DME Treatment Explorer

Given a patient's OCT history and current treatment, this model estimates how a typical responder on that medication would be expected to behave, based on the pharmacodynamic profile derived from each drug's pivotal trial. It then adjusts for individual response by comparing observed CST to what the model would have predicted at each visit. For treatment-naive patients, it uses population-average projections. In principle, this framework could quantify extension probability at each dosing interval, identify below-average responders, and compare expected outcomes across medications before a switch. It is intended as a proof of concept and starting point for future modeling efforts, not a clinical decision tool.

Examples:

First visit (no prior anti-VEGF)

Disease Baseline

Peak historical CST (µm)

The worst-ever OCT for this eye. Defines the treatable edema ceiling. Update only if a new peak is recorded.

Current Visit

Drug

Interval (wk)

Today's CST (µm)

Previous Visits (optional, improves accuracy)

New Treatment (today's plan)

Drug

Follow-up (wk)

Extension Probability Profile

Probability of CST ≤ 325 µm at trough for each interval. Based on predicted CST with ±1 SD uncertainty scaled to edema severity. Longer bars = safer to hold at that interval.

Single-Cycle CST Projection

How the model works

Start with the edema. Peak historical CST minus 270 µm (normal) = treatable fluid.

The drug removes a fraction. Peak effect decays exponentially. The effective half-life (72–88 days) captures clinical effect duration, not molecular clearance.

Calibrate to your patient. Observed CST vs expected CST reveals individual responsiveness. More visits = tighter estimate.

Probability, not prediction. At each interval, the model computes the likelihood of staying dry (≤ 325 µm). You decide what risk is acceptable.

Compare scenarios. Switch drugs, change intervals, add visits. The probability profile updates in real time.

Does it match real trials?

For each trial, the model estimates annual injection burden using trial-reported baselines and protocol-specific loading regimens at pScale = 1.0. Published values are compared to model predictions below.

Trial protocols differed from each other and from the model's framework. Protocol T used PRN retreatment; YOSEMITE used personalized treatment intervals (PTI); PHOTON used fixed intervals with shortening for disease activity. The model predicts for a typical responder (pScale = 1.0); trial means include non-responders. This is face validity, not independent validation.

References

¹ Wells JA, et al. Aflibercept, bevacizumab, or ranibizumab for DME. N Engl J Med. 2015;372(13):1193–1203. (DRCR Protocol T)
² Wykoff CC, et al. Efficacy, durability, and safety of faricimab with extended dosing up to every 16 weeks in DME. Lancet. 2022;399(10326):741–755. (YOSEMITE/RHINE)
³ Brown DM, et al. Intravitreal aflibercept 8 mg in diabetic macular oedema (PHOTON): 48-week results. Lancet. 2024;403(10432):1141–1152.
⁴ Dayneka NL, Garg V, Jusko WJ. Comparison of four basic models of indirect pharmacodynamic responses. J Pharmacokinet Biopharm. 1993;21(4):457–478.
⁵ Stewart MW, et al. Pharmacokinetic rationale for dosing every 2 weeks versus 4 weeks with intravitreal ranibizumab, bevacizumab, and aflibercept. Retina. 2012;32(3):434–457.
⁶ Diabetic Retinopathy Clinical Research Network. Reproducibility of spectral-domain OCT retinal thickness measurements in DME. JAMA Ophthalmol. 2014;132(9):1113–1122.
⁷ Starr MR, et al. Fluctuations in central subfield thickness associated with worse visual outcomes in patients with DME. Am J Ophthalmol. 2021;232:98–108.

▸ Detailed formulas & parameters

The Core Equation

CST(t) = 270 + Excess × (1 − Effect(t))

Normal CST: 270 µm

Structural floor. Midpoint of 250–300 µm dry range.

Excess = Peak CST − 270

Treatable edema. All drug effects operate on this.

Structural thickening (ERM, VMT) will cause overestimation.

Effect(t) = pScale × eff × e^{−0.693t/effHL}

Monoexponential clinical effect decay. effHL (72–88d) is a PD fitting parameter, NOT molecular vitreous half-life.

