Bayesian Therapeutic Drug Monitoring for Within-Trial Adaptation

Why Individual Concentrations Can Change a Regimen

A population pharmacokinetic model describes typical parameter values and between-participant variation. A participant’s measured concentrations provide additional information about that participant’s parameters. Bayesian therapeutic drug monitoring (TDM) combines those sources of information to form a participant-specific posterior distribution, then uses that distribution to predict exposure under candidate regimens.

The purpose here is not to treat a measured concentration as an exact estimate of clearance. Rather, a concentration is an observation generated by a pharmacokinetic model with uncertainty. The resulting posterior uncertainty matters: two participants with the same posterior median clearance can have different confidence in that estimate, and therefore different probabilities of reaching an exposure target.

The examples use a deliberately simple model for a continuous intravenous infusion at steady state. The same Bayesian structure applies to more elaborate pharmacometric models, but their likelihoods and predictions may require numerical computation.

Model and Posterior Distribution

The steady-state relationship \(C=R/CL\) assumes linear pharmacokinetics, a constant infusion, and enough time at that rate to reach steady state. The observation model assumes that the logarithm of the clearance proxy has mean \(\eta\), with known standard deviation \(\sigma\), and that observations are conditionally independent. The prior summarizes information available before this participant’s concentrations are used; it may be informed by a population model and relevant covariates.

These assumptions are consequential. If concentrations were measured before steady state, if clearance changes over time, or if the model does not describe the dosing and sampling conditions, this likelihood is not appropriate as written. In an operational trial, the likelihood, prior, and adaptation rule should be specified before the participant-level decision is made.

For this normal prior and normal likelihood, the posterior is normal. Its variance and mean are:

This result follows by multiplying the prior density by the likelihood. Apart from constants not depending on \(\eta\), the log posterior is

Collecting the terms involving \(\eta^2\) gives precision \(1/\tau_0^2+n/\sigma^2\). Collecting the linear terms gives \(\mu_0/\tau_0^2+\sum_i y_i/\sigma^2\). Completing the square therefore gives a normal posterior with the variance and mean above. The posterior mean is a precision-weighted compromise between the prior mean and the observed values: data with more precise measurements receive more weight.

Predicting Exposure and Choosing a Candidate Rate

For a candidate infusion rate \(R\), the predicted steady-state concentration is \(C=R/CL\). Because \(CL=5\exp(\eta)\) L/h, the posterior distribution of log concentration is normal:

where \(R/5\) is expressed in mg/L when \(R\) is in mg/h. This expression predicts concentration from uncertainty in the individual parameter; it does not add extra future measurement error. If a predictive distribution for a future observed concentration is required, the observation-error component must also be included.

A straightforward policy is to choose the rate that makes the posterior median concentration equal a protocol-defined target \(C^*\). Since the posterior median clearance is \(5\exp(\mu_n)\), the corresponding rate is:

This rule targets the median predicted concentration, not the posterior mean concentration. A different policy could target a posterior expectation or the probability of falling within a target range. The policy should match the clinical objective and be prespecified; it should not be selected after seeing which rule gives the most favorable answer.

Posterior Target Attainment, Not Just a Point Estimate

When the objective is to keep concentration within an acceptable interval \([C_L,C_U]\), a posterior target-attainment probability can be more informative than a single posterior median. For a candidate rate \(R\), the concentration lies in that interval precisely when:

The order of the bounds reverses relative to concentration because concentration decreases as clearance increases. The posterior probability of target attainment is therefore the posterior normal probability between these two log-clearance bounds. A protocol might select the smallest permitted rate whose attainment probability exceeds a prespecified threshold, while also imposing safety limits. The threshold and any tie-breaking rule are design choices, not consequences of Bayes’ theorem.

Shrinkage, Diagnostics, and Failure Modes

Bayesian individual estimation naturally shrinks noisy participant-level evidence toward the prior. This is often useful when only a few concentrations are available: it avoids treating a highly variable individual estimate as if it were precise. Shrinkage is not automatically beneficial, however. A prior that is poorly matched to the participant or trial population can pull estimates in the wrong direction.

Before acting on a posterior, examine whether the data are compatible with the model. Useful checks include comparing observed and model-predicted concentrations, checking residuals for systematic patterns across time or concentration, verifying dose and sampling records, and assessing whether the posterior is unusually sensitive to reasonable changes in the prior. If multiple observations are available, a pattern in which residuals consistently rise or fall over time may suggest that the steady-state or constant-clearance assumptions are failing.

A particularly important failure mode is to use a point estimate while ignoring posterior uncertainty. A rate that matches the posterior median may still leave substantial probability outside a clinically acceptable range. Conversely, an attainment probability can look reassuring only because the assumed model is too confident. Unmodeled variability, misspecified error, nonlinearity, time-varying clearance, or an unsuitable prior can make posterior intervals and probabilities misleading.

How TDM Fits Within a Trial

Bayesian TDM differs from simply assigning every participant a dose based on a population-average parameter. The population approach does not use an individual’s observed concentrations to update that participant’s clearance estimate. It also differs from an unregularized individual estimate, which may be unstable when data are sparse. Bayesian updating combines both sources, while retaining an explicit distribution for uncertainty.

For a within-trial adaptation, the statistical calculation is only one part of the decision. The protocol should specify the eligible observations, model and prior, target definition, candidate rates, adaptation threshold, safety constraints, and handling of missing or inconsistent data. Decisions must also respect the trial’s allocation and oversight procedures. TDM estimates an individual pharmacokinetic parameter and predicts exposure under a model; by itself it does not establish a treatment effect or justify a causal claim about clinical outcomes.

The introductory material on borrowing external pharmacometric data at interim concerns how external information can contribute to a model. In individual TDM, the practical question is how participant-specific concentration data update the applicable prior and how that posterior is translated into an allowed regimen. Borrowed information should not be treated as automatically exchangeable with the participant’s data: its relevance and any safeguards against conflict must be established in the model and decision procedure.

Common Mistake

• Wrong reasoning: “The concentration at 10 mg/h implies clearance of exactly 4 L/h, so the dose for a 2 mg/L target must be exactly 8 mg/h.”
• Correction: That calculation treats a noisy measurement as error-free and ignores prior information and posterior uncertainty. Under the stated model, the observation contributes evidence about clearance; the posterior distribution then supports a prespecified decision, such as a median-target rate or a target-attainment rule.