Why Sparse Sampling Can Change the Exposure Conclusion
A concentration-time profile is continuous, but a clinical study observes it at selected times. When samples are sparse or poorly timed, the resulting exposure endpoints can depend as much on the schedule and interpolation assumptions as on the underlying profile. This matters in drug-drug interaction (DDI) studies: a missed peak or inaccurate area estimate can make an interaction appear smaller or larger than it is.
The key distinction is between the exposure endpoint of interest and its estimate. AUC over a stated interval is the area under the concentration curve over that interval; Cmax is the maximum concentration on that curve. A finite set of samples does not, by itself, reveal every point on the curve. Direct calculations therefore approximate the area and can only identify the largest observed concentration, not necessarily the actual maximum.
This formula treats the concentration between adjacent samples as changing linearly. It does not assert that the true concentration curve is linear. The assumptions that matter are therefore about what the observations capture, what happens between them, and whether the chosen interpolation is adequate for the curve’s shape.
How Sampling Density and Curve Shape Affect AUC
With exact endpoint concentrations, trapezoidal AUC is exact on an interval where concentration changes linearly. For curved profiles, the direction of the interval error depends on curvature. A convex curve lies at or below the straight chord joining its endpoints, whereas a concave curve lies at or above that chord. Since the trapezoid integrates that chord, convexity implies an estimate at least as large as the true interval area; strict convexity implies it is strictly larger. Concavity gives the reverse inequality. If curvature changes within the interval, neither direction is guaranteed.
This is a local statement: it concerns the curve between two observed endpoints. It does not establish a universal direction for total AUC bias across an entire dose interval. An absorption phase, a peak, and an elimination phase can have different shapes, and a sparse interval may span more than one of them. Missing samples near rapid changes also makes the assumed interpolation less credible.
Worked Example: Trapezoidal AUC During Exponential Decline
Suppose a hypothetical participant has concentrations of 10 mg/L at 0 hours and 5 mg/L at 2 hours. With only these two observations, the trapezoidal estimate for this interval is:
To illustrate how curvature affects the estimate, additionally suppose the true profile over this interval follows exponential decline, \(C(t)=10e^{-kt}\). The observed endpoint at 2 hours implies \(5=10e^{-2k}\), so \(k=\ln(2)/2=0.3466\ \mathrm{h^{-1}}\). The exact interval area under this assumed curve is:
The trapezoidal estimate is about \(15-14.43=0.57\ \mathrm{mg\cdot h/L}\) higher, or roughly 4.0% of the exact area under the assumed curve. This direction is consistent with convexity: the second derivative is \(k^2C(t)>0\), so the chord is above the curve. The exponential assumption is needed for the exact comparison; the two observed concentrations alone do not establish the unobserved shape.
Increasing sampling density can reduce interpolation error when added samples resolve the curve’s changes, but it does not guarantee an unbiased endpoint. For example, a denser schedule that still misses a brief peak can underestimate Cmax, and a tail interval that remains long can still contribute appreciable AUC error. The earlier material on sampling schedules for AUC and Cmax addresses placing observations to capture important parts of a profile; the present issue is how to recognize and interpret bias when the available schedule is sparse or mistimed.
Why Cmax Is Especially Sensitive to Timing
If measured concentrations are exact and the sampling times fall inside the interval used to define Cmax, the largest sampled value cannot exceed the true profile maximum: each observation is one value of the profile. It can equal the maximum if a sample lands at the peak, but otherwise it may fall below it. The size of the shortfall depends on how far the nearest sample is from the peak and how sharply concentration rises or falls around it.
This distinction is important when comparing DDI conditions. If the perpetrator condition shifts the peak time, changes absorption rate, or makes the peak narrower, a schedule chosen around the control profile may miss the perpetrator peak more severely. Then the observed Cmax ratio can be biased even if the same nominal sample times are used in both conditions. Equal timing is useful, but it does not ensure equal endpoint bias.
Worked Example: A Missed Peak Distorts a Cmax Ratio
Consider a hypothetical probe study in which dense reference observations establish a reference Cmax of 10 mg/L, while dense observations during perpetrator treatment establish a Cmax of 15 mg/L. The profile maxima imply a DDI Cmax ratio of:
Suppose a sparse schedule misses both peaks. Its largest observed reference concentration is 9 mg/L, and its largest observed perpetrator concentration is 10.8 mg/L. The ratio calculated from those sparse observations is:
The sparse-schedule ratio differs from the ratio based on the profile maxima by \(1.2-1.5=-0.3\) ratio units. In this hypothetical example, the sparse schedule underestimates the relative increase in Cmax, even though it underestimates both condition-specific maxima. That result illustrates why the direction of bias in a ratio cannot be inferred merely from knowing that each observed maximum is below its respective true maximum: the relative underestimation matters.
When a Population-Model-Based Estimate Is Defensible
A population pharmacokinetic (popPK) or PBPK model can use information across participants, sampling times, and relevant covariates to estimate an individual concentration-time profile. An individual exposure estimate may then be calculated from that model-implied profile rather than from direct interpolation between that participant’s sparse observations. This can be useful when the study cannot collect dense samples, but it replaces a simple interpolation assumption with a more structured set of model assumptions.
