A Pupil Test for When Drying the Macula No Longer Restores Vision

Patients with wet macular disease receive eye injections every one to two months, often for life. Doctors decide when to inject, and when to switch to a stronger or longer-lasting drug, almost entirely from retinal scans that show whether fluid is present. A growing set of studies shows that these scans can mislead. Newer drugs dry the retina dramatically while vision stays flat, because the light-sensing cells underneath have already died. The question that matters for the patient's burden, which is whether another injection or a drug switch can still buy back function, is currently answered only by a letter chart. A letter chart is a crude, foveal-only measure of the light-sensing cells.

We propose to answer that question with a five-minute pupil test in a virtual-reality headset. The pupil constricts to light through two routes. One runs through the eye's light-sensing cells (cones and rods), which macular disease destroys. The other runs through a small population of ganglion cells that contain their own light pigment, melanopsin, and these cells largely survive outer-retinal disease. Colored light can drive the two routes separately. The ratio of the cone-driven response to the melanopsin-driven response, measured over the macula and compared against the fellow eye, should therefore isolate surviving photoreceptor function. The same measurement corrects for the lens yellowing, small pupils and general aging that confound raw pupil measures. The core bet is that this ratio predicts which treated eyes have hit a functional ceiling, and that it does so better than the microstructural scan features clinicians already have.

Core Claim

Chromatic pupillometry means measuring pupil responses to narrow-band colored light chosen to favor one receptor class. It has so far been validated almost only for glaucoma and systemic neurodegeneration. Head-mounted systems detect glaucoma with high accuracy: dilation time after a superior-quadrant blue pulse reached an AUC of 0.909 in Evaluation of the Glaucomatous Macular Damage by Chromatic Pupillometry. Combining inter-eye asymmetry with anisocoria duration reached 0.992 in Binocular head-mounted chromatic pupillometry can detect structural and functional loss in glaucoma. The first of those papers names two gaps itself. Current pupil measures have no disease-specific signature separating optic neuropathy from maculopathy, and the research file records that the method has never been tested as a monitoring endpoint during anti-VEGF therapy.

We claim that a light-adapted, macula-targeted ratio of cone-driven to melanopsin-driven pupil response supplies that missing signature and fills that missing role. The signature exists because glaucoma kills ganglion cells and so depresses both terms, whereas outer-retinal disease depresses mainly the cone term. In eyes being switched to faricimab or high-dose aflibercept, we expect the pre-switch ratio to predict functional recovery on microperimetry and visual acuity, with added value over deep-learning fluid volumes and ellipsoid-zone and inner-layer disorganization scores. Throughout, "added value" means that a model given the scan features plus the pupil index predicts recovery better than the same model given only the scan features.

Concretely, for stimulus sector $s$ (central disc or paracentral annulus) of eye $e$, let $A^{c}_{e,s}$ be the peak constriction to a red pulse on a rod-saturating blue background, and let $A^{m}_{e,s}$ be the sustained post-illumination constriction measured 6 seconds after a bright blue pulse. The index is

$$ R_{e,s} = \log \frac{A^{c}_{e,s}}{A^{m}_{e,s}}, \qquad \Delta R_s = R_{\text{study},s} - R_{\text{fellow},s}, $$

The log ratio cancels factors that scale both responses alike, such as pre-receptoral light loss and the iris and motor pathway. The fellow-eye difference $\Delta R_s$ cancels person-level aging. One caveat applies: lens yellowing attenuates blue more than red, so the cancellation is only partial. We therefore also record lens density from the device's age model, or from slit-lamp grading, as a covariate.

Research Program

The work has three stages, each with a stopping rule. The first stage is engineering and reliability. We will adapt the binocular headset protocol from the glaucoma studies (HTC Vive Pro Eye with 90 Hz Tobii tracking, 465 nm and 630 nm pulses) so that it replaces quadrant stimuli with a central 10° disc and a 10–20° annulus. Only light-adapted cone and melanopsin conditions are used, which avoids the 20–30 minute dark adaptation that rod isolation requires. In 30 healthy participants and 30 patients with unilateral nAMD, we measure same-day and one-week test–retest reliability of $R$ and $\Delta R$. The go/no-go criterion is that the inter-eye difference in unilateral disease must exceed the 95% repeatability limit in most patients.

The second stage is cross-sectional validity. In about 150 eyes spanning nAMD, DME, glaucoma and healthy controls, we test two things. The first is whether $\Delta R$ tracks sector-matched mesopic microperimetry sensitivity better than central subfield thickness does. Microperimetry maps point-wise retinal sensitivity under fundus tracking, and it has been shown to resolve function lost over specific lesions in Spatially Resolved Association of Structural Biomarkers on Retinal Function in Non-Exudative Age-Related Macular Degeneration Over 4 Years. The second is whether the two-term signature separates macular disease from glaucoma where single pupil measures cannot.

