Why Derive Exposure Measures From Simulated Profiles?
A simulated concentration-time profile describes drug concentration at selected times. Many pharmacometric questions, however, concern summaries of the whole profile: total exposure over a dosing interval, the highest concentration reached, or the time at which that maximum occurs. These are derived exposure measures. Their distributions can vary across simulated subjects, parameter draws, or both, even when every profile is generated from the same fitted model.
The calculation is part of the definition of the result. AUC calculated over one dosing interval is not interchangeable with AUC over a different interval; a maximum found on a sparse time grid may not equal the underlying curve's maximum. Therefore, state the time window, units, calculation method, and source of variation alongside each summary.
Calculating AUC From Discrete Profiles
Simulation output is often recorded at discrete times \(t_0<t_1<\cdots<t_K\), with concentrations \(C_0,\ldots,C_K\). When the curve between adjacent points is approximated by a straight line, the area on each segment is a trapezoid. Summing those areas gives the linear trapezoidal estimate.
This follows by integrating the linear interpolation on one segment. If the segment has duration \(\Delta t=t_{k+1}-t_k\), its concentration changes linearly from \(C_k\) to \(C_{k+1}\). The average concentration over that line is \((C_k+C_{k+1})/2\), so its area is average concentration times duration. Adding the segment areas yields the displayed sum. The approximation is more reliable when the grid captures important curvature and changes in the profile.
Worked Example: Interval AUC and Cmax From Four Profile Points
Suppose a hypothetical simulated profile has the following concentrations after a dose. Concentrations are in mg/L and times are in hours.
| Time (h) | 0 | 2 | 4 | 8 |
|---|---|---|---|---|
| Concentration (mg/L) | 0 | 6 | 9 | 4 |
Using the linear trapezoidal rule over 0–8 h:
The three segment areas are \(6\), \(15\), and \(26\) mg·h/L, respectively, so \(AUC_{0-8}\approx 6+15+26=47\) mg·h/L. The largest recorded concentration is 9 mg/L at 4 h, giving a grid-based \(C_{\max}=9\) mg/L and \(T_{\max}=4\) h. These maxima describe the recorded profile points; if the true peak lies between them, the grid-based Cmax can be lower than the continuous-curve maximum.
For repeated dosing, a common summary is \(AUC_{0-\tau}\), the area from the start of a dosing interval to its end, where \(\tau\) is the interval length. Summarize the interval specified by the question: an area from time zero to a later time can include more than one dose and should not be labelled as a single-interval AUC. If the model gives an analytic concentration function, numerical integration of that function can be used instead of integrating only at saved output points.
Interpreting Cmax and Other Exposure Summaries
For a discrete profile, calculate Cmax as the largest concentration among points in the stated window. Report the time of that point as grid-based Tmax. If there are multiple points tied for the maximum, state the convention used for Tmax, such as reporting the earliest time. When the peak is expected to occur between widely spaced points, save a finer grid around the peak or optimize the model-predicted curve over the time window. Do not describe a maximum on a coarse grid as an exact continuous-time maximum.
Other summaries require the same care about their definitions. A partial AUC is calculated over a specified subinterval, for example \(AUC_{2-8}\), and is not the total AUC over the dosing interval. For an oral concentration-time profile, the maximum may occur after time zero; after an idealized intravenous bolus, the maximum may occur at the instant of dosing. Thus, the observation window and dosing-time convention matter.
AUC to infinity is distinct from AUC over a finite simulated window. One common terminal extrapolation is \(AUC_{0-\infty}\approx AUC_{0-t_{\mathrm{last}}}+C_{\mathrm{last}}/\lambda_z\), where \(C_{\mathrm{last}}\) is the last observed concentration and \(\lambda_z>0\) is the terminal log-linear elimination rate constant. This approximation assumes the terminal portion is adequately represented by exponential decline and that the terminal rate has been estimated appropriately. A short or noisy tail can make the extrapolated component unreliable; report its contribution and do not treat it as directly integrated area.
Worked Example: Checking AUC Against an Analytic Profile
Consider a hypothetical one-compartment intravenous bolus profile with dose \(D=100\) mg, volume \(V=20\) L, and elimination rate constant \(k=0.10\) h\(^{-1}\). For \(t\geq 0\), the model concentration is \(C(t)=(D/V)e^{-kt}\). The initial concentration is \(100/20=5\) mg/L, so the maximum is at \(t=0\): \(C_{\max}=5\) mg/L and \(T_{\max}=0\) h.
Integrating this exponential profile gives:
The finite-window area is smaller. At 8 h, \(C(8)=5e^{-0.8}\approx2.247\) mg/L. The exact area through 8 h is \(5(1-e^{-0.8})/0.10\approx27.53\) mg·h/L. The remaining area is \(C(8)/k\approx22.47\) mg·h/L, and the two parts sum to 50.0 mg·h/L. This checks both the units and the distinction between finite-window AUC and AUC to infinity. The calculation relies on the specified exponential model; a discrete trapezoidal estimate need not equal the analytic result exactly.
Separating Parameter and Subject Variation
A simulated exposure measure inherits variation from the inputs that generated its profile. In nested simulation, parameter draws represent uncertainty in fitted model parameters, while newly simulated subject effects represent between-subject variability under a given parameter set. Covariates and between-occasion effects can also vary if the simulation design includes them. Keep these sources identifiable in the simulation output: a pooled distribution alone does not show which source is responsible for its spread.
