Detecting Misspecified Organ-Function Effects in PK Models

Why Organ-Function Effects Need Their Own Diagnostics

A PK model can describe the overall concentration data reasonably well and still misrepresent how clearance changes with renal or hepatic function. The average fit may hide a systematic error concentrated among patients with severe impairment. That matters because the patients for whom dose adjustment is most consequential may also be the patients whose exposures are least well predicted.

The diagnostic question is not simply whether residuals are small overall. It is whether prediction errors vary systematically with organ function, after accounting for dose, sampling time, other relevant covariates, and the model’s uncertainty. A visible pattern is evidence to investigate, not by itself proof that the organ-function relationship is the cause.

A useful starting model for clearance is a normalized power relationship:

where \(CL_i\) is individual clearance, \(R_i\) is a renal-function measure, \(R_{\mathrm{ref}}\) is its reference value, \(CL_{\mathrm{ref}}\) is clearance at that reference value, \(\beta\) describes the function relationship, and \(\eta_{CL,i}\) is an individual random effect. A corresponding model may use a hepatic-function measure, but the measure and functional form must be scientifically justified. This equation is an example, not a universal model: hepatic function may be represented by a composite clinical classification or several biomarkers, and renal function may not have a simple power relationship with drug clearance.

Interpret residuals in the context of the prediction used. For an observed concentration \(C_{ij}\) and a corresponding prediction \(\hat C_{ij}\), the raw residual is \(C_{ij}-\hat C_{ij}\). A positive value means the observation exceeds the prediction. A standardized residual divides the discrepancy by an appropriate predictive standard deviation; its exact construction depends on the model and residual-error assumptions. Population predictions show whether the model predicts the population distribution across covariates. Individual predictions condition on estimated individual effects and can be useful for checking within-person fit, but may conceal a population covariate problem because the individual effects absorb some of the discrepancy.

Assumptions and a Diagnostic Strategy

Before attributing a residual pattern to organ function, check that the comparison is meaningful. The function measure should be recorded reliably and at a clinically relevant time; patients being compared should not differ systematically in dose, sampling schedule, formulation, or other important predictors without those differences being addressed. The PK structural and residual-error models should be adequate enough that a time- or concentration-dependent error is not mistaken for a covariate effect. When using exposure measures such as AUC, ensure the exposure is estimated comparably across individuals and that assumptions about bioavailability and linearity are appropriate.

For example, a concentration residual that becomes positive late after dosing may indicate a distribution or elimination-phase problem, not necessarily an organ-function effect. Conversely, a residual pattern that is small at all sampling times in normal function but consistently positive at several times in severe impairment is more consistent with a function-stratified prediction problem. These patterns still require investigation of confounding and data support.

Worked Examples of Detecting and Interpreting Misspecification

Deriving the Exposure Consequence and Reading the Patterns

The dose implication above follows directly from the IV exposure relationship. If \(AUC_{\tau}=Dose_{\tau}/CL\), then holding the target \(AUC_{\tau}^{*}\) fixed gives \(Dose_{\tau}=AUC_{\tau}^{*}CL\). Thus, under these assumptions, a proportional change in clearance produces the same proportional change in the dose needed to maintain the target:

For oral dosing, this relationship must also account for bioavailability \(F\), since \(AUC_{\tau}=F Dose_{\tau}/CL\). If \(F\) changes with organ function or disease state, a clearance-only interpretation may mislead. More generally, if PK is nonlinear or the target depends on peak or trough concentration rather than AUC alone, the proportional rule does not apply without further analysis.

Several visual patterns help localize the problem. Residuals that are positive in the low-function group across early and late sampling times suggest systematically underpredicted concentrations there. Residuals that change sign over time may indicate an inadequate distribution or elimination profile rather than a simple clearance covariate effect. A smooth curve of residuals over function suggests a functional-form problem; a sharp transition may motivate examining a threshold, but only if the data are sufficiently dense around it. A few extreme residuals without a consistent stratum pattern are weaker evidence.

Stratified prediction plots are most informative when they display observed data, model predictions, uncertainty, and sample sizes together. Keep the function variable continuous as well as using clinically interpretable groups: categories aid communication, but can hide curvature and make results depend on arbitrary cut points. Where the data support it, compare predictions across the observed function range and mark regions with little or no data. Do not treat a curve beyond the observed range as validated by a good fit within the range.

Alternative Forms, Limitations, and Common Pitfalls

A power relationship is convenient and often interpretable, but alternatives may be more suitable. A linear additive model can represent a renal contribution plus residual nonrenal clearance. A spline can represent smooth curvature when there are enough observations across the range. A prespecified piecewise relationship can represent a scientifically motivated change in behavior. For hepatic effects, a single laboratory value may not represent the relevant physiology; a composite measure or a carefully justified set of covariates may be needed. Each added flexibility costs information and can produce unstable estimates when severe impairment is rare.

Compare alternatives using the full likelihood or another criterion appropriate to the model, parameter uncertainty, predictive checks, and plausibility—not fit improvement alone. Examine whether the alternative removes the organ-function pattern without creating a new time, dose, or subgroup pattern. Use uncertainty intervals for predictions in clinically relevant strata. If two models fit similarly but produce meaningfully different dose recommendations, the uncertainty itself is important; further data or a conservative decision may be needed.

• Confounding: Severe impairment may coincide with age, inflammation, co-medications, or disease severity. A residual pattern by renal function does not isolate a causal renal effect.
• Measurement error: An imprecise or mistimed organ-function measure can flatten or distort an estimated relationship.
• Limited support: Few patients at the most impaired levels make both the relationship and its uncertainty difficult to estimate reliably.
• Overfitting: A flexible curve may follow noise in a small data set and give unstable predictions between or beyond observed values.
• Prediction choice: Individual predictions can look acceptable after random effects absorb systematic bias. Assess population predictions across function strata as well.

Common mistake: “The alternative model has smaller residuals, so it is the correct renal model and its dose is the right dose.” Smaller residuals in the data used to fit a more flexible model are expected in many settings. The correction is to assess uncertainty, model complexity, residual structure, clinical plausibility, and out-of-sample predictive performance where available, then quantify how much the dosing decision changes. A better-fitting candidate is evidence about model adequacy, not proof of mechanism or a guarantee of improved clinical outcomes.