Retinal Vasomotion as a Cross-Species Proxy for Glymphatic Clearance

The brain clears its metabolic waste through a plumbing system rather than a filtration organ: cerebrospinal fluid is pumped along the spaces surrounding arteries, flushes through brain tissue, and exits along veins and lymphatic vessels. This "glymphatic" flow is not driven by a static structure but by rhythm — slow oscillations of blood vessel walls, strongest during deep sleep, that mechanically push fluid through the tissue. In mice this rhythm can be watched directly through a window in the skull, and it can even be strengthened with 40 Hz light-and-sound stimulation, which measurably speeds the clearance of amyloid Multisensory gamma stimulation promotes glymphatic clearance of amyloid. In humans, none of this can be measured without injecting a tracer into the spine, and the one non-invasive proxy that clinicians adopted — a diffusion-MRI index called DTI-ALPS — was recently shown to reflect the geometry of nerve fibers rather than fluid flow White Matter Geometry Confounds Diffusion Tensor Imaging Along Perivascular Space (DTI‐ALPS) Measures. There is, at present, no valid way to measure brain waste clearance in a living human without a lumbar puncture.

The eye offers a way around this, but not the way current "oculomics" uses it. The retina is an anatomical extension of the central nervous system; large-scale imaging-genetics work has shown that retinal traits share genetic architecture with brain MRI traits, with the shared heritability concentrated in glial rather than neuronal pathways Eye-brain connections revealed by multimodal retinal and brain imaging genetics in the UK Biobank. Yet every deep-learning system that reads the retina for brain disease — RETFound, DeepRETStroke, PolarNet+ — operates on static photographs or static 3D vessel maps A foundation model for generalizable disease detection from retinal images; A deep learning system for detecting silent brain infarction and predicting stroke risk; Beyond the Eye: A Relational Model for Early Dementia Detection Using Retinal OCTA Images. The signal glymphatics actually runs on — vessel-wall pulsation and slow vasomotion — is exactly what these pipelines average away. The retina is the only place in the human central nervous system where that rhythm can be recorded non-invasively, at high temporal resolution, through the pupil.

Core Claim

The core claim is that stimulus-evoked retinal vasomotion — the same slow rhythmic vessel-wall motion that mechanically drives glymphatic flow — can be recorded in the living retina and used as a non-invasive readout of glymphatic function, provided two things are done that no existing work does. First, the proxy must be validated where ground truth exists: in mice, by recording retinal microvascular dynamics and brain tracer clearance in the same animal. Second, the frequency content must be rescaled across species rather than copied, because rodent and human hemodynamics operate in different frequency regimes. The claim is testable and falsifiable: if retinal vasomotion power does not predict individual-animal clearance in mice, or if static retinal features predict human outcomes as well as dynamic ones, the idea dies.

This claim is deliberately different from the nearest existing work. The 2025 VCID biomarker workshop positions retinal imaging and glymphatic/perivascular imaging as two parallel, non-interacting biomarker tracks and explicitly leaves open whether retinal parameters correlate with clearance kinetics at all The future of biomarkers for vascular contributions to cognitive impairment and dementia (VCID): proceedings of the 2025 annual workshop of the Albert research institute for white matter and cognition. The UK Biobank eye-brain study links retinal traits to brain structure but only to static MRI metrics, never to fluid flow Eye-brain connections revealed by multimodal retinal and brain imaging genetics in the UK Biobank. PolarNet+ shows retinal capillary dropout detects early Alzheimer's disease but treats the vessels as frozen anatomy Beyond the Eye: A Relational Model for Early Dementia Detection Using Retinal OCTA Images. Nobody has recorded retinal vessel dynamics and brain clearance in the same organism, and nobody has attempted a quantitative mouse-to-human frequency calibration for sensory-evoked vasomotion — a gap the gamma-stimulation authors themselves flag, noting their mouse mechanics remain unconfirmed in a brain roughly 1,000-fold larger Multisensory gamma stimulation promotes glymphatic clearance of amyloid.

Research Program

The program has three linked components: establish the retinal–clearance coupling in mice, derive and test the cross-species frequency scaling, and build the human dynamic readout.