The third stage is the main longitudinal test. We will run a prospective observational cohort of recalcitrant nAMD and DME eyes switched to faricimab or aflibercept 8 mg. This is the population in which the anatomical-functional dissociation has been documented. In Real-World Data on Morphological and Functional Responses After Switching to Faricimab in Recalcitrant, Chronic Diabetic Macular Edema, CST fell from 468.5 to 339.5 µm while acuity plateaued. In Retinal and choroidal efficacy of switching treatment to faricimab in recalcitrant neovascular age related macular degeneration, 56% of eyes became dry with acuity only stabilized. In Deep learning-assisted analysis of biomarker changes after increase of dosing from aflibercept 2 mg to 8 mg in therapy-resistant neovascular age-related macular degeneration, fluid volumes fell while acuity did not change. Those papers motivate the study but are small retrospective series. We follow eyes for 12 months, with pupillometry, OCT, microperimetry and acuity at every visit. Treatment decisions stay with the clinician, so the study does not alter care.

Treatment burden enters as the application. We will analyze two possible decision rules counterfactually. The first would stop escalating doses in eyes whose index shows a fixed ceiling. The second would extend intervals in eyes with trace fluid but a stable index. For each rule we estimate injections avoided and the functional-loss events that the rule would have missed. This defines what a future randomized trial would test; it is not a recommendation.

Baselines

The naive reference is current practice: CST plus letter-chart acuity at baseline, used to predict 6- and 12-month microperimetry and acuity change. Any proposed index has to clear this floor first.

The strongest existing competitor is microstructural OCT. It combines deep-learning fluid volumes (as in the aflibercept 8 mg study above) with graded ellipsoid-zone and external-limiting-membrane integrity, disorganization of retinal inner layers, and hyperreflective foci counts. These are the features the dissociation literature blames for the ceiling, as in Predictive value of hyperreflective foci for anti-VEGF therapeutic outcomes in different subtypes of diabetic macular edema: a retrospective analysis. The nearest modeling precedent is A machine learning model for predicting anatomical response to Anti-VEGF therapy in diabetic macular edema, whose logistic regression reached an AUC of 0.90 for predicting a ≥20% thickness reduction. That target is anatomical, and the paper found intact ellipsoid zone to be a negative predictor of it, which illustrates how anatomical and functional targets diverge. We found no result on Lacuna for predicting functional recovery after an anti-VEGF switch from OCT biomarkers. We will therefore build that baseline ourselves on the same eyes: the same logistic and gradient-boosted models, nested cross-validation, and an equal hyperparameter budget. Beating it means a cross-validated AUC gain of at least 0.05 for the functional-ceiling target, defined as less than 2 dB mean sector-sensitivity gain at 6 months, with a confidence interval excluding zero by DeLong test. A gain in explained variance of continuous sensitivity change also counts.

Two further baselines isolate the key ingredient. The first is the best single raw pupil measure from the glaucoma work: dilation time, whose 0.909 AUC was for glaucoma versus control, not this task. The second is the cone term $A^{c}$ alone, without melanopsin normalization or fellow-eye differencing. If either matches $\Delta R$, the normalization idea fails even if pupillometry itself is useful.

Why not just…?

Why not just use microstructural OCT, which already images the dead photoreceptors? This is the most serious threat. Ellipsoid-zone loss is visible, and deep-learning segmentation is improving. Structure and function decouple in both directions, however. The cystoid DME subtype had the largest anatomical improvement and the worst acuity in the hyperreflective-foci study, and macular-hole eyes recover function only as the ellipsoid zone slowly reconstitutes, per The Healing Process and Functional Recovery of Neuroretina after Idiopathic Macular Hole Surgery without Internal Limiting Membrane Reversal Tamponade. Cells can look present and be dysfunctional, or look disrupted and recover. The incremental-AUC test above settles the question directly. If pupillometry adds nothing over OCT, we will report that it does not.

Why not just use microperimetry, the established functional test? Microperimetry is our reference standard, so the pupil test cannot beat it on accuracy. Its role in this study is to be the yardstick, not the competitor. It is slow, needs a trained operator and a cooperative patient, and is therefore rarely repeated at injection visits. Targeted grids such as Targeted Microperimetry Grids for Focal Lesions in Intermediate AMD: PINNACLE Study Report 7 exist precisely to shorten it. The pupil test claims to be an objective, fast proxy, so we report agreement with microperimetry, test duration, and the failure rate in low-vision eyes.

Why not just use existing glaucoma pupil metrics unchanged? These metrics are dominated by ganglion-cell and melanopsin dynamics, and they are confounded by aging, as Analysis for Extracted Features of Pupil Light Reflex to Chromatic Stimuli in Alzheimer’s Patients notes. In comorbid eyes they would conflate glaucoma with maculopathy. The ablation comparing raw measures with $\Delta R$ tests this directly.

Decisive Checks

Three results would end or redirect the program. First, if in unilateral nAMD the inter-eye difference $\Delta R$ does not exceed test–retest noise, the macular signal is too small relative to the whole-field pupil response, and we stop. Second, if $\Delta R$ does not correlate with sector-matched microperimetry more strongly than CST does, the index is not measuring outer-retinal function. Third, if in the switch cohort the pupil index adds no predictive value over the microstructural OCT model for the functional-ceiling target, the clinical case collapses, and the treatment-burden analysis should not proceed. The secondary specificity check is that glaucoma eyes should show reduced $A^{m}$ with near-normal $R$, while macular eyes show reduced $R$. This dissociation pattern, if found, would also be the first pupillary signature that distinguishes the two disease classes.