Why Individualize a Titration?
A population pharmacokinetic model describes typical drug behavior and the variability among individuals, but a particular patient may clear a drug faster or slower than the population typical value. Sparse concentration data can help refine that patient’s estimate. The practical question is not simply whether a measured concentration is high or low: it is how the measurement, its uncertainty, and prior population information together change the predicted exposure at candidate doses.
Bayesian individualization provides a coherent way to make that update. It combines a prior distribution for an individual parameter with a likelihood for the observed concentrations, producing a posterior distribution rather than a single supposedly exact estimate. That distribution can then be used to quantify the chance that exposure lies in a target band and to compare dose choices. As in target-based titration, the target and action rule must be defined in advance; the exposure calculation alone does not establish that a dose is clinically appropriate.
A Simple Individual PK Model and Its Assumptions
Consider a setting in which, at a specified dose and sampling time, predicted concentration is inversely proportional to clearance. Assume other relevant factors are known or adequately represented by the population prediction: dose, dosing history, sampling time, bioavailability, and the model’s structural form. This simplified relationship is useful for explaining the update; a real application should use the appropriate validated pharmacokinetic model.
The normal model for log clearance ensures positive clearance on the original scale. The observation equation says that, all else equal, lower clearance tends to produce a higher concentration. Its error term is assumed to be independent of the individual parameter and, for repeated samples in the formulas below, independent across samples conditional on that parameter. If residual errors are correlated, or if there is additional unexplained variability, the likelihood must reflect it rather than treating measurements as independent evidence.
The update also depends on the sampling information being correctly specified. A sample collected at the wrong recorded time, after a missed dose, or before the assumed steady state may not be comparable with the stated population prediction. A concentration can only inform clearance through this model if the dosing history and the factors that affect the prediction are sufficiently known.
Posterior Update and Its Justification
For one measurement \(y\), the prior density is proportional to \(\exp[-\eta_i^2/(2\omega^2)]\), and the likelihood is proportional to \(\exp[-(y+\eta_i)^2/(2\sigma_e^2)]\). Multiplying them gives the posterior density. Completing the square in \(\eta_i\) shows that it is normal:
The negative sign in \(m\) follows from the observation equation: a concentration above its population prediction, so \(y>0\), favors a negative clearance deviation. For \(J\) conditionally independent measurements with the same residual variance, the corresponding result is
These expressions make the role of shrinkage explicit. When measurement error is large relative to population variability, the posterior remains closer to the population prior. More precise and consistent information moves it further toward the individual data. Repeated measurements increase precision only to the extent that their errors are genuinely independent and the model remains appropriate.
Worked Example: Updating Clearance From One Sparse Sample
Suppose a hypothetical patient has a prior clearance deviation \(\eta_i\sim N(0,0.35^2)\). At a known dose and sampling time, the observed concentration is 25% above the population prediction, so \(y=\log(1.25)=0.22314\). Assume residual log-concentration error has standard deviation \(0.20\), hence variance \(0.04\). These assumptions make the normal update applicable.
The posterior variance and mean are
The posterior standard deviation is \(\sqrt{0.03015}=0.1736\). The posterior mean is negative, as expected from the higher-than-predicted concentration, but it is not as extreme as the data-only implication \(\eta_i=-0.2231\): the prior pulls the estimate toward zero. The posterior median clearance relative to typical clearance is \(\exp(-0.1682)=0.8452\). This is a model-based estimate, not a direct measurement of clearance.
From Posterior Clearance to Exposure and Dose
Under the same proportional assumptions, exposure at a fixed regimen is inversely proportional to clearance. If \(A_{\mathrm{pop}}\) is the population-predicted exposure at the current dose, then the individual’s exposure is \(A=A_{\mathrm{pop}}\exp(-\eta_i)\). Thus, if \(\eta_i\mid\text{data}\sim N(m,v)\), then \(\log A\mid\text{data}\) is normal with mean \(\log A_{\mathrm{pop}}-m\) and variance \(v\). The posterior median exposure is \(A_{\mathrm{pop}}\exp(-m)\); the posterior arithmetic mean is \(A_{\mathrm{pop}}\exp(-m+v/2)\). They differ because exposure is lognormal, not normal.
A 95% posterior credible interval for exposure is obtained by exponentiating the corresponding normal interval for \(\log A\). Conditional on the model, prior, and observed data, it contains 95% posterior probability. It is not a claim that there is a 95% probability that a fixed parameter lies in a particular frequentist confidence interval.
Worked Example: Exposure Probability and a Provisional Dose Adjustment
Continue the hypothetical example. Suppose population-predicted exposure at the current dose is \(80\ \mathrm{mg\cdot h/L}\), and the desired exposure band is \(80\) to \(120\ \mathrm{mg\cdot h/L}\). From the posterior above, \(\log A\) has mean \(\log(80)+0.1682159\) and variance \(0.0301538\). Its standard deviation is \(0.17365\), so the posterior arithmetic mean exposure is
For the target-band probability, standardize the two exposure limits on the log scale. The lower and upper standardized limits are \((\log(80)-\log(80)-0.1682159)/0.17365\approx-0.9687\) and \((\log(120)-\log(80)-0.1682159)/0.17365\approx1.3664\). Therefore,
The current regimen therefore has about a 74.8% posterior probability of placing exposure in this band under the stated model. Suppose, for illustration, that the decision objective is to set the posterior arithmetic mean exposure to \(100\ \mathrm{mg\cdot h/L}\), and exposure is dose-proportional. The continuous dose multiplier is \(100/96.09=1.0407\). At a current dose of \(200\ \mathrm{mg}\), this gives \(208.1\ \mathrm{mg}\). If only 10-mg increments are available, \(210\ \mathrm{mg}\) is the nearest feasible dose. This is a provisional model-based choice, not an automatic instruction: the target, safety constraints, dose limits, and consequences of uncertainty must also be considered. Dose rounding means the chosen dose will not exactly attain the continuous target.