A useful descriptive summary is to calculate the exposure measure for every simulated profile and then report its empirical mean, standard deviation, median, or selected quantiles, with units. To examine sources separately, compare exposure distributions across parameter draws while holding the subject-generation design fixed, and compare subjects within a parameter draw. If parameter and subject effects interact, their contributions need not separate into simple additive pieces. These descriptive summaries are not confidence intervals for a parameter.
Worked Example: Describing Two Sources of AUC Variation
Suppose an illustrative nested simulation produces the following AUC values in mg·h/L. The two parameter draws are denoted P1 and P2, and three newly simulated subjects are denoted S1–S3. Values are hypothetical and chosen to make the variation calculation transparent.
| Parameter draw | Subject S1 | Subject S2 | Subject S3 | Mean across subjects |
|---|---|---|---|---|
| P1 | 28 | 32 | 36 | 32 |
| P2 | 44 | 48 | 52 | 48 |
Across all six values, the mean is \((28+32+36+44+48+52)/6=40\) mg·h/L. The sample variance across the six profiles is the sum of squared deviations from 40 divided by \(6-1\):
Thus the pooled sample standard deviation is \(\sqrt{89.6}\approx9.47\) mg·h/L. The parameter-draw means are 32 and 48 mg·h/L, while the subject values within each draw differ by \(-4,0,4\) mg·h/L from that draw's mean. In this constructed additive example, the equally weighted population variance across parameter-draw means is \(\{(-8)^2+8^2\}/2=64\), and the equally weighted within-draw subject variance is \((16+0+16)/3=10.67\), both in squared AUC units. They sum to the pooled population variance \(448/6=74.67\;(\text{mg·h/L})^2\).
This exact decomposition works here because the table has a balanced crossed structure and the subject deviations are the same under both parameter draws. In an actual nonlinear model, parameter and subject effects can interact; do not assume that variance contributions add in this way without checking the design and structure. The table nevertheless illustrates why both conditional results (within a parameter draw) and pooled results can be useful.
Choosing and Checking the Calculation
Linear trapezoidal integration is simple and widely useful for sampled profiles. When concentrations decline approximately exponentially between two positive observations, a log-linear trapezoidal segment can better represent that decline: its area is \((C_{k+1}-C_k)/\ln(C_{k+1}/C_k)\times(t_{k+1}-t_k)\), provided \(C_k\neq C_{k+1}\) and both concentrations are positive. It is not appropriate across a zero concentration or a segment where the assumed log-linear decline is unsuitable. An analytic integral or numerical integration of the model curve can be preferable when available and sufficiently accurate. No method rescues a time grid that misses a rapid peak or an important change in slope.
Before interpreting exposure summaries, check the following:
- Time window: verify the start, end, and dosing interval. Confirm that the calculated AUC includes exactly the intended observations.
- Units: if concentration is in mg/L and time in hours, AUC is in mg·h/L. Convert time units consistently before calculating.
- Grid resolution: inspect whether the grid is dense enough around absorption, peak concentration, or rapid distribution changes. Refine it if Cmax or AUC changes materially when the grid is refined.
- Profile validity: investigate missing times, duplicate records, negative concentrations when the model requires nonnegative concentrations, or a profile that does not cover the stated interval.
- Variation source: label whether summaries pool parameter draws, subjects, occasions, or replicates. Do not attribute pooled spread to one source without a design that supports that comparison.
- Tail extrapolation: for AUC to infinity, check the terminal fit and report how much of the total is extrapolated. A large extrapolated fraction signals that the result depends strongly on tail assumptions.
A frequent mistake is to calculate the maximum of the recorded points and interpret it as the exact maximum of the continuous concentration curve. The recorded-point maximum is only the maximum on that grid; use a finer grid or optimize the curve when a continuous-time peak is required. Another mistake is to pool every simulated exposure and call its standard deviation “between-subject variability.” If parameter draws were also varied, that pooled standard deviation includes parameter uncertainty as well as subject variation.
Check Your Understanding
Answer each question, then compare your work with the solutions.
- A hypothetical profile has concentrations 2, 5, and 3 mg/L at 0, 1, and 3 h. Calculate the linear trapezoidal AUC from 0 to 3 h and identify the grid-based Cmax.
- Why can Cmax calculated on a sparse time grid be lower than the maximum of the continuous curve?
- A simulated profile has \(C_{\mathrm{last}}=1.5\) mg/L at 10 h and an estimated terminal rate \(\lambda_z=0.15\) h\(^{-1}\). Calculate the terminal extrapolated area and state one assumption needed to use it.
- What sources of variation are included in a pooled distribution of AUCs if both parameter draws and new subjects are simulated?
Solutions. (1) The segment areas are \((2+5)(1-0)/2=3.5\) and \((5+3)(3-1)/2=8\) mg·h/L. Therefore AUC is \(11.5\) mg·h/L, and the grid-based Cmax is 5 mg/L at 1 h. (2) The true peak can occur between saved times, so none of the recorded concentrations need equal the continuous maximum. (3) The extrapolated area is \(C_{\mathrm{last}}/\lambda_z=1.5/0.15=10\) mg·h/L. It assumes the terminal profile is adequately represented by exponential decline at the estimated rate. (4) The pooled distribution includes variation from parameter draws and between-subject differences; it may also include other simulated sources, such as occasion variability, if those were varied.