Nutrition studies keep reporting that older adults who eat more of certain nutrients score better on cognitive tests. Examples include an L-shaped vitamin E threshold near 9.3 mg/day in US survey data and lower incident decline at higher choline intake in a Chinese cohort. None of these studies can say what in the brain is being protected. A natural guess is the small blood vessels, and the retina is the one place where small vessels continuous with the brain can be photographed directly. The obvious study regresses brain MRI on diet with the retinal image as a "mediator", a variable on the causal path between them. That study is quietly broken. Diet measured from a few days of recall is noisy, and a retinal summary from one photograph is noisy. In mediation analysis, noise in the mediator makes the indirect path look smaller and the direct path larger. Noise in the exposure shrinks everything. The naive study would therefore under-credit the vascular route, and by an amount nobody knows.
This note proposes a design that corrects both noise sources with data that UK Biobank already contains. Repeated 24-hour dietary recalls give replicate measurements of each person's diet. Photographs of the left and right eye give replicate measurements of each person's systemic vascular state. The brain outcome is not an arbitrary MRI trait. It is an Alzheimer's-specific atrophy signature built and validated in the Alzheimer's Disease Neuroimaging Initiative (ADNI), where we can confirm that it tracks clinical progression. The product is a measurement-error-corrected split of the diet–brain association into a retinal-vascular part, a retinal-neural part, and a remainder, with confidence intervals that account for both sources of noise.
Core Claim
Replicates exist on both ends of the causal chain, so the attenuation of each can be estimated rather than assumed. Write the reliability of a diet measure averaged over $k$ recalls as
$$\lambda_k = \frac{\sigma_b^2}{\sigma_b^2 + \sigma_w^2/k},$$
where $\sigma_b^2$ is the true between-person variance in habitual intake and $\sigma_w^2$ is day-to-day within-person variance. This is the standard variance-components model behind multi-day averaging and the National Cancer Institute (NCI) usual-intake method. An analogous reliability $\rho$ for the retinal mediator comes from the covariance between left-eye and right-eye summaries relative to their variance. The shared, cross-eye component is exactly the systemic signal we want, because systemic vascular state affects both eyes while eye-local disease and imaging artifacts mostly do not. The claim is that once both $\lambda_k$ and $\rho$ are corrected, the estimated vascular share of the diet–brain association changes materially from the naive estimate. The second claim is that this share differs between explicit vessel geometry, neuroretinal layer thickness, and the diffuse signal captured by retinal foundation models. That would settle, at least within one cohort, the tension the literature currently leaves open.
The statistical ingredients are standard: regression calibration, variance components, and mediation with cross-fitting. What is new is the double-replicate design, its use for a cross-organ mediation question, and anchoring the outcome in an ADNI-validated disease axis instead of generic brain volume.
Where This Sits
Three bodies of work motivate the question without answering it. On the diet side, Zhang et al.'s NHANES vitamin E analysis averaged two 24-hour recalls in 2,255 adults aged 60 and over. It found an odds ratio of 0.88 per mg/day below the 9.296 mg/day inflection point. The CHNS choline cohort used multi-wave three-day recall averages and found 27.8–33.9% lower hazard of decline in middle quintiles. Both use cognitive screening only, with no imaging, so they cannot locate the effect. The UK Biobank DASH genome-wide study shows that repeated Oxford WebQ recalls yield a usable diet-quality score at scale, but it never touched the eye or brain imaging in the same cohort. Cowan et al. supply the method we reuse: a two-part mixed model that separates within-person from between-person variance for usual intake.
On the eye–brain side, the UK Biobank eye–brain imaging genetics study found 625 retinal–brain phenotypic associations in roughly 6,000 people imaged at both organs. It also reported shared genetic loci and Mendelian-randomization evidence linking Alzheimer's liability to retinal morphology. It is cross-sectional, ignores diet, and focuses on layer thickness rather than vessel geometry. The retinal foundation models RETFound and PRETI are large self-supervised image encoders trained by reconstructing masked image patches. They give strong systemic predictions, for example a heart-failure AUROC of 0.794 from RETFound on fundus photographs and a coronary calcium AUROC of 0.899 from PRETI. Neither has been tested on neurodegeneration or diet, and neither separates vessel from background signal. That separation matters because Kim et al. showed that fundus images with the vessels removed still predict age and sex. A foundation-model "mediator" may therefore not be vascular at all. PRETI's use of left/right eye pairs as a training signal is the closest methodological cousin to our use of the two eyes. PRETI uses the pairs to learn invariance; we use them to measure reliability.
