A single-file, offline web app for therapeutic drug monitoring (TDM) of three complement C5 inhibitors: eculizumab, ravulizumab and crovalimab. You enter a patient's weight, the doses they actually received and any measured serum concentrations. The app fits a published population pharmacokinetic model to that patient (Bayesian individualization), then forecasts the concentration–time curve, the trough before each upcoming dose, and the probability that the trough stays above a target.
Try it: https://robterheine.github.io/complementtdm/ — or download
complementtdm.html and open it in a browser. It works from file://
with no server, no network and no installation.
Research use only. Not a medical device. The models need prospective validation before routine clinical use. The app never suggests, recommends or optimizes a dose or dosing interval; it shows what the model predicts and the decision stays with the clinician.
- Individualizes a population PK model to one patient from their own samples (peaks, troughs, or both), including samples reported below the assay's lower limit of quantification.
- Forecasts the concentration–time curve with a 5–95 % credible band, and a table of the predicted trough before each dose with the median, the 5–95 % range and P(true trough > target), including a Monte-Carlo interval on that probability.
- Works without samples too: a population forecast from weight (and age or disease where the model uses them).
- Regimen explorer: reuses the fitted posterior to show the steady-state trough and P(> target) for other intervals or other doses, with no refit. It displays consequences only and never marks a regimen as preferred.
- Fit diagnostics: observed vs individual prediction, residuals, how much each individual parameter was actually informed by the samples (versus staying on the population prior), and sampler convergence.
- Prints an A4 report with the patient, dosing schedule, measurements, forecast, findings, your free-text advice and signature lines.
- Saves and restores sessions as JSON (and autosaves locally in the browser).
- Runs on a phone: charts switch to a genuinely different compact layout on narrow screens.
- Drug & patient: pick the drug, enter weight (plus age for crovalimab, disease for eculizumab) and, if you want, a different trough target.
- Doses & samples: enter the dosing history (full schedule) or, for a patient on a stable regimen, a dose + interval (steady state). Add measured concentrations. Optional: per-dose infusion duration, a planned regimen, LLOQ, an assay-error multiplier, and a flag to exclude a sample (for example one drawn after plasma exchange) while keeping it on record.
- Run forecast: chart, parameters, trough table and advisories.
- Explore regimen (optional): what a different dose or interval would do for this patient.
weight / age / disease ─┐
dosing history ─────────┼─► population PK model (ODE) ─► MAP-Bayes ─► Laplace ─► MCMC posterior ─┐
measured levels (+LLOQ) ┘ │
▼
troughs, P(> target), bands, half-life, diagnostics, regimen explorer, report ◄─ posterior draws
Models. Each drug is a self-contained specification (parameters, variability, residual
error, allometric scaling, hints) in src/model.js:
| Eculizumab | Ravulizumab | Crovalimab | |
|---|---|---|---|
| Structure | 2-compartment, parallel linear + Michaelis–Menten elimination | 2-compartment, linear | 3-compartment: SC depot → central ⇄ peripheral, linear |
| Source | ter Avest et al., Clin Pharmacokinet 2025;64:1139–1147 (doi:10.1007/s40262-025-01536-x) | Ultomiris FDA review, simplified (NDA 761108) | PiaSky FDA review, run05 (NDA 761388) |
| Covariates | weight; disease (PNH/aHUS) | weight | weight; age on the SC absorption rate |
| Variability | IIV on CL, V1; between-occasion CL in aHUS | IIV on CL, V1 (correlated) | IIV on CL, V1, Q, V2, ka, F (logit) |
| Routes | IV | IV | IV loading dose, then SC |
| Default trough target | 75 µg/mL | 150 µg/mL (≥ 100 minimum) | 75 µg/mL |
Time is in days, concentrations in µg/mL (= mg/L), doses in mg. The two FDA-review models are
transcriptions of the final NONMEM control streams shipped in this repository
(crovalimab.txt, ravulizumab.txt,
nonmemcontrolstream.txt); a model value is only ever changed by
re-transcribing its source.
Numerics. Concentrations come from an adaptive Dormand–Prince 5(4) ODE solver, with IV infusions and SC depot doses handled as events.
Individualization.
