Mean Exposure Does Not Validate Target Attainment Predictions in Children: A Comment on the Study by Wu et al [Letter]

Dear editor

Wu et al developed a physiologically based pharmacokinetic model of posaconazole oral suspension and used age-specific food simulations to propose meal conditions for pediatric hematology patients.1 Their reported mean absolute prediction error of 12.3% supports agreement for the concentration summaries evaluated. The more consequential inference is the probability of achieving a prophylactic trough concentration under each meal condition. That probability requires validation of the concentration distribution, particularly around the decision threshold, rather than agreement of its mean alone.

The distinction is apparent within the report. The methods define prediction error by comparison with literature-reported mean plasma concentrations when peak and area-under-the-curve data are unavailable. The limitations then acknowledge that simulated 90% intervals were wider than observed variability, suggesting overestimation of absorption-related variation, but state that population-level accuracy is preserved. This may be defensible for an adequately predicted population mean. It does not follow for the proportion below 700 ng/mL or above the modeled upper threshold of 3750 ng/mL. Distributions with the same mean can place substantially different fractions of patients on either side of those cutoffs.

This matters because the meal recommendations depend on those fractions. At 6 mg/kg three times daily with a moderate-fat, moderate-calorie meal, predicted attainment was 58.2%, 67.8%, 75.8%, and 90.8% across the four age groups. For children aged 2–6 years, a high-fat, high-calorie meal increased predicted attainment to 91.6%, supporting the statement that this meal was necessary. A wider-than-observed distribution could alter whether a regimen crosses a proposed reliability threshold, even when its mean exposure is accurate. The direction cannot be determined from interval width alone because it depends on the location of the cutoff and the shape of the concentration distribution.

A fit-for-purpose evaluation would compare observed individual troughs with the corresponding predictive distribution in an independent pediatric data set. For each age and meal stratum, the authors could report observed versus predicted lower-tail proportions, central quantiles, and interval coverage, rather than only whether observed means fall within a broad interval. Where repeated trough measurements are available, between-patient variation should be separated from within-patient variation and measurement error. Qualification guidance for PBPK applications emphasizes matching model evaluation to the intended use and assessing whether virtual populations reproduce relevant observed characteristics.2 Here, the intended use makes threshold calibration an essential performance measure.

Parameter uncertainty also needs to be distinguished from variability among virtual individuals. Simulating 500 individuals reduces Monte Carlo error conditional on the chosen parameter distributions; it does not establish that those distributions are correct for children with hematologic disease. Repeating the target-attainment analysis across clinically plausible absorption and variability parameters would show whether the proposed age-specific meal categories are stable. The validation should record actual meal intake and relevant co-medications because standardized meals and omitted interaction workflows are already acknowledged limitations. This is an extension of the authors’ own caveats, not a claim that these factors were overlooked. Pediatric oncology PBPK guidance emphasizes age-specific physiology, formulation and sampling constraints, supporting validation in the intended pediatric population rather than relying on virtual sample size alone.3

The model can usefully generate hypotheses about food effects and identify combinations that deserve clinical evaluation. Until its predictive tails are calibrated, the numerical attainment rates and statements that particular meals are necessary or sufficient should remain simulation-based hypotheses. Demonstrating acceptable mean concentration prediction is a different evidential achievement from showing that a recommended regimen reliably protects an individual child from subtherapeutic exposure.

Disclosure

The author declares no conflicts of interest in this communication.