Many patients with wet macular degeneration or diabetic eye swelling get monthly eye injections of drugs that block a blood-vessel growth signal (anti-VEGF drugs). In a large share of them, fluid keeps leaking into the retina anyway. Clinicians respond by switching these "recalcitrant" eyes to a newer drug: faricimab, which also blocks a second signal (angiopoietin-2), or a fourfold higher dose of aflibercept. The evidence that switching helps comes almost entirely from small, uncontrolled before-and-after series. In those series, thickness or fluid is measured just before and just after the switch. These eyes were chosen for switching because they looked their worst, so some of the measured improvement could be regression to the mean (eyes picked at a bad moment tend to look better later regardless), or simply the effect of continued injections.
We propose to answer the causal question these series cannot. For a given eye, does switching drugs improve fluid and vision more than staying on the current drug would have, among eyes whose recent fluid history looks the same? The key ingredient is matching on recent fluid history rather than on a single snapshot. We describe each eye by how its separate fluid compartments and its photoreceptor layer (the light-sensing cells) have been changing over the months before a switch decision, measured from routine optical coherence tomography (OCT, a cross-sectional laser scan of the retina) by automated segmentation. We then compare switchers with non-switchers who had matching trajectories at the same decision point. This is a retrospective "target trial emulation," meaning a historical dataset analyzed as if it were a randomized trial with explicit eligibility, time zero, and treatment strategies. It is built around the fact that clinicians decide to switch based on how fluid has been behaving over time, not on how it looks on one day.
Core Claim
The published switch studies report large anatomical gains. In recalcitrant wet macular degeneration (nAMD), central subfield thickness (CST) fell from 361.4 to 318.3 µm and 56% of eyes became dry after three faricimab injections (Eckardt et al., 6-month results). In chronic diabetic macular edema (DME), CST fell from 468.5 to 339.5 µm (Deiters et al.). After escalation to aflibercept 8 mg, the median intraretinal fluid fell from 8 to 2 voxels (Hafner et al.).
Our claim is that these numbers mix three things: the drug effect, regression to the mean, and the effect of the clinician's timing choice. Separating them requires conditioning on the pre-switch trajectory of each fluid compartment. Conditioning on baseline covariates or on central thickness is not enough. We expect two findings. First, a confounding-adjusted switch effect on fluid that is smaller than the uncontrolled estimate but nonzero. Second, a vision benefit concentrated in eyes whose photoreceptor layer was still intact and stable before switching. That second finding would turn the documented gap between drying and vision (Deiters et al.) into a usable rule for choosing which eyes to switch. As far as the file shows, no study has reported a confounding-adjusted switch effect. We found no reported result on Lacuna for a controlled or trajectory-adjusted estimate of the faricimab or aflibercept 8 mg switch effect.
Research Program
Data and measurements. The core data source is the LMU Munich Smart Eye Database (SmEyeDat), which already holds the faricimab and aflibercept 8 mg switch cohorts. The feasibility assumption we must verify first is that the same database contains enough contemporaneous non-switchers: eyes with persistent fluid that stayed on aflibercept 2 mg or ranibizumab. If it does not, the study needs a second center before it can start.
Segmentation. Every OCT volume is segmented into intraretinal fluid, subretinal fluid, fibrovascular pigment epithelial detachment, subretinal hyperreflective material, and photoreceptor-layer integrity (the ellipsoid zone and external limiting membrane). The segmentation method comes from existing models: an nnU-Net-style fluid segmenter (nnUNet RASPP) and outer-retina boundary extraction as in Song et al.. Before any segmenter is used, we check it across devices on the RETOUCH data.
Trajectory features. For each eye at each visit, we summarize the previous three visits per compartment. The summary includes the current level, the rate of change per week fitted against actual calendar dates, cumulative fluid burden, and the time since the last fully dry scan. Fitting against calendar dates matters because a short visit interval should not be misread as acceleration.
Sequential trial emulation. Every visit at which an eye meets the eligibility rule is a candidate time zero. The rule is persistent fluid on at least three consecutive visits at an interval of six weeks or less, after a loading phase. The two strategies compared are "switch now" and "continue current agent for the next 16 weeks." Switching is modeled with inverse probability weights given the trajectory features, systemic markers (renal function, motivated by Wu et al. and Lu et al.), and injection history. Later deviations from the assigned strategy are handled by censoring and reweighting.
Outcomes. The primary outcome is compartment-specific fluid volume at 16 weeks. Secondary outcomes are best-corrected visual acuity, photoreceptor-integrity change, and the achieved injection interval.
Handling irregular visits. Visit timing in treat-and-extend practice depends on how sick the eye is. We therefore also fit the Bayesian joint model of outcome, treatment, and visit processes with G-computation from the dynamic-regime target-trial work. That paper shows that ignoring the visit process biases estimates. Here it is a robustness analysis, not the headline method.
Effect modification. The effect modifier is the pre-switch photoreceptor trajectory. This directly tests whether the switch improves vision only when the light-sensing layer has not yet degraded.
Baselines
The naive reference. We reproduce the published uncontrolled pre/post contrast on our cohort: CST and compartment volumes at the last pre-switch visit versus after three injections, tested with a paired test as in Eckardt et al. and Hafner et al.. The reported numbers above are the target this reproduction should land near. If it does not, our cohort differs from theirs and that has to be explained before anything else.
