Passing Population PK Outputs Into Exposure-Response Models

Why the Handoff Matters

An exposure-response (ER) analysis often uses an exposure metric derived from a population pharmacokinetic (PK) model. That handoff is not just a matter of copying numbers into a new data set. The PK model may use observed concentrations to estimate each participant’s individual PK parameters; those parameters then generate predicted concentrations or a summary exposure metric. The ER analysis uses that derived quantity to describe or predict a response. Each step changes what the number represents.

A measured concentration at a sampling time, a model-predicted concentration at that time, and an exposure summary over an interval are different kinds of quantities. They may happen to have the same numerical value, but they are not interchangeable. Confusing them can lead to an ER analysis that treats predictions as measurements, pairs exposure and response from the wrong times, or hides uncertainty created by the PK estimation step.

From Concentrations to an Exposure Input

Let \(Y_{ij}\) denote the observed concentration for participant \(i\) at sampling time \(t_{ij}\). Let \(C_i(t)\) denote the concentration predicted by the PK model for participant \(i\) at time \(t\). A summary such as area under the concentration-time curve (AUC) over an interval is calculated from the predicted profile. For an interval from \(t_0\) to \(t_1\), define:

Here \(E_i\) is the exposure input for participant \(i\), with units of concentration multiplied by time. Its value depends on the model, the participant’s estimated PK parameters, the dose history, and the chosen interval. The observations \(Y_{ij}\) inform the PK fit; the predictions \(C_i(t)\) define the profile; and \(E_i\) summarizes that profile. None of these definitions requires their numerical values to differ.

In a population PK analysis, individual parameters are commonly estimated using both the population model and that participant’s concentration data. An empirical Bayes estimate (EBE) is a participant-specific estimate conditional on the fitted population model and the participant’s available observations. It is not a measured PK parameter. In data-sparse participants, EBEs are typically pulled toward the population distribution; this is often called shrinkage. The extent and consequences of shrinkage should be assessed before treating individual estimates as precise exposure inputs.

A sound handoff specifies, at minimum, the participant identifier, the exposure definition and units, the dose and time interval used, and the model version that generated the estimate. It also checks that the response record belongs to the same participant and is associated with an appropriate exposure period. The analysis population and clinical conditions need to be compatible with those used to obtain the PK estimates.

Using the Exposure in an ER Model

Once the exposure estimate is defined and aligned with the response, the ER model relates response to exposure. For example, a simple saturable mean-response model can be written as \(m(E_i)=E_0+E_{\max}E_i/(EC_{50}+E_i)\), where \(E_0\) is the baseline mean response, \(E_{\max}\) is the maximum increase above baseline, and \(EC_{50}\) is the exposure producing half of that increase. This describes a model relationship; it does not by itself establish that exposure causes the response. Interpretation also depends on the study design, relevant covariates, and assumptions of the response model.

The exposure definition must be fixed before interpreting the ER relationship. For instance, an AUC over one dosing interval and a concentration at a selected time are not the same input merely because both are expressed using concentration units. The metric, interval, units, and dose context need to be stated and used consistently. The specific choice among exposure metrics is a separate modeling decision; the transfer step should preserve, not obscure, the definition that was chosen.

Carrying PK Uncertainty Forward

An individual exposure estimate is uncertain because the PK parameters are estimated from finite, possibly sparse data and because the population model itself is estimated. Using a single EBE as if it were known exactly gives a convenient sequential analysis, but it conditions the ER fit on that one set of exposure values. This approximation may understate uncertainty in ER results when PK estimation uncertainty is material. The size and direction of resulting changes are not guaranteed to be identical in every analysis.

One practical representation is to calculate the exposure for each draw from the participant’s conditional PK parameter distribution, using the same dose history and interval for every draw. Each draw is then passed through the ER model. This preserves the fact that the exposure input is uncertain without relabeling it as an observed concentration. In a full analysis, uncertainty in ER parameters and response variability also matter; the example below isolates the effect of exposure uncertainty.

The first term represents average response variance conditional on exposure; the second represents variation in the conditional mean response across possible exposures. A plug-in analysis at one exposure value does not represent the second term as exposure uncertainty. This follows from the identity \(\operatorname{Var}(R)=\operatorname{E}[\operatorname{Var}(R\mid E)]+\operatorname{Var}(\operatorname{E}[R\mid E])\), which can be obtained by writing \(R-\operatorname{E}[R]\) as \([R-\operatorname{E}(R\mid E)]+[\operatorname{E}(R\mid E)-\operatorname{E}(R)]\). The cross-product has expectation zero because the first bracket has conditional expectation zero given \(E\).

Diagnostics, Alternatives, and Limitations

Before interpreting an ER model, check that the transfer is functioning as intended. Useful checks include confirming participant identifiers and record counts after merging; verifying units, dose history, and interval definitions; plotting observed concentrations against PK predictions over time; and examining whether individual estimates are strongly shrunk toward the population. Inspect ER fit by exposure range and by relevant design or clinical subgroups. A trend driven by a few extreme exposure estimates, a mismatch in timing, or poor PK predictive performance can make an apparently precise ER relationship misleading.

The main input choices have different meanings and limitations:

A common mistake is to say, “The PK model estimated this participant’s AUC, so AUC is observed without error in the ER data set.” The correction is that AUC is a model-derived estimate, and its precision depends on the PK model, data, and estimation. Another mistake is to use each participant’s nearest concentration sample as though it were the interval exposure. A sample is tied to a time; an interval summary describes a profile over time. Neither is automatically a suitable substitute for the other.

The handoff also has limits. A well-constructed exposure input cannot repair a misspecified PK model, nonrepresentative data, an inappropriate response model, or a misaligned response window. A sequential workflow is often practical and transparent, but a plug-in estimate does not fully communicate upstream uncertainty. More integrated approaches may be considered when the scientific question requires them; the choice depends on the analysis objective and data structure. Regardless of approach, predictions should be described as model-based quantities, not direct measurements or proof of a causal effect.