Deep networks trained on photographs of the back of the eye now predict stroke well. One system reports an AUROC of 0.901 for incident stroke, and its retina-based score detects silent brain infarcts far better than clinical risk factors do. The retina has also been described in this literature as a window onto the brain's small blood vessels. There is an uncomfortable alternative reading of these results. The same photographs reveal a person's age, sex and blood pressure with high accuracy, and those are the strongest stroke risk factors there are. A model can score well by working as an expensive age-and-blood-pressure meter.
We propose a direct test of that worry. The test asks what a retinal foundation model knows about future stroke and brain white-matter damage once everything predictable from measured risk factors has been removed, and where in the image that leftover knowledge comes from. The method borrows the oldest tool in epidemiology, adjusting for confounders by "holding them constant." We adapt it to 1024-dimensional image embeddings, using flexible, cross-validated regression so that a nonlinear age signal cannot pass for an independent vascular one. We then split the surviving signal into a part that maps onto measurable vessel geometry and a part that does not. The output is a calibrated number, the stroke information a retinal scan adds beyond a blood-pressure cuff and a birth date, together with a map of which retinal features carry it.
Core Claim
Current evidence that retinal AI "sees" cerebrovascular disease is not adjusted for confounders in the sense an epidemiologist would accept. The fix is to change the analysis, not to build a cleverer debiasing network. The proposed analysis has two steps.
The first step is cross-fitted partialling-out of confounders from the embedding. In plain terms, we predict each embedding dimension from the confounders with a flexible model fit on other folds, then keep only the part the confounders cannot explain. We then ask how much predictive value that leftover part adds over an equally flexible confounder-only model.
The second step is a vessel-versus-rest decomposition of what survives. It uses decoders that recover vessel measurements from the embedding.
Our hypothesis has two parts. First, a substantial share of the reported stroke discrimination will disappear once the confounder model is flexible rather than linear. Second, whatever survives will sit mainly in a vascular-geometry subspace, not in background texture of the retina.
Formally, let $e$ be a retinal embedding, $Z$ the measured confounders (age, sex, systolic and diastolic blood pressure, smoking, diabetes and its duration, BMI, lipids, antihypertensive use), and $Y$ the outcome. With $\hat m(Z)$ a cross-fitted flexible estimate of $\mathbb{E}[e \mid Z]$, the residualized embedding is $\tilde e = e - \hat m(Z)$. The target quantity is
$$\Delta = C\big(\hat f(Z, \tilde e)\big) - C\big(\hat g(Z)\big),$$
where $C$ is out-of-fold discrimination (Harrell's C for time-to-stroke, AUROC for binary endpoints), $\hat g$ is the best confounder-only model we can build, and $\hat f$ adds the residualized embedding. Let $P_V$ project onto the subspace of $\tilde e$ that linearly decodes vessel measurements. We then refit with $P_V\tilde e$ alone and with $(I-P_V)\tilde e$ alone. This tells us whether the surviving signal is vascular or non-vascular.
$\Delta$ answers a question about prediction beyond measured risk factors, not a causal question. Blood pressure narrows retinal arterioles and also causes stroke. Holding it fixed therefore asks what the retina adds beyond the cuff, which is exactly the clinically relevant question raised by the "$30 cuff" objection. It does not estimate a causal effect of retinal geometry.
Why Existing Work Does Not Answer This
The motivating evidence comes from end-to-end retinal stroke models. DeepRETStroke pre-trains a ViT-Large on 895,640 fundus photographs and distills knowledge from an MRI-confirmed silent-infarct detector. It reports AUROC 0.901 for incident stroke and 0.797 for silent infarct detection, against 0.633 for clinical traits. However, it compares the image model against risk factors rather than conditioning the image model on them.
RETFound is a masked-autoencoder foundation model pre-trained on 1.6M retinal images. It predicts three-year heart failure with AUROC around 0.79 and makes no attempt to separate that signal from age.
The graph model of Huang et al. represents vessels as graphs and contrastively aligns them with clinical tables. It reaches 71.92% AUROC for incident stroke in UK Biobank. Its alignment objective deliberately pulls age and blood pressure information into the image representation, which is the opposite of adjustment. That makes it a natural object of our audit.
Evidence that the shortcut is real comes from two papers. Kim et al. show that retinal background tissue predicts age and sex even with the vessels digitally removed. RetinaGuard shows age is so strongly encoded that it can be read to about 2 years error, and adversarially hidden to over 9 years. RetinaGuard never asks whether stroke prediction survives the hiding.