The strongest conventional competitor. This is 1:1 propensity score matching on baseline clinical variables (age, baseline acuity, prior injection count, CST at the decision visit), in the style of Tewari et al.. It is what a careful clinical epidemiologist would do first. It runs on the identical eligible person-visits, with the same outcomes, the same 16-week window, and the same bootstrap for intervals. The only difference from our method is the information used to adjust.
The key ablation. Our full method is run with CST trajectories in place of compartment trajectories. This isolates whether compartment-level segmentation, rather than trajectory conditioning per se, carries the adjustment. Hafner et al. motivate the test: CST missed changes that voxel-level fluid caught.
What counts as beating a baseline. A baseline is "beaten" if three conditions hold. Our estimate passes the falsification checks below where the baseline fails them. It achieves standardized mean differences below 0.1 on the pre-switch trajectory features. And it yields a materially different effect estimate, with the confidence interval excluding the baseline's point estimate.
Predictive reference points. There is no published causal estimate to compare against. For the secondary predictive task of identifying switch responders, the reference numbers in the file are these: the loading-phase model of Song et al. reached an AUC of 0.980 for 1-year persistent disease activity, versus 0.948 from baseline alone. Lu et al. reached an AUC of 0.90 with baseline logistic regression for at least 20% thickness reduction. The APTOS models in Zhang et al. reached an AUC of 80.06% for treatment continuation. None of these tasks is identical to ours, so they serve as context, not head-to-head targets.
Why not just…?
Why not just trust the before-and-after switch studies, since the effects are large? Eyes are switched at their worst moment, and injections continue throughout. Both regression to the mean and continued dosing would produce thinning even if the new drug added nothing, and Deiters et al. themselves leave this confound unaddressed. Our negative-control experiment measures the problem directly. It applies the same selection rule to non-switchers and measures how much they "improve" at a pseudo-switch date.
Why not just do propensity matching on baseline variables? The switch decision responds to how fluid has been changing. Two eyes with equal CST, one drying and one worsening, are not exchangeable. The joint-model simulations in the dynamic-regime paper show bias when this time-varying, visit-driven process is ignored. This alternative is a real threat, however, if clinicians in practice switch on snapshot criteria. In that case baseline matching would balance the trajectories too. We will check by reporting trajectory balance after baseline matching. If balance is already good, we will say so and our added machinery is unnecessary.
Why not just track central thickness with mixed-effects models? CST is free from the scanner and widely used. But Hafner et al. found that CST showed no significant change where intraretinal fluid and pigment-epithelial detachment fell significantly. Thickness also says nothing about the photoreceptor layer, which is the hypothesized modifier of the vision benefit. The CST-only ablation settles whether this matters for confounding adjustment.
Why not just fit a full Bayesian joint model or an image-forecasting model of the counterfactual retina? Models such as longitudinal flow matching or LTI–Gaussian-process treatment-response models are more expressive. But they need more data than a cohort of tens to a few hundred switchers can support, and their extra assumptions are harder to audit. We keep the joint model as a robustness check. If it and the simpler weighted emulation agree, the simpler one is the recommendation.
Decisive Checks
Segmentation transfer. Every segmenter must reach cross-vendor agreement comparable to the published leave-one-vendor-out Dice of 0.84 for nnUNet RASPP before its outputs enter the trajectory features. RETOUCH is the test set.
Pseudo-switch negative control. Among non-switchers, we pick dates at which the eligibility rule was met and estimate the "effect" of nothing happening. Our method should return roughly zero. The naive pre/post method is predicted to return a spurious improvement, and the size of that spurious improvement is the regression-to-the-mean component hidden in the published series.
Pre-period falsification. The estimated switch effect on fluid change during the eight weeks before time zero must be null. If it is not, residual confounding remains.
Effect modification by photoreceptor trajectory. This is the scientifically decisive result. If the vision benefit of switching is confined to eyes with a stable photoreceptor layer before switching, and the fluid benefit is not, the study explains the drying-without-seeing pattern and gives clinicians a concrete switching criterion. If the vision benefit is uniform, that hypothesis is wrong, and we report it as such.
References
- Retinal and choroidal efficacy of switching treatment to faricimab in recalcitrant neovascular age related macular degeneration
- Efficacy of switching treatment to faricimab in recalcitrant neovascular age related macular degeneration – 6 month results after completion of the loading phase
- Real-World Data on Morphological and Functional Responses After Switching to Faricimab in Recalcitrant, Chronic Diabetic Macular Edema
- 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
- nnUNet RASPP for Retinal OCT Fluid Detection, Segmentation and Generalisation over Variations of Data Sources
- Automatic quantification of retinal photoreceptor integrity to predict persistent disease activity in neovascular age-related macular degeneration using deep learning
- Renal dysfunction associated with clinical response to intravitreal conbercept therapy for diabetic macular edema
- A machine learning model for predicting anatomical response to Anti-VEGF therapy in diabetic macular edema
- Estimating Optimal Dynamic Treatment Regimes Using Irregularly Observed Data: A Target Trial Emulation and Bayesian Joint Modeling Approach
- Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach
- Predicting Diabetic Macular Edema Treatment Responses Using OCT: Dataset and Methods of APTOS Competition
- Longitudinal Flow Matching for Trajectory Modeling
- Learning Treatment-Response Models from Multivariate Longitudinal Data