A model-based estimate is justified only when the model is suitable for the intended population, dose, and DDI conditions; the sampling design contains enough information to update the relevant individual predictions; and the model’s structure and parameters are supported by appropriate data. In a PBPK application, that includes checking that the interaction mechanisms and physiological assumptions relevant to the study condition are represented plausibly. A model’s ability to generate a smooth curve is not evidence that the curve is accurate.
Define the AUC interval or Cmax target and the DDI conditions being compared. Confirm that the samples cover times informative for those targets.
Assess whether the model has been evaluated for the drug, dose range, population, and interaction mechanisms at issue. Check that key covariates and relevant absorption or disposition features are represented.
Determine whether the observations help estimate individual exposure or whether the estimate is driven mainly by population assumptions. Report uncertainty and avoid presenting a model prediction as if it were directly observed.
Compare predictions with observed profiles or external data where available, inspect residual patterns and influential observations, and assess whether reasonable alternative assumptions change the exposure conclusion.
Worked Example: Deciding Whether to Use a Model-Based Estimate
Suppose a hypothetical DDI study collects only two postdose samples per participant, at 1 and 8 hours, and needs to estimate AUC from 0 to 12 hours. A previously evaluated model for the same drug and dose range has been checked against denser profiles in a comparable population. It includes the perpetrator effect supported by prior data, and its predictive performance has been assessed separately in both study conditions.
What the samples contribute. The 1-hour observation informs the early profile and the 8-hour observation informs later concentrations. Neither observation directly measures the concentration at 0 or 12 hours, and two points alone cannot establish the shape between them or the tails. Direct trapezoidal integration would require endpoint handling and interpolation assumptions that these observations cannot verify.
What supports model use. The model’s prior evaluation in comparable conditions, explicit representation of the interaction, and checks in both DDI conditions make a model-based individual estimate more defensible than treating a straight line between two observations as a known profile. The estimate should still be reported as model-based, with its uncertainty and dependence on the model made clear.
What would weaken the case. If the perpetrator changes absorption in a way not represented by the model, if validation covers only the reference condition, or if the sparse observations are inconsistent with model predictions, the model-based estimate is not adequately justified. Additional informative samples or a better-supported model would be needed; smoothness alone cannot repair a structural mismatch.
Diagnostics, Alternatives, and Limits
Warning signs include large gaps around the expected peak, a sampling window that ends before the relevant exposure interval is covered, abrupt residual patterns when model predictions are compared with observations, and large changes in estimated exposure under reasonable alternative model assumptions. For direct estimates, inspect the actual times and concentrations rather than assuming nominal times were achieved. A sample delayed by several hours may be uninformative about the peak even if its scheduled time was appropriate.
The main alternatives serve different purposes. A denser, better-timed schedule provides more direct information and is generally preferable when feasible. Direct trapezoidal integration makes fewer structural assumptions about the full profile but relies on approximation between observed points. A population-model-based estimate borrows information across participants and time points, but depends on the model’s suitability and may be influenced strongly by prior population information when an individual has few observations. PBPK simulation can represent physiological and mechanistic assumptions, but likewise does not make an unsupported assumption true. None of these approaches automatically corrects a poorly chosen schedule or a mismatch between the model and the study.
Sparse schedules can also affect DDI comparisons unequally. If the two conditions have different curve shapes or peak times, their respective AUC or Cmax biases need not cancel in a ratio. Therefore, a reported exposure ratio should be interpreted in light of the sampling design and estimation method, not just the numerical result. A model-based ratio is still conditional on its assumptions; it is not an assumption-free substitute for observed exposure.
Common pitfall: “Both conditions used the same sample times, so any bias must cancel in the ratio.” This is not generally true. If one condition has a narrower or shifted peak, the same schedule can miss a different fraction of its Cmax. For AUC, condition-specific curvature can likewise produce different trapezoidal errors. Check the shape and timing in each condition and assess whether conclusions are robust to plausible estimation choices.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For a convex concentration curve between two exact observations, how does trapezoidal area compare with the true interval area?
- A schedule records a maximum of 7 mg/L, while the true profile reaches 8 mg/L. What is the observed-minus-true difference, with units?
- Why might using identical sample times in reference and perpetrator conditions fail to eliminate bias in the Cmax ratio?
- Name two considerations that support using a population-model-based exposure estimate from sparse data.
- What does a smooth model-predicted profile establish about its accuracy?
Solutions. (1) The trapezoidal area is at least the true area, and is strictly greater if the curve is strictly convex throughout the interval. (2) \(7-8=-1\ \mathrm{mg/L}\), so the observed maximum is 1 mg/L lower. (3) The DDI may shift or sharpen the peak, so the same times can miss different fractions of the two true maxima. (4) Examples include validation in a comparable population and DDI condition, an adequately represented interaction mechanism, informative observations, and acceptable predictive checks. (5) Smoothness alone establishes no accuracy; accuracy depends on the model assumptions and evidence supporting them.