Worked Case of Bayesian Dose Adaptation After a Trough

From a Measured Trough to a Dose Decision

A trough concentration is useful only when it changes a decision. After observing one, the clinician needs to account for what it says about this individual’s pharmacokinetics, how uncertain that information remains, and what exposure might follow each candidate dose. A dose that puts the forecast’s center near a target is not automatically the best choice: the spread of the forecast and the consequences of low or high exposure matter as well.

This hypothetical case uses a log-normal exposure model and three candidate daily doses. Exposure is expressed as a ratio to a chosen target trough, so a ratio of 1.0 is the target, and the acceptable range is 0.8 to 1.2. The example illustrates a decision calculation, not a universal therapeutic range or a prescribing recommendation.

Definition: A posterior predictive distribution for a future exposure is the distribution obtained by averaging the future-observation model over the posterior uncertainty in the individual parameters. It describes uncertainty in the exposure that could be observed under a specified dose, rather than uncertainty in a parameter alone.

Model and Assumptions for the Case

Let \(\theta\) denote this individual’s log trough ratio at a reference dose of 10 mg once daily. The measured trough is summarized by \(y\), the observed log trough ratio on that reference dose. For this simplified update, assume a normal prior for \(\theta\) and a normal measurement model:

\theta \sim N(0,0.2^2), \qquad y\mid\theta \sim N(\theta,0.15^2).

The model treats the trough as measured at the intended sampling time and assumes that the reference dose and dosing history are known. For the dose comparison, assume dose-proportional exposure over the candidate range, steady-state conditions, and no dose-related change in clearance or bioavailability. A future observed trough also has residual variability. Represent this by an independent log-scale error \(\epsilon\sim N(0,0.0895^2)\). These are working assumptions for the illustration; they should be checked against the drug, sampling record, and clinical context before using such a forecast.

The measured trough informs \(\theta\), while the future residual error represents variation between a modeled typical exposure and a subsequent measurement. Combining them is important: reporting only the posterior mean of \(\theta\) would omit uncertainty in both the estimated individual parameter and the future observation.

Worked Example: Updating the Individual Estimate From the Trough

Suppose the hypothetical observed log trough ratio is \(y=-0.02070\). The prior variance is \(0.2^2=0.0400\), and the measurement-error variance is \(0.15^2=0.0225\). Under the stated normal model, the posterior variance and mean are:

V_{\theta\mid y}= \left(\frac{1}{0.0400}+\frac{1}{0.0225}\right)^{-1} =(25+44.4444)^{-1}=0.01440,

E(\theta\mid y)=V_{\theta\mid y} \left(\frac{0}{0.0400}+\frac{-0.02070}{0.0225}\right) =0.01440(-0.9200)=-0.01325.

Thus the updated individual estimate is \(\theta\mid y\sim N(-0.01325,0.12^2)\), since \(\sqrt{0.01440}=0.12\). The posterior mean is closer to the prior mean of zero than the observed value is, because the prior contributes information and the trough is not treated as a perfectly precise measurement.

The observation corresponds to a trough ratio of \(\exp(-0.02070)\approx0.9795\), close to the target ratio. That fact alone does not establish that the current dose should continue: the posterior retains uncertainty, and a dose decision concerns future exposure and its clinical consequences.

Forecasting Exposure Under Candidate Doses

For a candidate dose \(d\) in mg, define \(D_d\) as the log of the future trough ratio to target. Under dose proportionality, changing from 10 mg to \(d\) adds \(\log(d/10)\) to the individual log exposure. The future log trough is therefore the updated individual parameter, the dose adjustment, and the future residual error combined:

D_d=\theta+\log(d/10)+\epsilon.

Conditional on the observed trough, \(\theta\) has posterior variance \(0.01440\), and \(\epsilon\) has variance \(0.0895^2=0.008010\). Assuming independence, their variances add. Hence the predictive standard deviation is \(\sqrt{0.01440+0.008010}=0.1497\), and:

D_d\mid y \sim N\left(-0.01325+\log(d/10),\,0.1497^2\right).

The target range \(0.8\leq \exp(D_d)\leq1.2\) is equivalent to \(\log(0.8)\leq D_d\leq\log(1.2)\). If \(m_d=-0.01325+\log(d/10)\), its predictive probability is:

P(0.8\leq\exp(D_d)\leq1.2\mid y) = \Phi\left(\frac{\log(1.2)-m_d}{0.1497}\right) - \Phi\left(\frac{\log(0.8)-m_d}{0.1497}\right),

where \(\Phi\) is the cumulative distribution function of a standard normal random variable. The lower-tail probability is the probability of exposure below 0.8; the upper-tail probability is the probability above 1.2. Reporting all three probabilities makes the direction of risk visible.

Worked Example: Comparing 8 mg, 10 mg, and 12 mg

Use the updated forecast to compare the three hypothetical doses. The calculations below use displayed means rounded to four significant figures and a predictive standard deviation of 0.1497.

At 8 mg: \(m_8=-0.01325+\log(0.8)=-0.2364\). The standardized target limits are approximately \((\log(0.8)-m_8)/0.1497=0.0885\) and \((\log(1.2)-m_8)/0.1497=2.7971\). Thus the probability in range is approximately \(0.9974-0.5353=0.4621\). The probability below range is about 0.5353, and the probability above range is about 0.002579.

At 10 mg: \(m_{10}=-0.01325\). The standardized limits are approximately \((-0.22314+0.01325)/0.1497=-1.4021\) and \((0.18232+0.01325)/0.1497=1.3063\). The probabilities below and above range are approximately 0.0805 and 0.0958, respectively. The in-range probability is therefore \(1-0.0805-0.0958=0.8237\).