The adjustment template comes from classical studies. Song et al.'s OCTA study in 246 older adults finds that lower arteriolar density relates to periventricular white-matter lesions (r = −0.17) after adjusting for age, sex, hypertension, diabetes, lipids and smoking. This establishes that some vascular signal survives adjustment, but only for two handcrafted metrics, with linear covariate terms, in one small cross-sectional cohort.
The cautionary tale comes from bibliometrics. Mishra et al. found that a 50% raw gender gap in self-citation shrinks to about 2% once career length is modeled. Our concern is that retinal stroke AUROCs may behave the same way once age is modeled properly.
Two tools supply the decomposition step. AutoMorph extracts interpretable vessel metrics: arteriolar and venular calibers (CRAE, CRVE), their ratio, fractal dimension, density and tortuosity. Feature Space Translation shows that frozen RETFound embeddings linearly or near-linearly decode these metrics, for example fractal dimension at R² = 87.5% on AI-READI. We reuse that decodability to define $P_V$. Neither paper links the decoded geometry to brain outcomes or adjusts for patient confounders.
Research Program
The primary cohort is UK Biobank. It has fundus photographs, linked hospital and death records for incident stroke, and full covariate panels. Huang et al. used 82,569 participants, which gives a published reference for split construction. A secondary endpoint is MRI-derived white-matter hyperintensity volume in the UK Biobank subset that also has brain imaging. The size of the overlap must be established at data access; we do not assume it. The UKBiobank_deep_pretrain brain-MRI networks are available if MRI-derived features beyond the released lesion volumes are needed. AI-READI's 1,067 paired multi-device participants with AutoMorph metrics and RETFound embeddings serve as the training set for the vessel decoders. This keeps $P_V$ independent of the outcome data.
We extract frozen RETFound embeddings and, if obtainable, DeepRETStroke embeddings. We also train an outcome-tuned score from RETFound to mirror how deployed models are built.
The confounder model $\hat g$ and the residualizer $\hat m$ are deliberately strong. We use gradient-boosted trees and spline-based Cox models with age-by-sex and age-by-blood-pressure interactions. Both are fit with five-fold cross-fitting, so no participant's residual is computed from a model that saw that participant.
$P_V$ is built from ridge decoders trained on AI-READI to predict each AutoMorph metric from the embedding. We take the span of the decoder weight vectors, kept separately for arteriolar metrics (CRAE, arteriolar fractal dimension and density) and venular metrics (CRVE, venular measures). This gives an artery-versus-vein split of the vascular subspace, which tests whether the compartment-specific pattern of Song et al. appears in fundus-photograph foundation features.
Before touching real outcomes, we validate the pipeline on semi-synthetic outcomes. These are generated from real covariates plus planted signals: one that is a nonlinear function of age only, and one that depends on decoded fractal dimension. The pipeline must report $\Delta \approx 0$ for the first and attribute the second to $P_V$.
Baselines
The naive reference point is the confounder-only model $\hat g$ on its own. This is the "cuff and birth date" predictor. It is fit with the same cross-fitting and the same hyperparameter search budget as every other model. Any retinal method must beat it, and its C-statistic is the floor against which $\Delta$ is measured. A linear-logistic version of it, matching the specification used in Song et al. and Zhang et al., is also reported. The gap between the linear and flexible confounder models is itself an informative result.
The strongest existing retinal stroke result on comparable data is the multimodal contrastive graph model of Huang et al.. It reports 71.92% AUROC for incident stroke in UK Biobank, versus 69.30% for its supervised multimodal baseline. We re-run it on our splits with its published configuration and then subject its embedding to the same partialling-out. DeepRETStroke's 0.901 internal AUROC (Jiang et al.) is on a different multi-country cohort and is not directly comparable. We cite it as the headline claim our audit interrogates and include its embedding only if the weights are obtainable. For a confounder-adjusted incremental AUROC of any retinal model on stroke, there is no reported result found on Lacuna. "Beating" the field therefore means producing the first credible estimate of $\Delta$ with a confidence interval that excludes zero, or showing that it does not.
The standard epidemiological competitor is the deep score added as a covariate. We take the unadjusted, outcome-tuned retinal risk score and enter it into a Cox model alongside $Z$. This is how incremental value is usually claimed, and it is the main threat to our proposal. We run it with both linear and flexible $Z$ terms on identical folds.
The interpretable competitor is AutoMorph metrics entered into the same adjusted regression. It tests whether a dozen transparent numbers carry everything the embedding carries after adjustment.