Absorbs drug clearance, VEGF rebound, BRB repair, and inflammatory recovery into one number.

Model derivation

This equation is a simplified indirect response model adapted for OCT-measured CST. The general framework for indirect pharmacodynamic response (IDR) models was established by Dayneka, Garg, and Jusko,⁴ who described how drugs that inhibit or stimulate the production or dissipation of a biological response can be modeled with first-order kinetics. Stewart et al.⁵ applied monoexponential binding-activity decay to compare anti-VEGF dosing intervals. This tool uses a similar monoexponential decay structure but substitutes VEGF binding activity with a directly observed clinical endpoint (CST), and replaces molecular vitreous half-life with an "effective half-life" that captures the full PK/PD chain from drug clearance to clinical edema recurrence. The result is an algebraic approximation of the full IDR differential equations, solved at trough (the time of the next clinic visit).

⁴ Dayneka NL, Garg V, Jusko WJ. Comparison of four basic models of indirect pharmacodynamic responses. J Pharmacokinet Biopharm. 1993;21(4):457–478.
⁵ Stewart MW, et al. Pharmacokinetic rationale for dosing every 2 weeks versus 4 weeks with intravitreal ranibizumab, bevacizumab, and aflibercept. Retina. 2012;32(3):434–457.

Drug Parameters

Drug	Peak Eff.	Eff. t½	Source
Faricimab 6mg	1.30	88d	YOSEMITE/RHINE
Aflibercept 8mg	1.50	75d	PHOTON
Aflibercept 2mg	1.07	75d	Protocol T
Ranibizumab 0.3mg	1.05	72d	Protocol T
Bevacizumab 1.25mg	0.80	75d	Protocol T

Peak efficacy = max fractional CST reduction. Aflibercept 8mg shares 2mg half-life (same molecule); longer duration from higher peak (4× dose).

Uncertainty Model

SD = 15 + 0.08 × (predicted CST − 270)
P(dry) = Φ((325 − predicted CST) / SD)

This SD represents total trough prediction uncertainty: OCT measurement noise, biological day-to-day fluctuation in edema, and model simplification error combined.

The 15 µm floor is anchored to published OCT reproducibility data. The DRCR.net reproducibility study⁶ reported a Spectralis coefficient of repeatability (CR) of ~7% for CST in DME eyes, corresponding to a per-measurement SD of ~7–10 µm for a dry macula. The 15 µm floor adds a margin for biological day-to-day variation on top of instrument noise.

The 0.08 proportional coefficient captures the observation that edematous maculae fluctuate more than dry ones. A post-hoc analysis of DRCR Protocols T and V⁷ found mean visit-to-visit CST SD of 56 µm in Protocol T (baseline ~412 µm) and 37 µm in Protocol V (lower baseline severity), consistent with severity-dependent variability. The specific coefficient 0.08 was chosen to produce plausible probability profiles across the CST range; it has not been empirically calibrated against patient-level trough data and should be treated as a tunable assumption.

⁶ Diabetic Retinopathy Clinical Research Network. Reproducibility of spectral-domain OCT retinal thickness measurements and conversion to equivalent time-domain metrics in DME. JAMA Ophthalmol. 2014;132(9):1113–1122.
⁷ Starr MR, et al. Fluctuations in central subfield thickness associated with worse visual outcomes in patients with DME in clinical trial setting. Am J Ophthalmol. 2021;232:98–108.

Parameter Calibration

Each drug's two parameters (peak efficacy eff and effective half-life effHL) are jointly calibrated against published trial data using a two-constraint approach.

Step 1: Peak Efficacy from Loading-Phase CST

At the end of the loading phase, the trough CST constrains the peak effect. The observed fractional excess removal at trough is:

f_obs = (baseline − CST_trough) / (baseline − 270)

Since this is measured at trough (typically 28 days post-injection), the decay has already acted. Solving for peak efficacy:

eff_initial = f_obs / e^{−0.693 × 28 / effHL}

Values > 1.0 indicate the drug can transiently reduce CST below the 270 µm structural floor, consistent with clinical observations of subfoveal thinning after anti-VEGF loading.

