Calibrated within-population patient individualisation of antimicrobial exposure: a pre-registered simulation and external-evaluation study

Technical Summary: Calibrated Within-Population Patient Individualisation of Antimicrobial Exposure

Problem Statement
The dosing of narrow-therapeutic-index antimicrobials (e.g., vancomycin, beta-lactams) in critically ill patients is challenged by high inter-patient variability in drug clearance, driven by factors such as augmented renal clearance, organ support, and fluid shifts. Fixed, label-based dosing often fails to achieve therapeutic targets, leading to treatment failure, resistance selection, or toxicity. While Model-Informed Precision Dosing (MIPD) or Bayesian forecasting is the established remedy, contemporary implementations frequently suffer from a lack of honest uncertainty quantification. Systems often return confident but incorrect estimates, failing to represent the limits of their own knowledge. The core problem addressed is not the novelty of the dosing arithmetic, but whether a specific engine can perform within-population individualisation with calibrated uncertainty, transparency, and honesty, particularly when drawing priors from a compiled medical-knowledge substrate.

Methodology
This study employed a pre-registered, cryptographically committed simulation and external-evaluation design to validate a specific individualisation engine. The protocol, pass/fail criteria, and sample sizes were fixed before any data generation.

• Validation Battery: The study evaluated eight published population pharmacokinetic (PK) priors across three antimicrobials (vancomycin, piperacillin–tazobactam, piperacillin) and diverse critical-care populations (neurosurgical, paediatric, ECMO, CKRT, augmented renal clearance). Each prior was characterized by its between-subject coefficient of variation for clearance ($\omega_{CL}$).
• Simulation Design:
  • Recovery Analysis: 500 simulated patients per prior (4,000 total) were used to test the recovery of individual clearance parameters from sparse, two-sample concentration data.
  • External Evaluation: Data were simulated from one adult vancomycin model and forecast using a different adult vancomycin prior to test cross-model prediction improvement.
• Engine Mechanics: The engine updated only the patient's clearance (holding volume of distribution and inter-compartmental clearance at population-typical values) using Maximum-A-Posteriori (MAP) estimation with a local-Gaussian (Laplace) approximation. Exposure metrics (AUC/MIC or %fT>MIC) were propagated via Monte Carlo sampling.
• Metrics and Criteria:
  • Bias: Mean error of recovered clearance on the log scale, required to be $\le 0.15 \cdot \omega_{CL}$.
  • Imprecision: Root-mean-square error (RMSE), required to be $\le 0.6 \cdot \omega_{CL}$ (representing ~64% variance explained).
  • Calibration: Empirical coverage of posterior credible intervals at 50, 80, 90, and 95% levels, required to fall no more than 10 percentage points below the nominal level.
  • External Prediction: Individualised forecasts were required to reduce RMSE to $\le 0.80$ times the population-typical forecast without worsening mean prediction error.
• Controls: The study included negative controls (e.g., drugs without priors, uninformative measurements) and fact-checking against external references to verify the engine's ability to refuse unsafe or unsupported estimates.

Key Results

• Calibration: The engine achieved calibrated coverage across all eight priors at all four nominal levels. Empirical coverage ranged from 47.8% to 54.4% at the 50% level and 92.2% to 97.6% at the 95% level, consistently meeting the pre-registered lower bound (nominal − 10 percentage points).
• Parameter Recovery:
  • Bias: Seven of eight priors met the bias criterion. The single exceedance occurred in the Thy paediatric prior (bias = −0.080 vs. bar of 0.078), a result attributed to the expected shrinkage of MAP estimates under sparse sampling in high-variability models.
  • Imprecision: Five of eight priors met the stringent imprecision criterion. Three exceedances occurred in the Wei, Hahn, and Thy priors. The authors note these failures are structural consequences of the sparse two-sample design interacting with either very tight priors (low variability to resolve) or high residual noise, rather than a deficiency in the update algorithm.
• External Evaluation: In the cross-model vancomycin test, individualisation reduced prediction error by 32% to 52% (RMSE ratios of 0.683 and 0.477) compared to the population prior, without introducing bias.
• Transparency and Controls: The engine successfully refused to generate numeric doses for drugs without priors (e.g., gentamicin) and for priors with unextractable error structures. It correctly flagged model-coverage gaps in demonstration patients (e.g., a patient on renal replacement matched to a prior not covering that modality) rather than concealing them.
• Comparison with LLMs: A frontier language model without the pharmacokinetic engine failed to produce calibrated exposure posteriors or abstain when data were insufficient, highlighting that the safety properties are intrinsic to the engine's architecture and substrate, not general language capabilities.

Significance and Claims
The paper claims to demonstrate that an individualisation engine, drawing priors from a compiled medical-knowledge substrate, can perform calibrated within-population individualisation with honest uncertainty quantification. The study does not claim the dosing arithmetic (Bayesian updating) is novel, as it is standard and implemented in commercial systems. Instead, the significance lies in:

1. Provenance and Discipline: The use of a compiled substrate for priors and clinical facts, validated under a pre-registered, cryptographically locked protocol with byte-for-byte determinism.
2. Safety-Critical Behavior: The engine's ability to provide calibrated credible intervals and to abstain or flag when inputs are insufficient or models are mismatched, rather than guessing.
3. Scope of Validation: The results bound the claims for the current scale of the substrate compilation. The study validates the method and the engine's behavior using published models, not the specific content of the substrate's future expansions.

The authors explicitly state that prospective validation against raw patient-level data is the primary outstanding step and that generalisation across populations (e.g., paediatric to adult) was not addressed. The observed exceedances in bias and imprecision are reported as pre-specified consequences of the estimator and sampling design, not as failures of the system's logic.