Vancomycin Renal Feedback 100k

Published Oct 7, 2026·1523 views·1 inquiries

At a glance

Licence

ai.market Standard Data Licence v1.1

ai.market updated the licence shown here. The seller approved the original listing.

Vancomycin clears through the kidneys — and damages them Vancomycin is excreted by the kidneys and injures the tubules, so exposure acts on its own elimination: dose → exposure → tubular damage → falling clearance → rising exposure. In every published population PK model and dosing program reviewed, renal function is a covariate fed in from outside. Here it is a state variable: the loop is closed inside the equations.

The cohort is 100,000 virtual adult inpatients whose baseline creatinine, ICU share, age and treatment duration match those of 128,993 real vancomycin-treated patients (Yu et al. 2021, AJHP). Each patient is simulated for up to 20 days of therapy and 45 days of follow-up. Two calibrated scenarios bracket the strength of vancomycin nephrotoxicity: scenario A is conservative, scenario B pushes the mechanism to its upper limit. The same patients, the same random streams — only the damage rate differs.

Every patient is recorded twice Observed (what the hospital sees) Truth (what actually happened) Noisy creatinine samples at 12-, 24-, 48- or 72-hour intervals, four laboratory regimes True hourly filtration (GFR) and vancomycin concentration Vancomycin trough and peak levels within the registered immunoassay range (4–80 mg/L) True daily and course AUC — the number no assay can measure Doses actually administered, KDIGO injury flag from observed creatinine Cause of each injury: vancomycin or competing. Detection lag of the clinical flag Creatinine-based clearance estimate used by the dose controller Error of three dosing estimators, patient by patient, in three time windows No real hospital archive contains the right column. Here it exists because every value was computed from a mechanistic model with 242 constants, each traced to a published source.

The same patients re-simulated under other conditions Without dose adjustment — what would have happened if the controller never reduced the dose. Without vancomycin toxicity — isolates the loop-induced bias in the dosing estimator. With intact kidneys — the exposure each patient would have had without any injury at all. Two monitoring regimes on the same patients — trough-guided vs. AUC-guided dosing, three emulated cohorts of 20,000 each, same noise. Nine tables per scenario Table Side Rows (A / B) Cols What it holds cohort observed 100,000 / 100,000 15 Baseline covariates and assigned regimen observed observed 1,802,987 / 1,333,673 10 Creatinine and vancomycin samples, dose, clearance estimate, injury flag observed_outcomes observed 100,000 / 100,000 15 Per-patient outcomes from observed data only truth truth 100,000 / 100,000 59 True parameters, exposure, estimator error, injury cause truth_traj truth 2,405,000 / 2,405,000 7 Hourly true states for 5,000 patients truth_counterfactual truth 100,000 / 100,000 17 Same patients without dose adjustment truth_noloop truth 100,000 / 100,000 9 Same patients without vancomycin damage truth_exposure_daily truth 100,000 / 100,000 41 True daily AUC with injury and with intact kidneys regime_arms truth 60,000 / 60,000 20 Paired trough/AUC arms, three emulated studies The model State variables: vancomycin mass in central and peripheral compartments, vancomycin-attributable damage Dv, competing damage Dc, serum creatinine. Integration: fourth-order Runge–Kutta, step 0.2 h. Eight physical invariants checked at every step.

C1 = A1 / V1 GFR(t) = max( GFR_0(t) · (1 − Dv) · (1 − Dc), 1 mL/min ) CL(t) = CL_nr + fr · GFR(t) · 0.06

dA1/dt = R_inf(t) − CL·C1 − Q·(C1 − A2/V2) dA2/dt = Q·(C1 − A2/V2) dDv/dt = k_dam · γ·C1 / (IC50 + γ·C1) · (1 − Dv) − k_rec · Dv dDc/dt = k_dam_c · (1 − Dc) − k_rec · Dc dCr/dt = ( G − GFR(t) · 0.06 · Cr ) / (0.6 · weight) Non-renal clearance 1.93 L/h, renal fraction 0.40 of filtration; central volume 36.3 L per 82 kg; peripheral 36.8 L, inter-compartmental clearance 2.66 L/h; damage follows Emax by tubular concentration (40–50× plasma), IC50 6.06–8.8 mg/mL from HK-2 cell studies; creatinine balance with Vd = 0.6 × body weight.