The first component adapts the gamma-stimulation paradigm in 5XFAD mice Multisensory gamma stimulation promotes glymphatic clearance of amyloid. Mice receive 40 Hz audiovisual stimulation (with 8 Hz and 80 Hz controls, matching the published frequency-specificity design), while retinal microvascular dynamics are recorded and brain glymphatic influx and efflux are tracked with intracisternal OVA-647 tracer under two-photon imaging, as in the original study. The retinal recording exploits a technical fact that makes this cheap rather than exotic: optical coherence tomography angiography (OCTA) works by repeatedly scanning the same tissue and registering change, so pulsatile vessel motion is already sampled at kilohertz rates — it is currently discarded as a motion artifact. The analysis re-purposes that discarded signal, extracting per-vessel caliber oscillation spectra in the ~0.1 Hz band where stimulation-driven arterial pulsatility power was shown to rise alongside clearance Multisensory gamma stimulation promotes glymphatic clearance of amyloid. A method for reconstructing flow dynamics from noisy tracer sequences already exists in the glymphatics literature — GlymphVIS, which fits advection–diffusion velocity fields with regularized optimal transport and achieved up to a 5-fold error reduction over plain optimal transport on rat DCE-MRI GlymphVIS: Visualizing Glymphatic Transport Pathways Using Regularized Optimal Transport — and will be reused to quantify clearance rates per animal. Because mice lack a macula and fovea entirely, so macula-centered human metrics cannot be back-translated Inhibition of AMD-Like Pathology With a Neurotrophic Compound in Aged Rats and 3xTg-AD Mice, the shared feature set is restricted to non-macular quantities: vessel-caliber oscillation spectra, capillary perfusion flicker, and layer-resolved flow pulsatility, all of which exist in both species.

The second component turns the cross-species gap from an obstacle into a quantitative model. Rodent resting heart rates run 400–600 beats per minute against the human 60–80, so the vascular oscillation frequencies that drive rodent perivascular pumping cannot be assumed to carry over A Porcine Model of Heart Failure With Preserved Ejection Fraction Induced by Chronic Pressure Overload Characterized by Cardiac Fibrosis and Remodeling. As a first-order allometric hypothesis, the human resonance band is predicted by rescaling the mouse band by the heart-rate ratio:

$$f_{\text{human}} \approx f_{\text{mouse}} \cdot \frac{HR_{\text{human}}}{HR_{\text{mouse}}},$$

where $f$ denotes the vasomotion frequency of interest and $HR$ the species resting heart rate. With $f_{\text{mouse}} \approx 0.1$ Hz Multisensory gamma stimulation promotes glymphatic clearance of amyloid and heart rates of ~70 versus ~500 bpm, this predicts a human band near 0.01–0.02 Hz — plausibly the regime of the infra-slow global brain activity coupled to cerebrospinal fluid movement that the VCID workshop identifies as a promising clearance marker The future of biomarkers for vascular contributions to cognitive impairment and dementia (VCID): proceedings of the 2025 annual workshop of the Albert research institute for white matter and cognition. The form is a crude first-order guess from cardiac allometry, stated as a hypothesis to be estimated in the mouse data (where both heart rate and the pulsatility band are measured) and then tested, not assumed, in humans.

The third component builds the human readout and the model behind it. A pilot cohort undergoes high-speed OCTA of the retina during 40 Hz audiovisual stimulation and during frequency-swept variants, testing whether stimulus-locked retinal vasomotion peaks in the allometrically predicted band. The predictive model is a temporal network operating on retinal image sequences, trained so that its spectral features — not its spatial texture — carry the signal; this is the design decision that the ablation in the next section is built to police. Validation in humans is necessarily indirect, because no non-invasive gold standard for clearance exists (that is the entire motivation). The endpoints are therefore the ones the field currently accepts: enlarged perivascular spaces and infra-slow activity coupled to CSF motion on MRI The future of biomarkers for vascular contributions to cognitive contributions to cognitive impairment and dementia (VCID): proceedings of the 2025 annual workshop of the Albert research institute for white matter and cognition, plus transfer evaluation on the ROAD and ROMCI OCTA cohorts of early-onset Alzheimer's disease and mild cognitive impairment used by PolarNet+ Beyond the Eye: A Relational Model for Early Dementia Detection Using Retinal OCTA Images. The UK Biobank cohort, which pairs retinal imaging with brain MRI and genetics Eye-brain connections revealed by multimodal retinal and brain imaging genetics in the UK Biobank, is useful for the static-association analyses but cannot supply dynamic retinal video, and the proposal does not pretend otherwise.

A structural reason to expect the retina to track clearance at all comes from the glial genetics: the UK Biobank eye-brain overlap is enriched in glial annotations Eye-brain connections revealed by multimodal retinal and brain imaging genetics in the UK Biobank, and the glial element in question — astrocytic endfeet bearing aquaporin-4 water channels — is precisely the machinery that gates glymphatic exchange and whose blockade with TGN020 abolishes stimulated clearance Covert cerebrospinal fluid dynamics dysfunction: evolution from conventional to innovative therapies; Multisensory gamma stimulation promotes glymphatic clearance of amyloid. Retinal Müller cells carry the same endfoot architecture, so the coupling hypothesis has a mechanism, not just a correlation.