Adversarial debiasing in the style of RetinaGuard is a further competitor. A RETFound head is trained with a gradient-reversal adversary predicting $Z$, then evaluated for stroke.
The key ablation swaps flexible residualization for linear residualization while holding everything else fixed. If our central claim is right, linear residualization will overstate $\Delta$ because nonlinear age information leaks through.
Why not just…?
Why not just add the deep risk score to a Cox model with the usual risk factors? This is the serious alternative. For the question "is there any incremental value," it may give nearly the same answer as our method if the confounder terms are flexible enough. We will report it side by side, and if the two agree we will say so. Our method is then justified only by the decomposition step, which the score-as-covariate approach cannot do because it collapses the embedding to one number. The decisive comparison is whether the score-as-covariate estimate shrinks as its age and blood-pressure terms move from linear to spline-with-interactions. Large shrinkage would show that standard practice overstates retinal value. The semi-synthetic age-only outcome will show this directly if it occurs.
Why not just use AutoMorph metrics in a multivariable regression? This is transparent and cheap, and if it matches the embedding's adjusted $\Delta$, it should be the recommended tool. Two pieces of evidence suggest it will fall short. Kim et al. show that predictive content lives outside the segmented vessels. Rastogi et al. show that handcrafted geometry such as tortuosity drifts across cameras. Whether the information outside the vessels is stroke-relevant after adjustment, rather than just age-relevant, is exactly what the $(I-P_V)$ arm measures. A null there would vindicate AutoMorph.
Why not just strip confounders adversarially, as RetinaGuard strips age? Adversarial removal has two weaknesses for this question. It only removes what a given adversary can detect, so a stronger probe can often recover the hidden variable. It also removes any image feature correlated with blood pressure, including arteriolar narrowing that may carry real information beyond the cuff reading. We test both weaknesses. We train a fresh, larger probe to predict age from the adversarially cleaned embedding, and we compare its stroke $\Delta$ with ours. RetinaGuard's preserved diabetic-retinopathy accuracy of 0.916 concerns lesions that are not age-correlated, so it does not tell us how this approach behaves on a vascular endpoint.
Proposed Experiments
The headline experiment estimates $\Delta$ for incident stroke in UK Biobank for four representations: frozen RETFound embeddings, the outcome-tuned RETFound score, Huang et al.'s graph embedding, and AutoMorph metrics. Each is evaluated under linear and flexible confounder models, with bootstrap confidence intervals over participants. We also report the adjusted hazard ratio per standard deviation of the residualized score, and the stroke AUROC within strata of age band, sex and treated hypertension. The within-stratum AUROC is an assumption-light check that the score separates people of the same age, sex and treatment status.
The decomposition experiment refits with $P_V\tilde e$ and $(I-P_V)\tilde e$ separately, and with arteriolar versus venular subspaces. On the MRI subset, it does the same for white-matter hyperintensity volume using adjusted partial R².
Two falsification checks guard against residual confounding by things we did not measure, such as cataract or image quality. The first is a negative-control outcome with no plausible microvascular route: hospital admission for accidental injury, which should give $\Delta$ near zero. The second is a negative-control feature: AutoMorph's image-quality grade, entered as if it were an exposure. If either shows sizeable adjusted "prediction," our adjustment is incomplete, and we will report the stroke estimate as an upper bound.
The result would be decisive in either direction. One possibility is that retinal stroke discrimination largely vanishes under flexible adjustment, which would reframe a decade of oculomics claims. The other is that a non-trivial $\Delta$ survives and concentrates in the vascular subspace, which would be the first evidence that the "retina as brain-vessel proxy" story holds inside a foundation model and not only in handcrafted metrics.
References
- A deep learning system for detecting silent brain infarction and predicting stroke risk
- A foundation model for generalizable disease detection from retinal images
- Predicting Stroke Through Retinal Graphs and Multimodal Self-supervised Learning
- Effects of Hypertension, Diabetes, and Smoking on Age and Sex Prediction from Retinal Fundus Images
- RetinaGuard: Obfuscating Retinal Age in Fundus Images for Biometric Privacy Preserving
- Lower Retinal Arteriolar Density Is Associated With Higher Cerebral Small Vessel Disease Burden: An Optical Coherence Tomography Angiography Study
- Self-citation is the hallmark of productive authors, of any gender
- AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline
- Feature Space Translation Framework for Cross-Device Alignment and Derivation of Clinically Relevant Retinal Biomarkers from Foundation Model Embeddings
- C-reactive protein to high-density lipoprotein cholesterol ratio: an independent risk factor for diabetic retinopathy in type 2 diabetes patients