Sequential Data, Diagnostics, and Failure Modes
With a new sample, the posterior from earlier samples can serve as the next prior, provided the model and conditional independence assumptions remain appropriate. The resulting update should be checked against the data rather than accepted mechanically.
Worked Example: Adding a Second Measurement
Suppose the patient in the first example has a second concentration measured at a known comparable dose and time. Its log ratio to the population prediction is \(y_2=0.05\), and its residual standard deviation is again \(0.20\). Assume its error is conditionally independent of the first sample’s error. The first posterior has mean \(-0.1682\) and variance \(0.03015\); the second measurement contributes likelihood precision \(1/0.04=25\).
Combining prior and likelihood precision gives
and the updated mean is
The second observation is less elevated than the first. It shifts the estimated clearance deviation toward zero and reduces its variance. The posterior standard deviation is about \(0.1311\). Before interpreting this as evidence of changing clearance, check whether dose, timing, adherence, and assay conditions were comparable. A discrepancy can reflect model mismatch or recording error rather than a true biological change.
Useful diagnostics include comparing each observed log concentration ratio with its posterior predictive distribution, checking whether residuals are systematically positive or negative across times, and investigating unexpectedly large discrepancies. A posterior predictive check asks whether a new observation is plausible under the model after accounting for parameter uncertainty and residual error. It should use the observation’s actual dose and sampling time. Repeatedly surprising measurements, dose-dependent residual patterns, or a concentration incompatible with the known dosing history are reasons to review the model and data.
A common mistake is to treat each concentration as an exact readout of clearance and convert it directly into a dose change. That ignores residual error, population variability, and exposure uncertainty. The correction is to update a distribution, predict exposure under candidate regimens, and make the action rule explicit. Another pitfall is counting correlated samples as independent: doing so makes the posterior artificially precise. The observation model must represent shared error or dependence when present.
How This Approach Differs From Alternatives
| Approach | What it uses | Strength and limitation |
|---|---|---|
| Population-based dosing | Typical parameter values and population variability | Useful before individual data are available; may not reflect an individual’s clearance. |
| Individual plug-in estimate | A point estimate derived from the individual’s measurements | Simple to act on, but can overreact to sparse or noisy data and does not by itself quantify predictive uncertainty. |
| Bayesian individualization | Population prior information plus an individual likelihood | Accounts for shrinkage and posterior uncertainty; depends on an adequate prior, likelihood, sampling information, and model. |
The Bayesian approach is not automatically superior if its assumptions are wrong. Its advantage is that it makes the combination of population and individual information explicit and carries uncertainty into predictions. In practice, more complex population PK models may be needed to account for multiple compartments, nonlinear elimination, changing bioavailability, covariates, or parameter correlations. A simple normal update should not be substituted for a suitable model merely because it is convenient.
Interpretation for Titration
A posterior exposure prediction supports a decision; it does not define the decision objective. A titration rule might aim to achieve a specified posterior mean, maximize the probability of entering a target band, or limit the probability of exceeding an upper exposure threshold. These objectives can recommend different doses, particularly when uncertainty is substantial. The intended criterion should be stated, including any safety constraints, allowable dose increments, and the time at which the next assessment is expected.
The analysis is limited by the quality and relevance of the population model, the information in sparse samples, and the appropriateness of the observation and exposure relationships. It may not separate clearance from other sources of variation when the data do not identify them. It also does not resolve whether a biomarker response is delayed or noisy; that requires modeling the response measurement and its timing, not just updating PK from concentrations. If the dose-response objective is uncertain or the model is misspecified, a precise-looking posterior can still lead to a poor action.
Check Your Understanding
Work through each question, then compare with the solutions below.
- With prior \(\eta_i\sim N(0,0.09)\), one observation \(y=0.20\), and residual variance \(0.04\), calculate the posterior mean and variance.
- If \(\log(A/A_{\mathrm{pop}})\mid\text{data}\sim N(0.10,0.0225)\), find the posterior probability that \(A/A_{\mathrm{pop}}\) is between 0.80 and 1.25.
- Current predicted mean exposure is \(75\ \mathrm{mg\cdot h/L}\). Under dose proportionality, what continuous dose multiplier targets a posterior mean exposure of \(90\ \mathrm{mg\cdot h/L}\)?
- Name two reasons a concentration may not be a valid update under the simple model, even if the assay result itself is accurate.
Solutions. 1. The posterior variance is \(0.09(0.04)/(0.09+0.04)=0.02769\). The mean is \(0.02769(-0.20/0.04)=-0.1385\). 2. The standard deviation is \(0.15\). The standardized limits are \((\log(0.80)-0.10)/0.15=-2.154\) and \((\log(1.25)-0.10)/0.15=0.8210\). Thus the probability is approximately \(\Phi(0.8210)-\Phi(-2.154)=0.7942-0.0156=0.7786\), or 77.9%. 3. The multiplier is \(90/75=1.20\), a 20% increase before dose rounding and other constraints. 4. Examples include an incorrect sampling time or dosing history, a missed dose, sampling before assumed steady state, and a model that does not represent the relevant concentration-time behavior.