a powerful, scalable, and accessible way to estimate fitness and health risk.

Why it matters: It turns a routine eye exam (which many people already have) into a powerful health checkup.
The Limitation: The authors are careful to say this is a new tool, not a replacement for exercise. They emphasize that RetFit captures a specific type of health information (vascular and nervous system health) that is distinct from, but complementary to, traditional fitness testing.

In short: Your eyes are a window to your heart, and this new AI tool is learning to read that window to tell us how healthy we really are.

Technical Summary: RetFit – A Deep Learning-Based Biomarker of Cardiorespiratory Fitness Derived from the Retina

Problem Statement
Cardiorespiratory fitness (CRF) is a robust, independent predictor of cardiovascular events and all-cause mortality, often outperforming traditional risk factors. However, its clinical utility is limited by the resource-intensive nature of gold-standard measurement methods, such as maximal cardiopulmonary exercise testing (CPET), which require specialized equipment, trained personnel, and maximal physical effort from subjects. While submaximal exercise tests (SETCRF) and non-exercise estimates offer alternatives, there is a need for scalable, non-invasive proxies that capture vascular health. Retinal Color Fundus Images (CFIs) provide a non-invasive window into systemic vascular health and have shown promise in predicting cardiovascular risk factors, yet their potential to estimate CRF remains unexplored.

Methodology
The study introduces RetFit, a novel CRF estimator derived from CFIs using state-of-the-art deep learning. The methodology involved three primary phases:

Model Development and Training:
- Data Source: The authors utilized the UK Biobank (UKBB) cohort, selecting 57,588 genotyped participants who underwent submaximal exercise testing (providing SETCRF values) and had high-quality CFIs. This yielded 102,826 images.
- Architecture: A Vision Transformer (ViT) model, specifically the pre-trained RETFound foundation model, was fine-tuned to predict SETCRF from CFIs.
- Training Strategy: The model was trained over 50 epochs with a warm-up phase followed by cosine annealing. A 5-fold cross-validation strategy with a leave-one-out design was employed. Separate models were fine-tuned for left and right eyes, and predictions were averaged for subjects with bilateral images.
- Quality Control: A CNN-based quality assessment, trained on human graders, was used to select the top 75% of images.
Validation and Benchmarking:
- Internal Benchmarking: RetFit was benchmarked against the ground-truth SETCRF within the UKBB.
- External Validation: The model's generalizability was tested in two independent cohorts where CRF was not measured: the Rotterdam Study (N=8,885) and the CoLaus Study (N=2,260).
- Associations: Regression analyses (linear and logistic) were performed to assess associations between RetFit and 21 binary disease outcomes and 15 continuous risk factors, adjusting for sex, age, BMI, and socioeconomic status.
- Prognostic Analysis: Cox Proportional Hazard models were used to evaluate RetFit's ability to predict cardiovascular events and all-cause mortality, comparing its performance to SETCRF in unadjusted and fully adjusted models (including Framingham-like risk factors).
Genetic and Interpretability Analysis:
- Genetics: Genome-Wide Association Studies (GWAS) were conducted for both RetFit and SETCRF using REGENIE. Gene and pathway enrichment analyses were performed using PascalX, and heritability was estimated using GCTA.
- Interpretability: Attention maps were generated from the Vision Transformer to identify image regions driving predictions. Additionally, 18 quantitative retinal vascular features (Tangible Image Features or TIFs) were extracted using the VascX pipeline to assess their correlation with RetFit and SETCRF.

Key Results

Prediction Performance: RetFit demonstrated a moderate positive correlation with SETCRF ( $r=0.45$ , $R^2=0.20$ ), indicating it captures a distinct but related signal.
Disease and Risk Factor Associations: In the UKBB, RetFit was significantly associated with 20 of 21 binary health outcomes and all 15 continuous risk factors. In many cases, RetFit provided unique information not captured by SETCRF (e.g., cataract, atherosclerosis, prospective memory). These associations were largely reproduced in the Rotterdam and CoLaus cohorts.
Prognostic Power:
- Cardiovascular Events: RetFit was prognostic for cardiovascular events in the UKBB (HR=0.936 per SD increase, adjusted) and the Rotterdam Study (HR=0.813, adjusted). Notably, in fully adjusted models, RetFit remained a significant predictor while SETCRF did not ( $p=0.065$ ). Adding RetFit to a model containing SETCRF significantly improved model fit, whereas the reverse was not true.
- Mortality: Both RetFit and SETCRF predicted overall mortality, though RetFit's performance degraded more significantly after full adjustment in the UKBB.
Genetic Architecture: GWAS revealed that RetFit and SETCRF have disjoint genetic architectures. RetFit was associated with 41 genes (enriched in eye abnormalities, cell development, and nervous system pathways), while SETCRF was associated with 25 genes (enriched in cardiovascular pathways). The genetic correlation between the two was low ( $r_g=0.09$ ).
Interpretability: Attention maps highlighted the superior vascular beds. Among vascular features, the number of arterial bifurcations showed the strongest association with RetFit ( $\beta=0.287$ ), whereas vascular density ratio was most strongly associated with SETCRF.

Significance and Claims
The paper claims that RetFit represents a novel, scalable, and accessible biomarker for estimating CRF and assessing systemic health. Its primary significance lies in:

Scalability: It offers a method to estimate CRF in large populations where exercise testing is unavailable, utilizing routine retinal imaging.
Distinct Biological Insight: The study posits that RetFit is not merely a proxy for exercise performance but captures a unique dimension of systemic health related to retinal microvasculature and nervous system development, evidenced by its distinct genetic architecture and superior prognostic power for cardiovascular events in adjusted models.
Clinical Utility: By demonstrating consistent associations with disease states and strong prognostic value in external cohorts without CRF measurements, RetFit supports the use of retinal imaging for risk stratification in both clinical and public health contexts, particularly for frail or elderly populations who cannot undergo maximal exercise testing.

The authors conclude that while RetFit does not fully replicate the physiological measurement of VO2max, it serves as a powerful, image-derived indicator of vascular health and disease risk that complements traditional fitness metrics.