On the ADNI side, Pennec et al. subtract a learned normal-aging deformation from longitudinal MRI to isolate an AD-specific morphological component. That component separates AD from controls with 91% sensitivity and 84% specificity, and this disentanglement supplies our outcome. Progression models such as missing-data SuStaIn and PerM-MoE represent vascular disease only through white matter hyperintensity volume on FLAIR. That gap motivates asking what the retina adds. These models also constrain the design: the outcome axis must survive transfer across scanners, which GSplit LBI shows is a real source of bias. ADNI's recruitment skew, documented in the HABS-HD comparison, and UK Biobank retinal model biases (Biases in a UK Biobank retinal classifier) limit how far any result generalizes. We report this rather than design around it.
Research Program
The analysis cohort is UK Biobank participants with at least one Oxford WebQ recall, fundus photographs of both eyes, and brain MRI. The research file puts concurrent eye and brain imaging at roughly 6,000 people. The diet-plus-retina subset without MRI is larger and is used for the diet-to-mediator leg. The exact counts and the time gaps between recall, retinal imaging and MRI must be extracted before any analysis is locked. If the gap between retina and MRI is long, the mediation is read as temporally ordered rather than simultaneous.
Exposures are a DASH adherence score and two nutrients chosen because prior cognitive findings exist: vitamin E and choline. Usual intake is estimated with the NCI-style two-part model, with $\sigma_w^2$ taken from participants with several recalls. Energy adjustment and a BMI-stratified sensitivity analysis address systematic under-reporting, which averaging cannot remove.
There are three mediator families, each computed per eye. The first is explicit vessel geometry: arteriolar and venular caliber, tortuosity, branching, and fractal dimension. These come from vessel maps produced by SGL vessel segmentation, following the caliber tradition of Cheung et al. and the UK Biobank deep-learning calibre study. The second is neuroretinal thickness of the ganglion cell and nerve fiber layers from OCT, with segmentation uncertainty available from RLS_PSDF. The third is frozen RETFound and PRETI embeddings. We run these embeddings twice: once on the full image, and once on vessel-masked images, following Kim et al. The difference shows how much of the foundation signal is vascular. Each high-dimensional embedding is reduced to a scalar mediator by a projection learned in held-out folds, using cross-fitting so that selection does not leak into the estimate. Mediator reliability $\rho$ is the cross-eye covariance of that scalar, estimated after excluding eyes with local pathology and adjusting for image quality. Image quality matters here because dementia itself is linked to poorer OCT scans (OCT scan quality and dementia).
Outcomes are built in ADNI and then transferred. We train the Pennec-style AD-specific deformation score on ADNI longitudinal MRI and verify that it predicts change in the Clinical Dementia Rating sum of boxes (CDR-SB) there. We then apply it to UK Biobank T1 scans with harmonization, alongside white matter hyperintensity volume as the conventional vascular outcome. Mediation is estimated in a structural model where exposure and mediator are latent variables with replicate-based error variances. A naive model with no correction is fit on identical data for comparison.
Baselines
The do-nothing reference is the naive mediation analysis that every existing diet–cognition study would run. It uses the mean of whatever recalls each person has, a single-eye mediator, and ordinary regression. Corrections are judged against it. The proposal "beats" it if the corrected indirect effect differs from the naive one by more than its own confidence interval half-width, and the recall-subsampling check below confirms that the correction is calibrated. A large difference that fails calibration counts as a failure.
The strongest competing explanation for any retinal mediation is that the retina is just a camera for blood pressure and glycemia. The systemic-marker baseline therefore puts measured blood pressure, HbA1c and lipids in the mediator slot, using the same exposures, outcome, covariates and cross-fitting, with their own reliability estimated from repeated visit measurements where available. The retinal mediators must explain diet–brain association beyond these markers, tested by including both. Otherwise the vascular-retina story adds nothing clinically. The only quantitative evidence in the file that the retina outperforms systemic markers is Dai et al.'s qualitative statement about blood glucose. No reported effect size was found on Lacuna for this comparison in a brain outcome.
The strongest existing retina-to-brain predictor is the feature set of the UK Biobank eye–brain study. Adding its retinal traits to brain MRI raised fluid-intelligence prediction from r = 0.314 to 0.344. We reproduce that style of engineered-plus-CNN retinal features as a mediator family and require any foundation-embedding mediator to at least match its cross-validated association with our ADNI-anchored outcome. No reported result was found on Lacuna for diet-to-retina mediation of any brain outcome, so there is no published number to beat on the main estimand. The comparison is internal and pre-registered.
Two ablations isolate the key ingredient. The exposure-only correction fixes $\lambda_k$ but leaves the mediator uncorrected. The mediator-only correction does the reverse. If one correction does all the work, the double-replicate design is unnecessary, and we will say so. The vessel-masked versus full-image foundation embedding is a third ablation. It isolates whether foundation-model mediation is vascular. All mediator families share one tuning budget for the projection step, and all are fit on the same folds.
Why not just…?
Why not just regress brain MRI on diet and skip the retina? That estimate is reported as the total effect and is part of the analysis. It cannot say whether the route is vascular, and that is the open question left by the vitamin E and choline studies. If the total effect is null after exposure correction, the mediation question is moot. We would report that as the main finding, which is still informative given the positive cognitive literature.
