Predicting AUC and Cmax From Individual Posteriors

From Individual Parameter Uncertainty to Exposure Uncertainty

A concentration forecast describes exposure over time, but many pharmacometric decisions are expressed in terms of a summary of that curve. AUC describes total exposure over a stated interval; Cmax describes the highest concentration over that interval. When an individual’s pharmacokinetic parameters are uncertain, neither summary should be calculated only from a single “best” parameter estimate. Instead, calculate the exposure metric for each draw from the individual’s posterior distribution. The resulting draws approximate the posterior distribution of that metric under the specified model and regimen.

This is a transformation of the individual parameter posterior, not a new data update. The observed concentrations, residual-error model, and prior information have already contributed to that posterior. The forecast regimen and the definition of the exposure metric must now be specified: a change in dose, route, dosing interval, or AUC time window can change the distribution being calculated.

Definition: Let \(\theta_i\) denote the pharmacokinetic parameters for individual \(i\), and let \(y_i\) denote that individual’s observed data. A posterior draw \(\theta_i^{(m)}\) is a draw from \(p(\theta_i\mid y_i)\), conditional on the model and its assumptions. For a specified regimen \(r\) and interval \([a,b]\), the model-based AUC for draw \(m\) is \(\mathrm{AUC}_i^{(m)}(a,b)=\int_a^b C_i(t;\theta_i^{(m)},r)\,dt\). The corresponding Cmax is \(C_{\max,i}^{(m)}(a,b)=\max_{t\in[a,b]} C_i(t;\theta_i^{(m)},r)\), where \(C_i\) is the model-predicted concentration.

AUC is reported in concentration multiplied by time, such as mg·h/L, and Cmax in concentration, such as mg/L. The interval is part of the definition. For example, AUC from zero to the end of one dosing interval is not automatically the same quantity as AUC from zero to infinity after a single dose. Cmax also depends on whether the interval includes the absorption phase, a prior dose, or both.

Transform Each Draw, Then Summarize

Suppose \(M\) posterior draws are available. For every draw, use the same forecast regimen and calculate the full model-predicted profile needed to obtain the requested metric. This produces paired values \((\mathrm{AUC}_i^{(m)},C_{\max,i}^{(m)})\). Posterior means, medians, standard deviations, and quantiles can then be computed from those metric draws. Pairing matters: AUC and Cmax from draw \(m\) both use the same \(\theta_i^{(m)}\), so their joint variation is retained.

\widehat{E}[\mathrm{AUC}_i\mid y_i] =\frac{1}{M}\sum_{m=1}^{M}\mathrm{AUC}_i^{(m)},\qquad \widehat{E}[C_{\max,i}\mid y_i] =\frac{1}{M}\sum_{m=1}^{M}C_{\max,i}^{(m)}

These are Monte Carlo estimates of posterior means. A posterior median is the median of the metric draws, and a central 95% posterior credible interval is formed from their 2.5th and 97.5th percentiles. Conditional on the model, observed data, and forecast regimen, that interval contains 95% of the posterior probability for the individual’s model-based metric. This is a posterior probability statement, not a frequentist coverage statement.

A key computational distinction is that the metric must be evaluated draw by draw. In general, the metric evaluated at posterior mean parameters is not the posterior mean of the metric. For example, under a one-compartment IV bolus model, AUC is \(D/CL\); because taking a reciprocal is nonlinear, \(D/E[CL\mid y_i]\) need not equal \(E[D/CL\mid y_i]\).

Worked Example: AUC and Cmax After an IV Bolus

Consider a hypothetical individual receiving a 400 mg IV bolus. Assume a one-compartment model with instantaneous input, linear elimination, and concentration \(C(t)=D/V\exp(-CLt/V)\). For each posterior draw, calculate AUC from zero to infinity and the maximum concentration. The model-based formulas are \(\mathrm{AUC}_{0-\infty}=D/CL\) and \(C_{\max}=D/V\), with the maximum occurring immediately after the bolus.

Draw	CL (L/h)	V (L)	AUC (mg·h/L)	Cmax (mg/L)
1	4	40	400/4 = 100	400/40 = 10
2	5	50	400/5 = 80	400/50 = 8
3	6	60	400/6 = 66.67	400/60 = 6.667
4	8	80	400/8 = 50	400/80 = 5

For these four illustrative draws, the Monte Carlo mean AUC is \((100+80+66.67+50)/4=74.17\) mg·h/L, and the mean Cmax is \((10+8+6.667+5)/4=7.417\) mg/L. These four values demonstrate the calculation; they are too few to characterize posterior tails reliably. In an analysis, summaries and credible intervals should use a sufficiently large, adequately sampled posterior.

