Bayesian Forecasting With Time-Varying Clearance

Why Clearance May Need to Change Over Follow-Up

An individual forecast based on one fixed clearance value can become misleading if the patient’s clearance changes during follow-up. A change in organ function, for example, may alter drug elimination even when the prescribed dose is unchanged. Serial concentrations and clinical measurements can provide evidence about that change, but neither necessarily measures clearance without error. A Bayesian state model offers a way to combine prior information, evolving clearance, and new observations while retaining uncertainty in the forecast.

The model developed here treats log clearance as a latent state that can move over time. Concentrations are linked to that state through a pharmacokinetic prediction, and a clinical measurement may also provide information through a calibrated relationship with the state. The aim is not to assume that every observed fluctuation represents a true clearance change: the model separates estimated state change from measurement error.

This random-walk specification has a clear interpretation: in the absence of new information, the expected next log-clearance equals the current log-clearance, while uncertainty increases by \(q\) per transition. It does not assert that clearance is constant, nor does it specify a systematic trend. A trend or a different transition model may be needed when clinical knowledge supports one.

Observation Models and Assumptions

The concentration observation model must reflect the dosing history and pharmacokinetic structure. In a full analysis, the predicted concentration is calculated from the dosing record and the clearance path, together with other relevant PK parameters. For a transparent scalar illustration, suppose a concentration has been measured under conditions where the predicted concentration is locally proportional to inverse clearance. Let \(C_k^0\) be the prediction at \(CL_{\mathrm{ref}}\). On the log scale, write

This is an illustrative observation model, not a universal concentration equation. It is appropriate only when the relevant PK prediction has approximately the stated sensitivity to log clearance. More generally, a local approximation may use \(r_k=-s_kx_k+\epsilon_k\), where the sensitivity \(s_k\) comes from the PK model. The model also assumes the stated error distribution is adequate and that the observation is correctly aligned with dose and sampling times.

A clinical measurement can be incorporated if its relationship with the latent clearance state is specified and calibrated. For example, after an appropriate transformation and scaling, let \(b_k=x_k+\eta_k\), with \(\eta_k\sim N(0,R_B)\). Here \(b_k\) is a proxy for log clearance, not clearance itself. In practice, the relationship may include a slope, intercept, and additional uncertainty; treating an uncalibrated marker as a direct clearance measurement would be unjustified.

The simple Gaussian updates below rely on a Gaussian state distribution, Gaussian independent measurement errors, and known transition and observation variances. They also presume that the reference clearance and other PK quantities used to form \(C_k^0\) are sufficiently well specified. If those quantities are uncertain, their uncertainty should be represented rather than silently ignored. Earlier material on sequential Bayesian updating provides the general logic of revising an individual distribution as observations arrive; here the state itself is allowed to evolve between observations.

Sequential Bayesian Updating of the State

For a scalar Gaussian state with prior \(x\sim N(m,P)\) and an observation \(u=x+e\), where \(e\sim N(0,R)\) independently of \(x\), the posterior is Gaussian. Its mean and variance are

For a concentration observation \(r=-x+\epsilon\), set \(u=-r\). For the calibrated clinical proxy \(b=x+\eta\), set \(u=b\). If the two observations are conditionally independent given the state, they can be applied one at a time. Before the next occasion, the random-walk transition gives \(m_{\mathrm{next}}=m_{\mathrm{post}}\) and \(P_{\mathrm{next}}=P_{\mathrm{post}}+q\).

Forecasting and Checking the Model

A forecast begins by propagating the current state distribution through the transition model. One week after the last update in the example, the state has mean \(0.2342\) and variance \(0.004840+0.01=0.01484\). Because clearance is an exponential transformation of the state, its distribution is lognormal: its median is \(CL_{\mathrm{ref}}\exp(m)\). A central 95% interval can be obtained by transforming \(m\pm1.96\sqrt{P}\), provided the Gaussian state approximation is appropriate.

Diagnostics should address whether the state model and observation links are credible, not just whether the arithmetic update is correct. Forecast residuals that remain systematically positive or negative may indicate a trend, incorrect dosing or sampling times, an inadequate concentration observation model, or a clinical proxy relationship that does not apply to this patient. Residual variation much larger than predicted can indicate underestimated assay or process variability. Conversely, a model that responds sharply to every small concentration fluctuation may have an implausibly large \(q\) or an error variance that is too small. Posterior predictive checks and calibration across patients help assess these features; patient-level checks alone cannot establish that the population-level model is adequate.

The random walk is only one approach. A fixed-clearance model is simpler and may be suitable when there is no credible evidence of change, but it cannot represent gradual evolution. Independent occasion-specific clearance values permit abrupt changes but discard the continuity encoded by a state transition. A mechanistic model driven by a measured physiologic process may be more interpretable when that process is well characterized, but it requires additional assumptions and data. The random walk is a flexible compromise, not a substitute for a mechanistic explanation when one is available.

Common Pitfalls and Limitations

• Interpreting a noisy concentration as a clearance measurement. The concentration is evidence about clearance conditional on the dosing history, sampling time, PK model, and error variance; it is not a direct measurement of clearance.
• Using a clinical marker without a calibrated link. A marker can inform the state only through a defensible measurement model. Its numerical value cannot be substituted for \(x_k\) without justification.
• Choosing process variance by convenience. A small \(q\) suppresses change and can produce overconfident forecasts; a large \(q\) allows rapid movement and may overreact to noise. The value should be supported by data or prior clinical knowledge, with sensitivity checks where appropriate.
• Confusing a forecast distribution with certainty about future clearance. The forecast combines uncertainty about the current state, future state evolution, and possibly residual error. Omitting any of these can make an interval too narrow.

The scalar Gaussian update is not exact for every PK model. When concentrations depend nonlinearly on a clearance path, or when multiple PK parameters and states are uncertain, numerical Bayesian methods may be needed. A simple state model also does not identify the cause of changing clearance, and it may extrapolate poorly if the process changes direction or follows a strong trend. The forecast should therefore be interpreted as conditional on the specified transition, observation models, dosing record, and assumptions.