Why not just use an FFQ or a single recall? Food frequency questionnaires rank people but quantify absolute intake poorly (Sausenthaler et al.). A single recall gives no internal estimate of $\sigma_w^2$, so its attenuation cannot be corrected, only assumed. The subsampling experiment shows directly how much naive single-recall estimates are biased in this cohort.
Why not just use engineered caliber and layer thickness instead of foundation models? This is a real threat. These measures are interpretable and may carry all the vascular signal. They are a full mediator family here, not a strawman. If they match or exceed the foundation embeddings once the embeddings are restricted to vessel-masked or vessel-only signal, the recommendation will be to use engineered measures. That result would also tell the field that RETFound-style systemic gains come from non-vascular background.
Why not just correct exposure error with published reliability coefficients? Published reliabilities come from other populations and instruments. UK Biobank recall counts vary by person, so reliability varies by person. The two-eye mediator correction has no published analogue at all. Estimating both reliabilities internally is cheap with the data at hand.
Proposed Experiments and Decisive Checks
First, a simulation calibrated to the estimated $\sigma_b^2$, $\sigma_w^2$ and cross-eye reliability. It shows the size and direction of naive bias in the indirect and direct effects and confirms that the corrected estimator recovers planted effects with nominal interval coverage.
Second, the recall-subsampling check, which is the decisive test of the exposure correction. Among participants with four or more recalls, we recompute estimates using $k = 1, 2, 3$ randomly chosen recalls. Naive estimates should shrink roughly in proportion to $\lambda_k$, because regression dilution scales a slope by the reliability. Corrected estimates should stay flat across $k$. Drift in the corrected estimates means the variance model is wrong. Note what this check does not do: it validates the correction, not causality. Stable between-person reporting bias and confounding survive averaging. The analogous mediator check compares one-eye and two-eye estimates.
Third, the main estimates. For DASH, vitamin E and choline, we report total, indirect and direct effects on the ADNI-anchored AD signature and on white matter hyperintensity volume, for each mediator family, both with and without the systemic markers. The decisive pattern for the vascular hypothesis is retinal vessel-geometry mediation that survives systemic-marker adjustment, appears for the white matter outcome, and appears more weakly for the AD-specific atrophy signature. Mediation carried by layer thickness instead would point to a neural rather than vascular route.
Fourth, outcome validity and threats. We confirm in ADNI that the transferred signature tracks CDR-SB change, and in UK Biobank that scanner harmonization removes site effects. We run unmeasured-confounding sensitivity analyses for both mediation assumptions, plus a negative-control exposure with no plausible vascular pathway. The negative control should show no indirect effect. If it does, the retinal mediator is picking up general health or socioeconomic status rather than vascular biology, and the main conclusions are withdrawn.
References
- Association between dietary vitamin E intake and cognitive decline among old American: National Health and Nutrition Examination Survey
- Age and Gender Disparities in the Association of Long-Term Dietary Choline and Choline Compound Intakes with Incident Cognitive Decline in Middle-Aged and Older Chinese Adults: A Prospective Cohort Study
- Genome-Wide Association Analysis of Over 170,000 Individuals from the UK Biobank Identifies Seven Loci Associated with Dietary Approaches to Stop Hypertension (DASH) Diet
- Total Usual Micronutrient Intakes Compared to the Dietary Reference Intakes among U.S. Adults by Food Security Status
- Eye-brain connections revealed by multimodal retinal and brain imaging genetics in the UK Biobank
- A foundation model for generalizable disease detection from retinal images
- PRETI: Patient-Aware Retinal Foundation Model via Metadata-Guided Representation Learning
- Effects of Hypertension, Diabetes, and Smoking on Age and Sex Prediction from Retinal Fundus Images
- Disentangling normal aging from Alzheimer's disease in structural magnetic resonance images
- Multimodal subtypes identified in Alzheimer’s Disease Neuroimaging Initiative participants by missing-data-enabled subtype and stage inference
- Flexible Multimodal Neuroimaging Fusion for Alzheimer's Disease Progression Prediction
- GSplit LBI: Taming the Procedural Bias in Neuroimaging for Disease Prediction
- Assessing the Generalizability of ADRDs Subtypes Across Diverse Cohorts HABS‐HD and ADNI
- On Biases in a UK Biobank-Based Retinal Image Classification Model
- Influence of secondhand smoke exposure on the retinal vasculature of children in Hong Kong
- Association between deep learning measured retinal vessel calibre and incident myocardial infarction in a retrospective cohort from the UK Biobank
- Association Between Dementia and Optical Coherence Tomography Scan Quality
- A deep learning system for predicting time to progression of diabetic retinopathy
- Regional and socio-economic differences in food, nutrient and supplement intake in school-age children in Germany: results from the GINIplus and the LISAplus studies