The paired draws show that higher clearance and larger volume coincide in this small illustration. Calculating AUC from one set of clearance draws and Cmax from a separately sorted set of volume draws would lose that pairing and could distort their joint posterior relationship.

Analytical Exposure Formulas When Available

For some models, the metric can be calculated directly from parameters, avoiding numerical integration. For a single oral dose in a one-compartment model with first-order absorption and elimination, assume \(k=CL/V\), \(k_a>k\), constant bioavailability \(F\), linear kinetics, and no contribution from another dose. The predicted concentration is

C(t)=\frac{F D k_a}{V(k_a-k)} \left(e^{-kt}-e^{-k_a t}\right),\qquad t\geq 0.

Integrating this curve from zero to infinity gives \(\mathrm{AUC}_{0-\infty}=FD/CL\). The peak occurs where the derivative of concentration with respect to time is zero. Setting that derivative to zero gives \(k_a e^{-k_a t}=k e^{-kt}\), and rearranging yields the time of peak concentration:

t_{\max}=\frac{\ln(k_a/k)}{k_a-k}.

Substituting this time into \(C(t)\) gives Cmax for that draw. The assumptions matter: these formulas do not automatically apply to nonlinear elimination, multiple compartments, a finite AUC interval, or a regimen with overlapping doses. For such cases, evaluate the appropriate model profile over the specified interval.

Worked Example: Oral Dose With First-Order Absorption

For a hypothetical single oral dose, let \(D=100\) mg, \(F=0.80\), \(CL=5.0\) L/h, \(V=40\) L, and \(k_a=1.20\) h\(^{-1}\). Assume the one-compartment first-order model above, with \(k_a>k\), linear elimination, and no other doses. First, \(k=CL/V=5.0/40=0.125\) h\(^{-1}\). The total AUC is

\mathrm{AUC}_{0-\infty}=\frac{FD}{CL} =\frac{0.80(100\ \mathrm{mg})}{5.0\ \mathrm{L/h}} =16.0\ \mathrm{mg\cdot h/L}.

The time to peak is

t_{\max}=\frac{\ln(1.20/0.125)}{1.20-0.125} =\frac{\ln(9.60)}{1.075} =2.104\ \mathrm{h}.

At that time, \(e^{-0.125(2.104)}\approx0.7687\) and \(e^{-1.20(2.104)}\approx0.0801\). Therefore,

C_{\max} =\frac{0.80(100)(1.20)}{40(1.20-0.125)} (0.7687-0.0801) \approx 1.537\ \mathrm{mg/L}.

In posterior forecasting, this calculation is repeated with \(F\), \(CL\), \(V\), and \(k_a\) from each posterior draw. The resulting AUC and Cmax values vary across draws, even though the dose and model form are fixed.

Numerical Calculation From Predicted Profiles

When no convenient analytical expression is available, calculate AUC from model-predicted concentrations on a sufficiently fine time grid. The trapezoidal rule approximates the area between successive times \(t_j\) and \(t_{j+1}\) by a trapezoid. Cmax can be approximated by the largest predicted concentration on the grid. Both calculations should be repeated separately for each posterior draw.

\mathrm{AUC}_{0-T}^{(m)} \approx \sum_{j=0}^{J-1} \frac{C^{(m)}(t_j)+C^{(m)}(t_{j+1})}{2} (t_{j+1}-t_j), \qquad C_{\max}^{(m)}\approx\max_{0\leq j\leq J} C^{(m)}(t_j).

The units follow from multiplying concentration by time in each trapezoid. A grid-based Cmax is only an approximation if the actual peak falls between grid points. A sparse grid can also underestimate AUC, particularly when concentrations change rapidly between sampled times.

Worked Example: Trapezoidal AUC and Grid-Based Cmax

Suppose three hypothetical posterior draws produce concentrations at times 0, 1, 2, and 4 h as shown below. The predictions are from the structural model, not noisy measured concentrations. Use the trapezoidal rule for AUC from 0 to 4 h and take the largest listed concentration as the grid-based Cmax.