Key findings Quantity Scenario A Scenario B Injury flagged, % of patients 19.8 20.6 Vancomycin as primary cause / competing cause 4,570 / 15,188 18,452 / 2,198 Injury without vancomycin channel, % 14.3 3.4 Naive slope inflated vs. causal, × 5.0 3.1 Reverse arm share (kidney → drug), % 35.5 27.3 Detection lag, median (IQR), hours 14 (3–35) 12 (1–28) Causal slope, random dose as instrument (95% CI) 1.44 (1.15–1.82) 3.17 (2.82–3.51) The exposure-mediated injury difference between monitoring regimes is bounded at 0.8–1.9 pp (scenario A) to 1.8–4.1 pp (scenario B) and at most 5.5 pp — against published observational differences of 13.7 and 17.2 pp.

Validation 113 structural checks, 0 failed — nulls, schema, key integrity, flag recomputation, exposure sums, observed/truth column separation. Bit-for-bit reproducibility — all 18 tables regenerated from the published generator, SHA-256 match confirmed. Calibration bands passed on the reference arm (therapy ≥ 5 days): injury rate 17.8% / 19.1%, creatinine rise ratio 1.86 / 2.10, time to injury 4.0 / 4.0 days, recovery 11.0 / 12.0 days. External validation against three published regimen comparisons the model was never fitted to — six criteria recorded before any run. Scenario A passes 3/6, scenario B passes 5/6. The null study is reproduced in both. The long-therapy study is not reproduced by either — that is the ceiling result. Benchmark — injury prediction at 48 h, 5-fold CV, observed tables only: logistic regression AUROC 0.745 (A) / 0.773 (B); XGBoost 0.767 / 0.797. Published on real patients: 0.735 (0.638–0.833) for logistic regression. Companion products and preprint An accompanying preprint describes the model, calibration, both scenarios, the three main findings and the external validation in full — it is the scientific paper behind this dataset and will be posted to a preprint server. The dataset is the companion data release of that preprint.

A generator — the complete Python source code of the model (17 modules) — is available as a separate product. It reproduces every table bit for bit with one command (python run.py --scenario both --verify) and can build new cohorts of any size from the same 36 calibrated parameter sets. ~25 minutes for both scenarios on a single CPU core, 4 GB RAM. The generator comes with its own passport, constants audit and full documentation.

Limitations Data are synthetic. No dosing recommendations follow, and no AUC value in the dataset is a diagnostic threshold for injury. Tubular concentration is a fixed multiple of plasma: no renal cortical tissue compartment. Non-creatinine chromogen interference with the Jaffe reaction is not modeled. The dose controller acts on the creatinine-based clearance estimate; it does not use measured vancomycin concentrations. The competing channel does not switch off once started. Scenario B is an upper bound: it overstates the published dose-response increment and the vancomycin share in the randomized comparison. Provenance 57 peer-reviewed sources, 242 audited constants. Every constant is listed in constants_audit.csv with author, year, DOI/PMID, search queries and verification date. Every source is listed in literature_registry.csv with a working link. Specification hash (SHA-256): 08d02bab628c0b1f93b41088c5a11e725709a717d68c09fb1ca9f40ca2be29a8. Seed 20260929.

Original seller-approved version
Healthcare & Life Sciencessynthetic-datapharmacokineticsvancomycintherapeutic-drug-monitoringTDMnephrotoxicityacute-kidney-injuryAKIrenal-functionprecision-dosingODE-modelin-silicoclinical-pharmacologypharmacometricsAUC-estimationcausal-inferenceparquetclosed-loop-simulationdose-responseKDIGO

The seller can request a verified scan of this data.

ai.market standard terms

$390.00

1active

Published Oct 7, 2026 · 186.77 MB · 1 objects

Download window: 7 days after purchase.

Publisher

ai.market Seller

Stripe connected

No completed purchases yet

No ratings yet

Ask a Question

allAI checks current public information to answer questions about this listing.

Non-Custodial

Purchased files go directly from seller to buyer. We can store and show a public sample only when the seller chooses to publish it.