Step 2: Effective Half-Life from Injection Burden

The effective half-life is calibrated so that the model's predicted annual injection count matches the published trial value. For each candidate effHL, the model computes the equilibrium interval (longest interval where P(dry) > P(recurrence)) and derives:

inj/yr = loading_inj + (52 − loading_wks) / equilibrium_interval

The effHL that minimizes |model − published| injection count is selected. Since eff depends on effHL (Step 1), both parameters are co-determined iteratively.

Pharmacological Constraint

Same molecule → same effHL. Aflibercept 2mg and 8mg share effHL = 75 days because the same fusion protein implies the same pharmacodynamic decay kinetics. The 4× dose increase manifests only as higher peak efficacy (1.50 vs 1.07), not a longer decay rate. This is consistent with PHOTON design rationale: higher doses extend effect duration by starting from a higher peak, not by altering clearance.

Calibration Summary

Drug	Source Trial	Baseline µm	eff	effHL days	Pub. inj/yr	Δ
Aflibercept 2mg	Protocol T	412	1.07	75	9.2	-0.4
Ranibizumab 0.3mg	Protocol T	412	1.05	72	9.4	-0.6
Bevacizumab 1.25mg	Protocol T	412	0.80	75	9.8	+0.9
Faricimab 6mg	YOSEMITE	491	1.30	88	~7	+0.6
Aflibercept 8mg	PHOTON	480	1.50	75*	~6.3	+0.0

* Constrained to aflibercept 2mg value (same molecule). Only eff is free.

Calibration limitations. Each drug's parameters are fit to aggregate trial means, not individual patient data. The calibration data (loading trough CST and injection burden) are reported at the arm level. Different trial designs (PRN in Protocol T, PTI in YOSEMITE, fixed intervals in PHOTON) introduce protocol-specific biases that the model absorbs into its drug parameters. Effective half-life is not independently measurable. It is a summary parameter that combines molecular clearance, VEGF rebound kinetics, blood-retinal barrier repair, and inflammatory resolution into a single exponential decay rate.

▸ Limitations

Not validated. Not tested against individual patient outcomes.

Effective half-life ≠ molecular half-life. A PD fitting parameter, not a measured quantity.

No systemic factors. HbA1c, blood pressure, renal function not modeled.

CST ≠ vision. DRCR data shows modest, inconsistent CST–VA correlation.

Structural disease. ERM, VMT, fibrosis treated as fluid. Overestimates treatable component.

Cross-drug pScale. Assumes responsiveness carries across mechanisms. May not hold for dual-pathway agents.

Purpose

Diabetic macular edema is the leading cause of vision loss in working-age adults with diabetes. Treatment requires repeated intravitreal injections, often for years, placing a substantial burden on patients, caregivers, and health systems. The decision of which drug to use and how frequently to give it is shaped by a growing number of anti-VEGF agents with distinct pharmacologic profiles, each studied under different trial designs, dosing protocols, and patient populations. Direct comparisons across all clinically relevant scenarios are not feasible through randomized trials alone, and the potential expansion of DME treatment to include multimodal approaches such as fenofibrate will only increase the combinatorial complexity.

These conditions make DME an ideal candidate for pharmacodynamic modeling. A well-calibrated model can synthesize published trial data into a common quantitative framework, allowing clinicians to explore patient-specific scenarios that no single trial could address. This tool is a proof of concept toward that goal. It does not replace clinical judgment, and its predictions have not been prospectively validated. My hope is that it demonstrates the feasibility and potential value of this approach, and that it encourages further efforts to bring pharmacokinetic and pharmacodynamic reasoning into routine DME care.

What this is not

Not a validated clinical tool. Not a medical device. Parameters derive from published trials with varying designs. Use clinical judgment.

Future directions

Adjunctive therapy modeling (topical NSAIDs, fenofibrate, dexamethasone implants). Stochastic T&E simulation. Dose-response modeling connecting molecular PK to effective half-life.

Author

Henry Bison, MD, Ophthalmology Resident PGY3, University of Maryland. Independent academic project.

[email protected]

Disclaimer

Provided "as is" without warranty.

Experimental research tool. Not a medical device. No patient data collected or transmitted. All computation runs in your browser.