- MAP-Bayes: minimise
Σ (obs − pred)² / σ² + log σ² + ηᵀ Ω⁻¹ ηover the full random-effects vector (2 for eculizumab and ravulizumab, 6 for crovalimab). Residual variance follows the published model (σ² = add + prop · pred²); the assay-error multiplier scales it. Below-LLOQ samples use the M3 method,−2 log Φ((LLOQ − pred)/σ). Multi-start Nelder–Mead reports when the starts disagree. - Laplace approximation: the curvature at the optimum gives the proposal covariance
(
2 H⁻¹). - MCMC: adaptive random-walk Metropolis, four chains from over-dispersed starts (8 000 iterations; 10 000 for the 6-eta crovalimab posterior), about 2 400 kept draws. Every prediction is computed by pushing the posterior draws through the model, so uncertainty in the individual parameters carries through to the troughs.
- Diagnostics: split R̂ and split-chain bulk effective sample size per parameter, acceptance, and a per-parameter "informed by your data" measure. A fit that does not converge is shown as not reliable and its predictions are withheld.
What the headline probability means. P(true trough > target) is computed from the
latent concentration of each posterior draw. Assay (residual) error is deliberately left out,
and the table, tooltips and report say so, so a measured level can differ from the band.
Safeguards for real-world data entry. The app asks whether the patient was treated before the first entered dose (missing earlier doses make a patient look like a slow eliminator and the forecast falsely reassuring). During a drug's loading phase, or when the last two dosing intervals disagree, it does not guess the next interval: it shows no assumed next-dose trough and no projected doses until you enter a planned regimen. Dose, date, concentration and unit plausibility are checked, and imported session files are validated before use.
Reproducibility. The sampler seed is fixed, so the same inputs give the same numbers.
- Suggest, optimize or titrate a dose or interval.
- Model anti-drug antibodies (the crovalimab stream is the ADA-negative population).
- Model a patient who was switched from one drug to another: a session models one drug throughout.
- Validate the models in children, pregnancy or dialysis-dependent renal failure. Crovalimab refuses age < 12 years and flags 12–16; eculizumab accepts weights down to 3 kg, with a caution that the lowest weights rely on allometric extrapolation.
- Propagate uncertainty in the population parameters (only between-patient and posterior uncertainty).
- Send, store or sync anything. Nothing leaves the device; the only persistence is an optional local autosave and JSON export you control.
- Transcription:
node tools/verify_control_streams.mjsre-implements each NONMEM$DES/ADVANsystem independently and compares it with the app's model: maximum relative deviation ≤ 2.7 × 10⁻⁴ % for all three drugs. - Against NONMEM: for 14 simulated subjects across the three drugs (rich, sparse, trough-only,
single-trough and below-LLOQ designs), the app's MAP estimates agree with NONMEM
POSTHOCempirical-Bayes estimates to within 0.006 of NONMEM's own standard error (validation/results_table.md). - Calibration:
npm run calibrationruns a simulate-then-fit study (300 patients per scenario, five scenarios, the production MCMC). All five scenarios give 89.3–90.0 % coverage of the nominal 90 % trough interval with flat PIT histograms (docs/VALIDATION.md). - Test suite:
npm testruns 50 suites, about 1 670 assertions, dependency-free. New checks are shown to fail against the broken code before they are trusted.
Requires Node ≥ 16 for development only. The app itself has no dependencies.
npm test # full test suite (about 5 minutes)
node tools/verify_control_streams.mjs # app vs NONMEM control streams
npm run build # inline index.html + src/* into complementtdm.html
npm run calibration # the long simulate-then-fit calibration studyindex.html here is the build template; complementtdm.html is the generated single-file
app (never edit it by hand). The version lives in one place, src/version.js, and must match
package.json.
src/model.js drug specifications, ODE solver, simulation
src/bayes.js MAP, Laplace, MCMC, diagnostics, runFit, regimen explorer
src/chart.js dependency-free SVG charts
src/diagnostics.js fit-diagnostic panels
src/ui.js forms, session handling, tables, report
src/app.css styles (including phone layout and print)
tests/ 50 suites (Node, no install)
tools/ control-stream cross-check, calibration study, golden-vector generator
validation/ NONMEM post-hoc comparison, run files and report
docs/ developer guide, validation write-ups, audit and review notes
The developer guide (docs/DEVELOPER_GUIDE.md) covers the
architecture, public APIs, conventions and the reasoning behind each release.
The code is released under the MIT License. The pharmacokinetic models are transcriptions of published or regulatory sources listed above; please cite those sources when you use the corresponding model. Trough targets are taken from PMID 35238929 (eculizumab), PMID 33512711 (ravulizumab) and PMID 39835421 (crovalimab). The eculizumab article itself is not redistributed here.
Rob ter Heine, Radboud Applied Pharmacometrics, Radboudumc, Nijmegen — r.terheine@radboudumc.nl