Draw	C(1) (mg/L)	C(2) (mg/L)	C(4) (mg/L)
1	4	3	1
2	3	2	0.5
3	5	4	2

For draw 1, the interval widths are 1, 1, and 2 h. Thus AUC is \(\frac{0+4}{2}(1)+\frac{4+3}{2}(1)+\frac{3+1}{2}(2)=2+3.5+4=9.50\) mg·h/L, and the grid-based Cmax is 4 mg/L. For draw 2, AUC is \(\frac{0+3}{2}(1)+\frac{3+2}{2}(1)+\frac{2+0.5}{2}(2)=1.5+2.5+2.5=6.50\) mg·h/L, with grid-based Cmax 3 mg/L. For draw 3, AUC is \(\frac{0+5}{2}(1)+\frac{5+4}{2}(1)+\frac{4+2}{2}(2)=2.5+4.5+6=13.0\) mg·h/L, with grid-based Cmax 5 mg/L.

These three AUC values and three Cmax values are an empirical approximation to their posterior distributions. Here, the concentration maximum happens to appear at 1 h for each draw. In general, refine the grid around a possible peak or optimize the predicted concentration over continuous time; otherwise, the calculated Cmax may be too low.

Diagnostics, Alternatives, and Limitations

Before interpreting metric summaries, check that the calculations represent the intended forecast. Confirm the dose history, route, dosing interval, time origin, and AUC limits. Verify parameter units: for example, dose in mg and clearance in L/h produce AUC in mg·h/L. Check that concentrations are nonnegative when the model requires this, and that the numerical AUC is stable when the time grid is refined. Compare numerical results with an analytical formula when both are available.

Posterior sampling quality also matters. If draws are strongly autocorrelated, the nominal draw count overstates the amount of information available for tail quantiles. Inspect sampling diagnostics and ensure enough effective draws for the summaries of interest. Implausible or extreme metric draws may reflect a genuine broad or skewed posterior, but they can also reveal coding errors, unit mismatches, or parameter draws near a boundary. Investigate their source rather than silently removing them.

Analytical formulas are usually faster and may be more accurate when their assumptions apply. Numerical integration is more flexible for complex models and finite intervals, but it introduces time-grid and numerical-integration error. A plug-in calculation using posterior mean parameters is simpler, but generally does not represent the posterior distribution of the exposure metric. The draw-by-draw approach preserves parameter uncertainty and, when calculations are paired, the dependence among metrics.

These are model-based exposure distributions. They describe uncertainty in the exposure implied by posterior uncertainty in the pharmacokinetic parameters, conditional on the chosen model and regimen. They are not automatically distributions of future measured concentrations with residual observation error added. Nor do they account for model uncertainty unless that uncertainty is explicitly represented. In particular, an interval computed from these draws cannot correct a misspecified structural model or an incorrectly recorded regimen.

Pitfall: plug in first, transform later. Using \(D/E[CL\mid y_i]\) as though it were the posterior mean AUC ignores the nonlinearity of \(1/CL\). Calculate \(D/CL^{(m)}\) for every draw, then summarize.
Pitfall: use the largest measured concentration as model-based Cmax. Sparse observed samples can miss the peak, and measured values include residual error. Use the predicted concentration curve for each posterior draw and search over the defined interval.
Pitfall: report an interval without its conditioning and units. State the metric, interval, units, regimen, and that the credible interval is conditional on the model and data.

Key takeaway: Specify the regimen and exposure window, transform every individual posterior draw into AUC and Cmax, and summarize the resulting metric draws. Check analytic assumptions or numerical accuracy before interpreting posterior uncertainty.

Check Your Understanding

Answer each question, then compare with the brief solutions.

Under a one-compartment IV bolus model, what formula gives AUC from zero to infinity for each posterior draw?
If an oral model has \(F=0.75\), \(D=120\) mg, and \(CL=6.0\) L/h in one draw, calculate AUC from zero to infinity.
Why should one calculate \(D/CL^{(m)}\) for each draw rather than calculate \(D\) divided by the posterior mean clearance?
What is one consequence of using a sparse time grid to estimate Cmax?
What does a 95% posterior credible interval for model-based AUC mean in this setting?

Solutions. (1) \(\mathrm{AUC}_{0-\infty}^{(m)}=D/CL^{(m)}\), assuming linear elimination and complete IV input. (2) \(\mathrm{AUC}=FD/CL=0.75(120)/6.0=15.0\) mg·h/L. (3) AUC is nonlinear in clearance, so the transformed posterior mean need not equal the posterior mean of the transformed draws; draw-by-draw calculation preserves the posterior distribution of AUC. (4) The true model-predicted peak can lie between grid points, causing the grid maximum to underestimate Cmax. (5) Conditional on the model, observed data, and forecast regimen, the interval contains 95% of the posterior probability for the individual’s model-based AUC; this is not a frequentist coverage statement.