Baselines

The naive reference point is DTI-ALPS, the diffusion-MRI index the field currently uses as a non-invasive glymphatic marker in insomnia, Parkinson's lateralization, stroke recovery, and traumatic brain injury The change of MRI indexes of brain glymphatic function and sleep status before and after repeated transcranial magnetic stimulation in insomnia disorder patients; Diffusion–tensor MRI study of the relationship between glymphatic system asymmetry and onset lateralization in Parkinson's disease; Repetitive transcranial magnetic stimulation promotes the recovery of upper limb motor dysfunction in ischemic stroke patients: a DTI-based glymphatic system imaging prospective study; Reduced glymphatic function after traumatic brain injury measured using diffusion MRI. It is included not because it is strong but because it is the incumbent, and beating it is a low bar that must nonetheless be cleared explicitly. Its collapse is documented in detail: radial asymmetry ratios of 1.3 to 1.8 appear in all 35 white-matter regions examined, not only near medullary veins, the signal persists at high b-values where free perivascular fluid would be suppressed, and 7T imaging shows medullary veins deviate by up to 30 degrees from the axis the index assumes White Matter Geometry Confounds Diffusion Tensor Imaging Along Perivascular Space (DTI‐ALPS) Measures. The comparison protocol is to run DTI-ALPS on every human participant who gets the dynamic retinal scan, and to regress the retinal marker against both DTI-ALPS and fiber-geometry covariates. Beating this baseline means the retinal marker predicts MRI endpoints while remaining unexplained by white-matter orientation metrics — no reported result found on Lacuna for DTI-ALPS as a predictor of any independent clearance ground truth, which is itself the point.

The strongest existing competitor is static retinal deep learning, represented by two systems that must both be run. PolarNet+ is a 3.92M-parameter network that converts OCTA volumes into sector grids and links vessel changes across retinal vascular layers, detecting Alzheimer's disease at AUC 88.69% and mild cognitive impairment at AUC 88.02% on ROAD/ROMCI, with external multi-center AUC 86.55% Beyond the Eye: A Relational Model for Early Dementia Detection Using Retinal OCTA Images — the best published static-OCTA number on the exact cohorts used for transfer evaluation here. RETFound is a foundation model pre-trained on 1.6 million retinal images that predicts 3-year heart failure at AUROC 0.794 from fundus photos and 0.799 from OCT, and diabetic retinopathy at AUROC 0.943 on APTOS-2019 A foundation model for generalizable disease detection from retinal images — the strongest general-purpose retinal representation available. Both will be fine-tuned on the new dynamic human dataset with identical tuning budget, compute, and label access as the proposed temporal model, evaluated on the same held-out endpoints (perivascular-space burden, and ROAD/ROMCI disease labels). Beating them means higher AUROC on at least one endpoint with confidence intervals that do not overlap, and — more importantly for the claim — retention of that advantage in the key ablation described next.

The ablation that isolates the proposal's key ingredient is architectural, not dataset-level: the proposed temporal network is run twice, once on the full dynamic sequence and once on temporally collapsed static frames (the per-frame average, which is what PolarNet+ and RETFound effectively consume), with identical parameters, training data, and tuning. The proposal's central bet is that the static version loses predictive power for clearance-related endpoints while the dynamic version keeps it. If the two perform identically, the dynamic ingredient adds nothing and the proposal's claim is refuted regardless of headline AUROC. A second ablation regresses out systemic hemodynamics — heart rate, blood pressure, respiration — recorded simultaneously during scanning, to test whether retinal vasomotion carries brain-specific information rather than a systemic cardiovascular shadow.

Why not just…?

Why not just use DTI-ALPS, since it is already non-invasive and published across four clinical populations? Because the evidence says it does not measure what it claims to measure. The confounding study found the asymmetry it relies on everywhere in white matter, a signal origin inside axons rather than fluid spaces, and up to 30 degrees of vein misalignment violating the method's own assumption White Matter Geometry Confounds Diffusion Tensor Imaging Along Perivascular Space (DTI‐ALPS) Measures. Its authors conclude it should not be used as a standalone glymphatic measure. This alternative is not a real threat to the proposal's motivation; it is the reason the proposal exists.

Why not just fine-tune RETFound or PolarNet+ on clearance endpoints instead of building a dynamic model? This is the real threat, and the proposal is designed so it can lose to it honestly. Static retinal features demonstrably encode cerebrovascular pathology — DeepRETStroke detects silent brain infarction at AUC 0.797 versus 0.633 for clinical traits alone A deep learning system for detecting silent brain infarction and predicting stroke risk, and PolarNet+'s deep vascular complex is the most sensitive layer for early neurodegeneration Beyond the Eye: A Relational Model for Early Dementia Detection Using Retinal OCTA Images. It is entirely possible that static vessel dropout already carries enough signal for disease classification. The decisive experiments are the temporal-collapse ablation and the mouse coupling study: if static frames match dynamic ones on human endpoints, or if retinal vasomotion fails to predict per-animal tracer clearance in mice, then static fine-tuning is the right answer and this proposal should be abandoned in its favor. Conversely, if the dynamic signal predicts clearance where static features do not, the fine-tuning route is capped at disease correlation without a mechanism.

