Sampling Around Lag Time and Transit Absorption

Why Sampling Time Must Follow the Input Profile

A sample can be analytically precise yet contribute little to the parameter of interest. For an oral dose with an absorption lag, a measurement taken well after the concentration peak may describe elimination clearly but provide weak information about when absorption began. Conversely, an early sample close to the lag can be highly sensitive to lag time, but that advantage depends on knowing the lag closely enough to schedule the sample on the intended side of onset.

Transit absorption creates a different timing problem. Rather than a single sharp onset followed by first-order absorption, a sequence of transit compartments spreads drug input over time. The sampling plan should therefore distinguish the beginning of the input profile, its central region, and the later concentrations shaped primarily by absorption and elimination. The objective is not to sample at one universally “best” time: it is to obtain complementary information about the parameters the individual model is intended to estimate.

Lag-Time Model and Local Information

Consider a one-compartment oral model with first-order absorption and elimination, a lag time \(L\), and additive concentration error. Let \(u=t-L\) denote time since the lag has elapsed. For \(t>L\), write the concentration as

where \(k_a\) and \(k_e\) are the absorption and elimination rate constants, respectively, and \(B\) is a positive concentration scale. For \(t\leq L\), the model prediction is zero. Assume \(k_a>k_e>0\), a fixed dose, a correctly specified model, and observation errors that are independent Gaussian errors with mean zero and known variance \(\sigma^2\). For local sensitivity calculations, treat \(B\), \(k_a\), and \(k_e\) as fixed while assessing information about \(L\).

Because \(u=t-L\), differentiating with respect to \(L\) is the negative of differentiating with respect to \(u\). Thus,

This derivative explains why samples at different phases have different roles. Before the lag, the prediction is constant at zero, so its ordinary local derivative with respect to \(L\) is zero. Just after onset, the concentration can change rapidly when \(L\) shifts, making the derivative large in magnitude. At the concentration peak, the time derivative is zero, so the lag-time derivative is also zero. Later, the absorption contribution fades and the lag sensitivity generally becomes smaller.

For several independent samples at times \(t_i\), with the same error variance, the scalar information for \(L\) is

This is local information: it describes curvature around a specified nominal value of \(L\) and the other model parameters. It is not a guarantee that the estimate will be precise if the nominal model is wrong or the data do not support the assumed error structure.

Transit Absorption: A Distributed Input

A standard transit model represents absorption input as the output of a sequence of transit compartments. Let \(n\) be the number of transit compartments before the absorption depot, and let \(k_{\mathrm{tr}}\) be the common transit rate constant. The transit delay density is a gamma density:

It integrates to one, has mean transit time \(\mathrm{MTT}=(n+1)/k_{\mathrm{tr}}\), and, for \(n\geq1\), has its mode at \(n/k_{\mathrm{tr}}\). These describe the distribution of transit delays, not the time of the plasma concentration peak. After the transit delay, first-order absorption and elimination further shape the concentration.

For example, with a lag \(L\), the concentration can be expressed as the transit density convolved with the one-compartment response to a dose entering the absorption depot:

Here \(F\) is bioavailability and \(V\) is apparent volume. The transit profile broadens input, so sampling only at an anticipated sharp onset is less compelling than in a simple lag model. Useful times typically span the early rise, the central input region, and a later declining period. The precise concentration peak depends on transit, absorption, and elimination together; it cannot be read directly from the transit-density mode.

Onset, Peak, and Post-Absorption Samples

The timing distinction is particularly clear in the lag model. The concentration peak occurs at time \(t_{\max}=L+u_{\max}\), where setting the time derivative to zero gives

At this time, \(\partial C/\partial L=0\): moving the lag shifts the whole profile horizontally, but a sample exactly at the profile’s maximum has no first-order change in its predicted concentration under a small lag shift. A peak sample can still help estimate scale or other parameters; it is specifically weak for local lag information. Samples on the declining limb can inform elimination, while very late samples may contain little information about absorption because the absorption exponential has nearly vanished.

Assumptions, Diagnostics, and Design Choices

The calculations above assume a correct structural model, known nominal parameter values, accurately recorded sampling times, and independent Gaussian additive errors with constant known variance. The local information formula follows from the derivative of the model mean under those assumptions. If assay variability changes with concentration, the variance in the information calculation must reflect that error model; if observations are correlated, summing independent one-sample contributions is not valid.

A lag estimate can be fragile when a sample is scheduled very close to nominal onset. A small error in actual dose time, sample time, or nominal \(L\) can move the observation from just after onset to before it. Moreover, at exactly \(t=L\), the piecewise model is not smoothly differentiable in the ordinary two-sided sense. Do not use the steep post-lag derivative as if it applied across the boundary.

Useful diagnostics include plotting predicted concentrations and parameter sensitivities against time, checking whether the proposed samples fall on the intended side of the nominal lag, and recomputing the design over plausible values of \(L\), \(k_a\), and \(k_e\). If the high-information time shifts substantially across plausible values, include samples that bracket the onset rather than relying on one narrowly timed draw. For transit models, inspect both the transit density and the full predicted concentration profile; the former alone does not establish where concentration information is greatest.

A peak-focused schedule is easy to motivate when the clinical interest is maximum concentration, but it is a poor stand-alone strategy for lag estimation. Evenly spaced samples are less dependent on a single nominal peak or onset, though they may spend draws in uninformative regions. A sensitivity-based schedule targets parameter information directly; in practice, it should be balanced across parameters and checked for robustness. None of these approaches removes the need to consider feasible clinic windows, assay limits, or model misspecification.