Predicting Clinical Biomarkers at Prespecified Visit Times

From Mechanistic Trajectories to Scheduled Biomarkers

QSP models often generate continuous-time trajectories, while a clinical protocol records biomarkers at particular visits: for example, baseline and weeks 4, 12, and 24. The prediction target is therefore not merely a curve. It is the distribution of biomarker values at specified times, and of summaries such as change from baseline or the probability of crossing a response threshold. Those quantities depend on evaluating each simulated trajectory at the same clinically defined times before summarizing across trajectories.

The target must also say what the biomarker value represents. A model state may be a latent biological quantity, whereas a clinical endpoint may be a measured laboratory value after a measurement process. Decide which scale is required, and ensure the model output has been mapped to that scale before extracting visits. The predictions below are conditional on the ensemble of trajectories supplied; they do not, by themselves, establish that the model is clinically validated.

Before calculating these quantities, specify the visit times and their time origin, the biomarker units, the baseline definition, and the direction and scale of a clinically meaningful response. Confirm that every trajectory uses the same dosing and visit-time conventions. If the ensemble is weighted, those weights must also be carried into the summaries; the examples here use equally weighted trajectories.

Evaluating a Trajectory at a Visit

If the model solver can return output exactly at a scheduled time, use that value. If the solver provides values only on a time grid, an interpolation rule may be needed. Linear interpolation between adjacent output points is one option when the trajectory is sufficiently smooth over that interval. Interpolation is a numerical approximation, not a new model prediction; it should not conceal sharp transitions, discontinuities, or inadequately resolved dynamics.

For an ensemble, apply the same visit-time rule to every trajectory. The empirical predictive distribution at visit \(t_k\) is represented by the set of values \(Y_1(t_k),\ldots,Y_M(t_k)\), where \(M\) is the number of trajectories. Its empirical cumulative distribution is

where \(I\{\cdot\}\) equals 1 when the condition is true and 0 otherwise. Quantiles of this empirical distribution can summarize predictive spread; a prediction interval made from its quantiles describes the represented predictive distribution under the chosen ensemble and model assumptions. It is not a confidence interval for a fixed parameter.

Changes from Baseline and Responder Thresholds

Calculate a change within each trajectory before summarizing across trajectories. The absolute change \(D_{jk}\) retains the biomarker's units. A percent change is \(100D_{jk}/Y_j(t_0)\), provided the baseline value is nonzero and the ratio is meaningful for the biomarker. For a biomarker where lower values are favorable, percent reduction is often reported as \(100\{Y_j(t_0)-Y_j(t_k)\}/Y_j(t_0)\); this reverses the sign relative to percent change. State the convention explicitly.

A responder definition turns a continuous trajectory-level outcome into an indicator. For example, if response means at least a 15% reduction from baseline, define \(R_j=I\{100[Y_j(t_0)-Y_j(t_k)]/Y_j(t_0)\geq15\}\). For equally weighted trajectories, the predicted responder proportion is \(\hat p=M^{-1}\sum_j R_j\). This is the fraction of represented trajectories meeting the stated criterion. It is not the probability that any particular patient will respond unless the ensemble and weighting support that interpretation.

Longitudinal Contrasts Preserve Pairing

A longitudinal contrast compares values at two times within a trajectory. For visits \(t_a\) and \(t_b\), define \(C_j=Y_j(t_b)-Y_j(t_a)\). Computing this for each trajectory preserves the fact that both values came from the same virtual patient or uncertainty draw. The ensemble of \(C_j\) values gives the distribution of individual within-trajectory contrasts.

The importance of pairing follows directly from the variance identity. Write \(A=Y(t_a)\) and \(B=Y(t_b)\) for values on the same trajectory. Then

Thus, the distribution of a within-trajectory contrast depends on the covariance between visits. Subtracting values drawn independently from the two visit-specific marginal distributions discards that covariance and generally gives the wrong spread. The mean contrast equals the difference of the two means when calculated on the same equally weighted ensemble, but that equality does not make the contrast distributions interchangeable.

Assumptions, Diagnostics, and Alternatives

Scheduled-time predictions are interpretable only if the simulated ensemble matches the intended predictive target. Check that baseline, visit times, biomarker definition, units, dosing schedule, and any clinical visit-window convention are consistent across trajectories and with the question being answered. If the target is a latent model state, do not label it as a measured laboratory outcome; if a measured outcome is required, use an appropriate mapping to that scale.

• Check time alignment. Verify time origin, units (such as days versus weeks), and whether each visit is measured from treatment start or another event.
• Check interpolation. Compare grid spacing with the speed of the modeled dynamics. Refine solver output near rapid changes, dose events, or discontinuities instead of relying on coarse linear interpolation.
• Check baseline and denominator. Confirm that the same protocol-defined baseline is used for each trajectory. Percent change is unstable or undefined for zero or near-zero baselines.
• Check ensemble pairing and weights. Keep each trajectory's visit values together. Preserve any ensemble weights rather than treating unequal weights as equal counts.
• Check summaries against trajectories. Inspect trajectories and empirical distributions for implausible values, multimodality, or a few influential outcomes. A mean alone may obscure these features.

There are several reasonable choices, but they answer different questions. Direct solver evaluation is preferable when available; interpolation is a practical approximation when justified by adequate temporal resolution. Absolute change is useful when the biomarker's units and scale matter, while percent change expresses change relative to baseline but can behave poorly near zero. A threshold indicator answers a binary clinical question but discards information about how far each trajectory lies above or below the cutoff. A paired contrast answers a within-trajectory longitudinal question; comparing two marginal distributions alone does not retain that information.

A common mistake is to take the median or mean at each visit and then treat the resulting summary curve as if it were a typical trajectory. In general, a summary at each time can combine different trajectories, so its changes need not represent the ensemble's individual changes. The correction is to calculate the desired change, threshold indicator, or contrast separately for every trajectory first, and only then summarize those derived outcomes.

These calculations do not remove limitations in the mechanistic model, the uncertainty ensemble, or the clinical endpoint definition. They also do not establish that an estimated tail probability or responder proportion is precise merely because it is numerically reported. Monte Carlo precision and broader predictive uncertainty are distinct, and the ensemble must be sufficiently large and representative for the intended summary.