Diffusion Run
seed: 9612
Denoising convergence curve
Observed control N: 120
Borrowed ESS: 65
Estimated control reduction: 29.8%
prototype · diffusion mock
Ophthalmology Digital Twin · Diffusion Synthetic Control (Mock)
Mock diffusion pipeline for synthetic control arm generation: latent noise -> denoising steps -> calibrated patient-level output.
Setup
Endpoint: BCVA change at Week 24 (ETDRS letters) · Historical pool: AMD-118, AMD-122, AMD-OBS-19 (680)
Observed vs Synthetic
BCVA change: observed 2.10 ± 7.40 | Synthetic arm 2.52 ± 4.39 ·Δ +0.42
CST change: observed -32.0 ± 58.0 | Synthetic arm -30.7 ± 35.4 ·Δ +1.3
Treated arm vs Synthetic arm Distribution alignment: BCVA 58.9 / 100 · CST 64.7 / 100
Mock method note
This is a product mock: generated patients are simulated with a deterministic pseudo-random denoising loop, then calibrated to scenario-level control moments.
Treated arm vs Synthetic arm · Distribution alignment
Histogram overlay
Treated armSynthetic arm
Q-Q plot
x-axis: Treated arm quantiles · y-axis: Synthetic arm quantiles
External Control Methodology Pipeline
Step 1: Data cleaning
Raw N: T 120 / C 120
After cleaning: T 115 / C 115
Step 2: Cohort definition
Cohort eligible: T 115 / C 115
Step 3: Propensity score / weighting / matching
Balance before (mean |SMD|): 0.299
Balance after (mean |SMD|): 0.313
Step 4: Primary endpoint estimation
Primary effect (treated - external control): 5.45
95% CI: [3.98, 6.92]
Step 5: Sensitivity analysis
n(T/C): 69 / 61
| Scenario | Effect | Delta vs primary | Interpretation |
|---|---|---|---|
| Strict cohort window | 5.44 | -0.02 | Stable |
| More aggressive PS trim | 0.09 | -5.36 | Moderate shift |
| Hidden bias stress (Gamma-like) | 4.65 | -0.8 | Robust |
Conditional Twin Generator (X→Y(t))
Baseline condition X
age 70.9 · Sex F · baseline BCVA 60.5 · baseline CST 379 · Lesion type occult · Prior injections 4
Counterfactual control trajectory
Mean with 95% prediction interval · K=50
Synthetic Patient Samples
| ID | Age | Base BCVA | Base CST | Duration (y) | BCVA Δ | CST Δ | Weight |
|---|---|---|---|---|---|---|---|
| SYN-001 | 73.1 | 60.7 | 381 | 3 | -7.3 | 27 | 0.86 |
| SYN-002 | 67.9 | 56.8 | 389 | 3 | 2.9 | -43 | 0.84 |
| SYN-003 | 71.7 | 49 | 497 | 2.8 | -4.2 | -20 | 0.81 |
| SYN-004 | 81.3 | 47.3 | 368 | 3.2 | 3.2 | -45 | 0.79 |
| SYN-005 | 76.3 | 55.1 | 365 | 3.4 | 5.5 | -40 | 0.82 |
| SYN-006 | 75.2 | 61.2 | 404 | 2.4 | 3.4 | 60 | 0.84 |
| SYN-007 | 76.3 | 50.9 | 501 | 1.5 | 4.9 | -60 | 0.83 |
| SYN-008 | 72.8 | 62.6 | 373 | 4.2 | 0.5 | -56 | 0.83 |
| SYN-009 | 83.3 | 62.2 | 423 | 3.6 | 6.5 | -91 | 0.85 |
| SYN-010 | 77.1 | 61.1 | 383 | 3.6 | 2.3 | -34 | 0.82 |
| SYN-011 | 68.5 | 54.8 | 383 | 4.9 | 3.5 | -41 | 0.81 |
| SYN-012 | 70.3 | 57.7 | 446 | 3.3 | -2.2 | 15 | 0.85 |
update log
Prototype Change Log
2026-03-06 · v0.4.0
Conditional Twin Generation (X→Y(t))
- - Added patient-level baseline conditioning fields: age, sex, baseline BCVA/CST, lesion type, prior injections.
- - Added patient picker and K-counterfactual generation flow (default 50 twins) for selected treated patient.
- - Added trajectory output with mean and 95% prediction interval confidence band.
2026-03-05 · v0.3.0
External Control Methodology Pipeline
- - Added end-to-end methodology pipeline: data cleaning -> cohort definition -> PS weighting/matching -> primary endpoint estimation -> sensitivity analysis.
- - Added configurable analysis controls for IPTW vs PS matching, caliper, and propensity trimming.
- - Added balance diagnostics (mean |SMD| before/after), treatment effect with 95% CI, and sensitivity scenario table.
2026-03-05 · v0.2.0
Distribution Alignment + CSV Export
- - Added treated-arm vs synthetic-arm distribution alignment panel.
- - Added histogram overlay and Q-Q plot for BCVA/CST metric toggles.
- - Added CSV export buttons for alignment dataset and synthetic patient-level table.
2026-03-05 · v0.1.0
Ophthalmology Diffusion Twin (Mock) Launch
- - Added a one-page mock prototype for synthetic control arm generation with diffusion-style denoising loop.
- - Included ophthalmology scenarios (nAMD, DME, GA) with configurable cohort size, denoising steps, and calibration strength.
- - Added side-by-side observed-vs-synthetic metrics and patient-level synthetic sample table for review.