Constructing Age-Stratified Pediatric Virtual Cohorts

Why a pediatric virtual cohort must be joint

A pediatric PBPK simulation often requires many virtual individuals, each with an age and a corresponding set of physiological parameters. Those parameters are not independent entries in a lookup table. Age, body weight, organ size, and maturation are related: for example, organ size generally changes with body size, while metabolic maturation follows its own developmental pattern. A cohort assembled by sampling each parameter separately can therefore contain values that are individually plausible but implausible in combination.

The goal is not to make every simulated child resemble an average child. It is to represent the intended population’s variability while preserving the relationships that make the individual parameter sets physiologically coherent. The earlier material on confounding between age, size, and maturation motivates keeping these dimensions distinct: a larger child of a given age is not automatically more mature for every enzyme, and a high maturation value does not itself imply a particular body weight.

A cohort is not defined only by its marginal distributions. A weight distribution and an organ-size distribution can each look reasonable on their own while their pairing is unrealistic. Cohort construction must therefore specify the target age composition, the within-age variability, and how parameters depend on one another.

Specify the target population and age strata

Start by defining the population the cohort is intended to represent: for example, children in a particular age range, clinical setting, or dosing context. Choose nonoverlapping age strata that support the intended comparisons and the available physiological information. Boundaries should be explicit, so every individual belongs to exactly one stratum. A design might use birth to less than 1 month, 1 to less than 12 months, 1 to less than 2 years, 2 to less than 12 years, and 12 to less than 18 years. These are design choices, not universal biological boundaries.

Record age at a resolution suited to development. Days or weeks may matter in early infancy, whereas years may be sufficient in older children. For preterm children, state whether age means chronological age, corrected age, or another defined measure, and apply the same rule to the reference data and the simulated cohort. Mixing age definitions can create apparent mismatches between age and maturation.

The number simulated in each stratum should reflect the intended use. For a population-representative cohort, use the target population’s age proportions. For an analysis requiring precise comparisons in a small age band, deliberate oversampling may be useful, but the resulting cohort is then not proportionally representative unless appropriate population weights are retained.

Generate parameters conditionally, not independently

A general construction order is: assign a stratum, sample an individual age within it, then sample physiological parameters from distributions conditional on age and relevant covariates. For example, body weight can be sampled from an age-specific reference distribution, and organ size can then be sampled conditional on age and body size. Maturation may be sampled from an age-dependent developmental distribution, with additional dependence on gestational history or other covariates when supported by the model and data.

Conditional sampling can use measured individual-level reference data, age-specific empirical distributions, or fitted conditional models. If a conditional model is used, check that it represents the intended population and does not extrapolate beyond the support of the data. A relationship estimated across all ages may be misleading if age groups have different patterns. Conversely, fitting each narrow age band separately may discard useful information or produce unstable estimates when data are sparse.

A simple way to preserve dependence between two approximately normal parameters is to generate correlated standardized residuals. Let \(Z_W\) be a standardized weight deviation and let \(\epsilon\) be an independent standard normal draw. For a chosen correlation \(\rho\), set \(Z_O=\rho Z_W+\sqrt{1-\rho^2}\epsilon\). Both residuals have mean zero and variance one, and their correlation is \(\rho\). Parameters can then be formed by applying age-specific means and standard deviations. This construction is only appropriate when the transformed distributions and dependence assumption are defensible; it is not a substitute for checking biological plausibility.

Many physiological quantities are positive, and maturation fractions lie between zero and one. A normal distribution on the original scale may generate negative organ sizes or maturation values outside their allowed range. Possible alternatives include modeling log organ size or using a bounded transformation for maturation, then transforming simulated values back. The chosen distribution should be checked against data and the model’s physiological constraints rather than selected only for mathematical convenience.

Why independent sampling can fail

If weight and organ size are sampled independently from their marginal distributions, the cohort retains the specified marginal means and variances but removes their dependence. Some resulting pairs may be physiologically discordant, such as low body weight paired with unusually large organ mass. Whether a particular pair is implausible depends on the target population, age, and the physiological quantity, so diagnostics should use suitable references rather than a universal cutoff.

The issue can be seen in the bivariate-normal illustration above. Independent standard normal deviations have correlation zero; with a positive weight–organ-size relationship, that independence underrepresents children who are jointly low or jointly high and creates more discordant pairings. The marginal histograms alone would not reveal this failure. Scatterplots, conditional summaries, and checks against age-specific reference ranges are needed.

Age stratification also matters to joint summaries. Even if two parameters have little association within each narrow age stratum, their pooled association can be substantial because both change with age. Conversely, a pooled correlation may conceal a different or weaker relationship within strata. For a parameter \(Y\), the overall variance separates into within-stratum and between-stratum components:

Here \(k\) indexes strata, \(\pi_k\) is the proportion in stratum \(k\), \(\mu_k\) is its mean, and \(\mu\) is the overall mean. The first term describes within-stratum variability; the second describes differences between stratum means. This decomposition follows by writing \(Y-\mu=(Y-\mu_k)+(\mu_k-\mu)\) within stratum \(k\), squaring, and taking expectations. The cross-term has expectation zero because \(E(Y-\mu_k\mid k)=0\). Thus pooled variation should not be mistaken for within-age variability.

Representing maturation without confusing it with size

Maturation should be assigned using the developmental measure relevant to the PBPK process, not inferred solely from body weight. Enzyme, transporter, renal, and binding processes can have distinct developmental profiles; an age-dependent value for one process should not be treated as a universal maturation factor. The earlier material on pediatric ontogeny supplies the process-specific context for selecting such profiles.

A practical cohort may assign a process-specific expected maturation value at each age and then represent supported between-child variability around that value. If individual-level evidence indicates dependence between maturation and another covariate, model it conditionally. Otherwise, avoid imposing a correlation merely because it seems biologically intuitive. In particular, do not duplicate an age trend by sampling an age-dependent maturation value and then applying an additional unsupported age or size adjustment.

Diagnostics, alternatives, and limitations

A cohort should be checked before it is used in PBPK simulations. Useful checks include age-specific histograms and quantiles for continuous parameters; scatterplots of weight against organ size; maturation against age; and cross-tabulations of clinically meaningful categories such as age band and prematurity status. Review extreme records individually. Check for impossible values, duplicated age adjustments, abrupt jumps at stratum boundaries, and combinations that fall outside supported physiological ranges. A smooth overall distribution is not sufficient if important age-specific patterns are wrong.

When diagnosing failure, distinguish a genuine tail observation from an invalid combination. A rare but supported child should not be removed simply because the value is unusual. Conversely, clipping every out-of-range value can pile observations at the boundary and distort the distribution. If constraints are required, build them into the conditional model or use evidence-based eligibility rules, and document how they affect the cohort.

No single approach is best in every setting. Resampling is attractive when the source data are representative and contain the required joint measurements. Conditional models can be more flexible but may depend strongly on assumptions, especially in sparse age ranges. Independent sampling can be a useful debugging comparison, but should not be treated as a realistic cohort unless independence is justified for the quantities being sampled. A fixed typical value may suit a deterministic scenario, but it is not a variable virtual population.

Finally, cohort realism is limited by the quality and scope of the supporting data. Sparse measurements, inconsistent age definitions, and unmeasured covariates can leave dependence uncertain. State these limitations and keep the simulated cohort’s purpose clear: its age composition, parameter distributions, and dependence structure should all correspond to the population and question being modeled.