Chorus
Chorus is a vancomycin simulator that fits 45 published population PK models at once, making it possible to compare model divergence and aggregate predictions.
For Educational and Research Use Only. Chorus is an experimental simulator and is not a validated clinical tool. Do not use it for patient care or clinical decision-making. See the Terms.
Why Chorus
Model-informed precision-dosing (MIPD) tools typically bet on a single population model, chosen up front. Chorus explores a different question: what happens if you fit many published models at once and average them? It fits up to 45 models to the same patient data and shows where the models agree, where they diverge, and which ones the data supports.
Privacy
Chorus runs entirely in your browser. Patient files are read from and written to a folder you choose on your own disk — nothing is ever sent to a server.
Workspaces are de-identified by default, following HIPAA Safe Harbor, so the app can run with no protected health information on disk. See Privacy & security for the full policy.
What Chorus does
Given a patient's demographics, dosing history, measured vancomycin concentrations, and serum creatinine measurements, Chorus:
- Runs every selected population PK model (up to 45 active models) in parallel, producing a predicted concentration–time curve for each model.
- Fits each model to the patient using Bayesian maximum-a-posteriori (MAP) estimation against the observed drug levels.
- Combines the models into a single prediction by model averaging.
- Computes error metrics (RMSD, SSE, bias) for every model and for the aggregate.
- Detects acute kidney injury (AKI) from the creatinine time series using KDIGO-style criteria.
- Searches candidate regimens — evaluating dose/interval combinations against configurable targets (AUC/MIC, trough thresholds).
A separate machine-learning layer can predict, before any levels exist, which models are likely to fit a given patient best — useful for scenarios where there is nothing yet to fit against.
45 active population PK models, run in parallel
One / two / three-compartment analytic solutions
Creatinine interpolated across the timeline
Optional steady-state initial conditions
Bayesian MAP estimation, L-BFGS optimizer
Combined additive + proportional residual error model
Weighted observations, explicit below-limit-of-quantification handling
Equal-weight, SSE-weighted, or SSE + overfitting penalty
RMSD / SSE / bias per model and aggregate
Dose / interval search against AUC/MIC and trough targets
Probability of target AUC, toxic trough, and effective trough
Or evaluate a specific regimen of your own directly
KDIGO-style detection from the serum-creatinine series
Predicts which models fit a patient best before any levels exist
Exact SHAP explanation for every prediction
Trainable on your own patients, entirely on your machine
Numerical core written in Rust, compiled to WebAssembly
Full 45-model fit in a fraction of a second
Workspace files written to a folder you choose, via the File System Access API
Nothing sent to the server, under any configuration
De-identified by default — HIPAA Safe Harbor, as if every input were PHI
Relative dates, ages over 89 capped to "90+"
Cooperative workspaces — a team can share one folder on a shared drive
Per-patient file locks with a heartbeat, so concurrent sessions never collide
Workspace-wide report of how each model performs across your patients
Chrome or Edge on desktop — requires the File System Access API
Nate Van Veldhuizen, PharmD