At 12 mg: \(m_{12}=-0.01325+\log(1.2)=0.1691\). The standardized limits are approximately \((-0.22314-0.1691)/0.1497=-2.620\) and \((0.18232-0.1691)/0.1497=0.08832\). The probability below range is about 0.0044, and the probability above range is about 0.4648. The in-range probability is \(0.5352-0.0044=0.5308\).

The 10 mg option has the greatest probability of falling in the stated range. The 8 mg forecast places substantial probability below range, while the 12 mg forecast places substantial probability above range. These probabilities are predictive statements under the model and assumptions, not guarantees about an individual future measurement.

Include Clinical Consequences in the Choice

A probability of being in range is a useful summary, but it assigns no distinction between different kinds of out-of-range exposure. To illustrate an explicit decision rule, suppose the hypothetical clinical loss is 2 units for an exposure below 0.8, 0 units for an exposure in range, and 5 units for an exposure above 1.2. The units are arbitrary decision weights, not measured clinical outcomes. Expected loss is the sum of each outcome’s probability multiplied by its assigned loss.

Worked Example: Selecting a Dose by Expected Loss

For 8 mg, the expected loss is \(2(0.5353)+0(0.4621)+5(0.002579)=1.0835\) loss units. For 10 mg, it is \(2(0.0805)+0(0.8237)+5(0.0958)=0.6400\) loss units. For 12 mg, it is \(2(0.0044)+0(0.5308)+5(0.4648)=2.333\) loss units.

Under these stipulated consequences, 10 mg has the smallest expected loss and is preferred among the three candidates. This choice reflects both the predictive probabilities and the greater assigned penalty for above-range exposure. A different, clinically justified loss function could lead to a different decision; the model does not determine the consequences.

A related consideration is whether a candidate exceeds a specific safety-risk threshold. For example, if a hypothetical policy requires \(P(\exp(D_d)>1.2\mid y)\leq0.10\), then 8 mg and 10 mg satisfy the constraint (about 0.002579 and 0.0958), while 12 mg does not (about 0.4648). A risk constraint can be used alongside an expected-loss comparison, but it should be justified for the clinical setting rather than selected to make a preferred dose pass.

Diagnostics, Alternatives, and Limitations

Before interpreting the result, check the inputs and the model’s fit. Confirm that the trough was collected at the intended time relative to the dose, that the dose history is reliable, and that steady state is plausible. Check units, the target definition, and whether the recorded concentration is compatible with the assay and model. A mistimed sample or missed doses can make the measured trough a poor indicator of stable individual PK.

The calculation also depends on dose proportionality, the log-normal predictive form, and the assumed uncertainty components. If exposure is not proportional to dose, the term \(\log(d/10)\) is not an adequate dose adjustment. If the individual posterior is skewed or multimodal, a single normal approximation may hide important alternatives. Posterior predictive checks and residual diagnostics can help reveal systematic lack of fit or unusual residuals; calibration assesses whether predictive intervals attain their stated coverage over suitable cases. These checks do not prove that the assumptions hold for this particular decision.

The approach differs from selecting a dose using only a point estimate. A plug-in calculation might use the posterior mean and ignore the predictive spread; it cannot show the probability of clinically important tails. Conversely, a fully simulation-based posterior predictive calculation can represent nonlinear PK, parameter correlation, and non-normal posteriors more flexibly, but requires a model and computation appropriate to those features. The normal calculation here is transparent and convenient when its assumptions are reasonable; it is not a substitute for a more realistic model when they are not.

Finally, the dose grid is limited to 8, 10, and 12 mg. The best choice among these does not prove that one is optimal among all feasible doses, nor does a forecast alone determine whether to change therapy. Clinical context, other observations, practical dose increments, and the consequences of acting or waiting may matter. The calculation makes its assumptions and trade-offs explicit so that they can be evaluated.

Common Pitfall

Wrong reasoning: “The posterior mean exposure is near the target, so the dose is safe and should be chosen.” This ignores predictive uncertainty and can overlook a substantial probability of harmful exposure.
Correction: Calculate the predictive distribution for each candidate dose, including posterior parameter uncertainty and future residual variability. Then compare relevant probabilities or expected clinical loss under clearly stated consequences.
Another error: Treating a high in-range probability as a universal decision rule. The probability is meaningful only relative to a defined target range; asymmetric clinical costs or risk constraints may favor a different option.

Key takeaway: A trough updates an individual’s PK forecast, but adaptation requires predicting exposure under each dose and connecting that distribution to clinically justified consequences. Keep posterior uncertainty, future variability, target definitions, and model assumptions distinct.

Check Your Understanding

Answer each question using the model and hypothetical quantities in this case.

Why is the predictive variance larger than the posterior variance of \(\theta\) alone?
Under the stated loss weights, which of 8 mg, 10 mg, and 12 mg has the smallest expected loss?
If the measured trough is taken at an uncertain time after dosing, which assumption is in question and why does it matter?
Suppose an upper-exposure risk limit is 0.10. Which candidate doses satisfy it?

Solutions. (1) The future observation includes both uncertainty about \(\theta\) and independent residual variability, so the variances add: \(0.01440+0.008010=0.02241\), giving predictive SD 0.1497. (2) The expected losses are approximately 1.0835, 0.6400, and 2.333 units for 8, 10, and 12 mg, respectively; 10 mg is smallest. (3) The assumption that the sample represents a correctly timed trough is questionable. A timing error can change the observed concentration for reasons not captured by the individual parameter update, making the forecast unreliable. (4) The upper-tail risks are approximately 0.002579, 0.0958, and 0.4648. Therefore 8 mg and 10 mg satisfy the stated limit, while 12 mg does not.