Why not just apply the 40 Hz stimulation protocol directly to humans without any scaling model? Because qualitative frequency copying is a documented mechanism of translational failure. A meta-analysis of animal models in cerebral ischemia attributes preclinical-to-clinical breakdown partly to unaccounted physiological mismatch between species Treatment of acute cerebral ischemia using animal models: a meta-analysis, and the gamma-stimulation authors themselves caution that their mechanics are unconfirmed in a brain with roughly a thousandfold larger volume and an order of magnitude slower heart rate Multisensory gamma stimulation promotes glymphatic clearance of amyloid. The frequency-sweep arm of the human pilot settles this empirically: if the retinal vasomotion response peaks at 40 Hz-evoked frequencies in humans exactly as in mice, the scaling model was unnecessary; if it peaks in the allometrically predicted slower band, direct copying would have targeted the wrong regime.

Why not just use clinical risk factors, which are cheap and guideline-embedded? Because they demonstrably miss subclinical cerebrovascular disease. In the DeepRETStroke validation, clinical traits alone detected silent brain infarction at AUC 0.633 against 0.797 for retinal imaging, and retinal-image-guided stratification was associated with 82.44% fewer recurrent stroke events over one year than trait-guided care A deep learning system for detecting silent brain infarction and predicting stroke risk. Clinical traits are retained as a covariate baseline in every human analysis, but the evidence already shows where their ceiling sits.

Decisive Checks

Three experiments carry the burden of proof. First, the mouse coupling study: across animals and stimulation frequencies, retinal vasomotion power in the ~0.1 Hz band must predict tracer clearance rate measured by two-photon imaging, with the relationship surviving the systemic-hemodynamics regression. The sharpest control is pharmacological: TGN020 blockade of aquaporin-4 abolishes stimulated clearance in mice Multisensory gamma stimulation promotes glymphatic clearance of amyloid; if retinal vasomotion persists undiminished under blockade while brain clearance dies, the retinal rhythm is merely an epiphenomenon of sensory drive and the proxy claim is false — if it attenuates in lockstep, the coupling is real and glially gated. Second, the human frequency-sweep pilot must show stimulus-locked retinal vasomotion in the predicted band, distinguishing the allometric hypothesis from direct frequency copying. Third, the static-collapse ablation on the human cohort must show a predictive gap between dynamic and static inputs on perivascular-space endpoints, and the dynamic model must be checked against the fiber-geometry covariates that sink DTI-ALPS White Matter Geometry Confounds Diffusion Tensor Imaging Along Perivascular Space (DTI‐ALPS) Measures. Only if all three pass does the program proceed to the longitudinal question that ultimately matters: whether the retinal rhythm flags failing clearance before cognitive symptoms appear — the question the whole oculomics enterprise, currently running on frozen anatomy, has left unanswered.

References

  1. Multisensory gamma stimulation promotes glymphatic clearance of amyloid
  2. White Matter Geometry Confounds Diffusion Tensor Imaging Along Perivascular Space (DTI‐ALPS) Measures
  3. Eye-brain connections revealed by multimodal retinal and brain imaging genetics in the UK Biobank
  4. A foundation model for generalizable disease detection from retinal images
  5. A deep learning system for detecting silent brain infarction and predicting stroke risk
  6. Beyond the Eye: A Relational Model for Early Dementia Detection Using Retinal OCTA Images
  7. The future of biomarkers for vascular contributions to cognitive impairment and dementia (VCID): proceedings of the 2025 annual workshop of the Albert research institute for white matter and cognition
  8. GlymphVIS: Visualizing Glymphatic Transport Pathways Using Regularized Optimal Transport
  9. Inhibition of AMD-Like Pathology With a Neurotrophic Compound in Aged Rats and 3xTg-AD Mice
  10. A Porcine Model of Heart Failure With Preserved Ejection Fraction Induced by Chronic Pressure Overload Characterized by Cardiac Fibrosis and Remodeling
  11. Covert cerebrospinal fluid dynamics dysfunction: evolution from conventional to innovative therapies
  12. The change of MRI indexes of brain glymphatic function and sleep status before and after repeated transcranial magnetic stimulation in insomnia disorder patients
  13. Diffusion–tensor MRI study of the relationship between glymphatic system asymmetry and onset lateralization in Parkinson’s disease
  14. Repetitive transcranial magnetic stimulation promotes the recovery of upper limb motor dysfunction in ischemic stroke patients: a DTI-based glymphatic system imaging prospective study
  15. Reduced glymphatic function after traumatic brain injury measured using diffusion MRI
  16. Treatment of acute cerebral ischemia using animal models: a meta-analysis