Joint Longitudinal PK/PD and Clinical Outcome Models

Why Link PK, PD, and Clinical Outcomes?

Repeated drug concentrations describe exposure; repeated pharmacodynamic (PD) measurements describe biological response; and a clinical endpoint records whether the response matters to patients. These data are related, but they are not interchangeable. A concentration measurement is not itself a clinical benefit, and a biomarker change does not establish a clinical effect. A joint model makes the proposed links explicit, so that predictions for candidate regimens can use all three types of information while preserving their distinct meanings.

At an interim analysis, the model can predict the distribution of clinical outcomes under each candidate regimen, conditional on the information available and the trial’s modeling assumptions. An adaptation rule can then compare those predictions with pre-specified efficacy and safety criteria. The predictions are conditional model-based forecasts, not guarantees that a regimen will produce the forecast result.

A Linked Model and Its Assumptions

Let \(i\) index participants, \(j\) PK measurement times, and \(k\) PD measurement times. Let \(C_{ij}\) be the concentration for participant \(i\) at time \(t_{ij}\), \(M_{ik}\) the PD measurement at time \(s_{ik}\), \(D_i\) the assigned regimen, and \(Y_i\) a binary clinical outcome measured over a defined follow-up window. Let \(X_i\) denote baseline covariates. A possible model is:

Here \(C_{\mathrm{pop}}\) is the population concentration-time function, \(\theta\) its PK parameters, and \(h\) a specified exposure transformation (for example, the identity or a saturating function). The participant-specific terms \(b_{Ci}\) and \(b_{Mi}\) allow participants to differ in PK and PD response. The time-average \(A_i\) summarizes the modeled biomarker trajectory over the clinically relevant interval \(0\) to \(T\); another scientifically justified summary could be specified instead. The direct regimen term \(\gamma_3D_i\) is included only if a direct effect not captured by the modeled PD summary is scientifically justified and specified in advance.

A common assumption is that \((b_{Ci},b_{Mi})\) follow a mean-zero bivariate normal distribution with covariance matrix \(\Sigma_b\). Conditional on these effects and the model parameters, the concentration and biomarker measurement errors may be assumed independent, with mean zero and specified variances. The clinical outcome is conditionally Bernoulli with the stated probability. These are assumptions to assess, not facts guaranteed by collecting repeated measurements. The model also requires appropriate timing and measurement units, a defined clinical follow-up window, and a concentration and biomarker structure adequate for the intended predictions. If observations are missing, the assumed observation process must also be defensible; for example, likelihood-based analyses commonly require missingness to be explainable by observed modeled information.

To interpret predictions under a candidate regimen as relevant to adaptation, the regimen must be well defined, the trial’s treatment assignment and data collection must be represented correctly, and the model must adequately describe how the regimen affects concentrations, PD, and outcome. A randomized regimen assignment supports comparisons of assigned regimens, but does not by itself make an association between a post-randomization exposure and outcome causal. Such a causal interpretation requires additional assumptions about the exposure-response structure and the absence of unmodeled pathways or confounding.

Joint Likelihood and Dependence

Write \(b_i=(b_{Ci},b_{Mi})\), and let \(f_b(b_i;\Sigma_b)\) be its density. Conditional on \(b_i\), a simple likelihood contribution for one participant is the product of the concentration, PD, and outcome contributions. The marginal contribution integrates over the unobserved participant effects:

where \(\psi\) collects the model parameters. The trial likelihood is the product of these participant contributions when participants are independent conditional on the design and model. Estimation may use likelihood-based methods or a Bayesian model; either way, predictions must account for uncertainty in the parameters and in unobserved participant effects. Treating the observed concentration or PD value as error-free can understate this uncertainty.

The shared effects give the model a direct mechanism for dependence. For example, under the concentration equation, the residual part of \(\log C_{ij}\) includes \(b_{Ci}\); under the PD equation, the residual part of \(M_{ik}\) includes \(b_{Mi}\). If the measurement errors are independent of each other and of the participant effects, then:

This follows because all cross-covariances involving independent, mean-zero errors are zero, leaving only the covariance of the shared participant effects. Thus, a participant’s PK profile can inform the model’s estimate of their latent PD response, and both can inform their outcome prediction. Omitting relevant dependence, or imposing dependence without support, can distort predictive uncertainty.

From the Joint Model to an Interim Prediction

For each candidate regimen, predict concentrations over the relevant time course, propagate them through the PD model, calculate the specified biomarker summary, and then predict the clinical outcome. In a fitted model, this calculation should integrate over uncertainty in model parameters and participant effects, as well as any relevant measurement or outcome variability. For instance, the predicted probability of response under regimen \(d\) is an average of the response probability over the predictive distribution of the participant’s biomarker summary and covariates, not necessarily the response probability evaluated at an average biomarker value.

The following derivation explains why. If \(A\) is uncertain and the clinical model is logistic, then the marginal response probability is:

This follows by conditioning on the quantities that determine the outcome probability and averaging that conditional probability over their predictive distribution. In general, this is not equal to applying the logistic function to the mean of its input, because the logistic function is nonlinear. A plug-in calculation can therefore give a different answer and omit predictive uncertainty.

Using Predictions in an Adaptation Rule

Predictions inform an adaptation; a rule specified before reviewing the interim data determines the action. For example, a protocol might retain candidate regimens whose predicted toxicity risk is below a specified limit, then select among those using a pre-specified clinical benefit criterion. The prediction target, follow-up window, covariate population, uncertainty summary, and handling of unavailable measurements should be defined in advance. If predictions are updated as data accrue, the analysis and decision process should preserve the trial’s intended error-rate properties.

Diagnostics, Alternatives, and Limitations

A joint model can fit one data stream while misrepresenting another. Examine the concentration-time profiles and residual patterns, whether observed biomarker trajectories are compatible with predicted trajectories, and whether predicted clinical event frequencies align with observed outcomes at the interim. Systematic time trends in residuals, substantial lack of fit across dose groups, implausible parameter estimates, or predictions that change sharply under small modeling choices are reasons to question the adaptation. These checks can reveal problems, but passing them does not prove the model is correct.

A two-stage alternative first estimates individual exposure or biomarker summaries and then relates those estimates to clinical outcomes. It can be easier to explain, but treating estimated summaries as known may ignore measurement error and uncertainty. Separate PK, PD, and outcome models may also be simpler, but without a linking structure they do not fully propagate dependence. A joint model can use shared information and quantify dependence, at the cost of stronger structural assumptions and greater computational complexity. The choice depends on the prediction target and the quality and quantity of the available data.

Important limitations include sparse sampling that cannot identify the PK or PD shape, a biomarker that is weakly related to the clinical endpoint, extrapolation to regimens or participant profiles poorly represented in the trial, and misspecified random-effects or error distributions. Predictions may be unstable when data provide little information about a model component. A model that predicts well within observed regimens may still be unreliable for a substantially different regimen. Consequently, predictions should be reported with uncertainty and interpreted within the model’s supported range.